What Is Vocabulary | Word Embeddings and Semantic Vectors — How Computers Place Words in Meaning Space

WORD EMBEDDINGS · SEMANTIC VECTORS · STATIC EMBEDDINGS · CONTEXTUAL EMBEDDINGS · COSINE SIMILARITY · SEMANTIC SPACE · VECTOR SEARCH

Word embeddings are numerical representations learned from language data. A word or token becomes a point or vector in a high-dimensional space, and items used in similar contexts tend to occupy nearby regions. The geometry makes broad semantic relationships computationally measurable.

Embeddings are one implementation of distributional semantics, not a replacement for dictionary meaning or human concepts. Static embeddings give one vector per word, while contextual embeddings produce different representations for the same form depending on the sentence. That distinction is essential for polysemous vocabulary.

This guide explains semantic vectors, dimensions, cosine similarity, nearest neighbours, static and contextual embeddings, subword tokenization, cross-lingual spaces, bias, analogy geometry and vector search. Existing eduKateSG Distributional Semantics and Hybrid Retrieval articles remain untouched.

An embedding turns patterns of word use into geometry.

2. Word embeddings

Word embeddings are numerical vector representations of words or tokens learned from language data. Similar usage patterns place words near one another in a high-dimensional space.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

3. Semantic vectors

A semantic vector is a list of numbers representing contextual or relational properties. The dimensions are usually not human-readable dictionary features; meaning is distributed across many dimensions.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

4. Static embeddings

Static embeddings assign one vector to each vocabulary item regardless of sentence context. They are useful for broad similarity but struggle with polysemy.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

5. Contextual embeddings

Contextual models produce different representations for the same token depending on surrounding words. This allows bank in a financial sentence to differ from bank beside a river.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

6. Training from co-occurrence

Many embedding methods learn from predicting or reconstructing words from context. Repeated distributional regularities shape geometry in semantic space.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

7. Dimensions

Embedding dimensionality determines how many numerical coordinates represent an item. More dimensions can capture richer patterns but do not automatically produce better meaning.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

8. Cosine similarity

Cosine similarity compares the direction of vectors. It is widely used because direction often reflects contextual similarity better than raw vector length.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

9. Nearest neighbours

A word’s nearest vectors often reveal semantic, topical, grammatical or collocational neighbours.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

10. Analogy geometry

Some embedding spaces show geometric relations that support analogies. These patterns are useful but imperfect and sensitive to model design and data.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

11. Polysemy problem

Static vectors merge several senses into one average representation. Contextual embeddings reduce this problem by representing each occurrence separately.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

12. Tokenization

Modern models often represent subword tokens rather than whole words. Vocabulary analysis must therefore distinguish human lexical items from model token units.

The key educational distinction is representation versus meaning: the vector captures statistical structure in language use, but its coordinates are not a human-readable definition of the concept.

13. Out-of-vocabulary words

Older static models can fail on unseen words. Subword-based systems construct representations from pieces, improving coverage but changing what ‘word vector’ means.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

14. Semantic similarity

Vector proximity can approximate human judgments of semantic similarity, but it may also reflect topic, syntax or shared discourse environment.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

15. Relatedness versus similarity

Coffee and cup may be close because they co-occur, even though they are not the same kind of thing. Embeddings often encode relatedness alongside similarity.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

16. Bias

Embeddings inherit statistical biases from training data. Nearby vectors can reflect stereotypes and historical inequities as well as useful semantic structure.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

17. Temporal change

Embeddings trained on different historical corpora can reveal shifts in usage and semantic neighbourhood over time.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

18. Cross-lingual embeddings

Words from different languages can be mapped into shared vector spaces, enabling translation-like comparisons and multilingual search.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

19. Sentence and document embeddings

The embedding idea can extend from words to larger units such as sentences, paragraphs and documents.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

20. Retrieval systems

Vector search uses embedding similarity to retrieve semantically related content even when exact keywords differ.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

21. Education

Embeddings can visualise semantic neighbourhoods, reveal near-neighbours and support corpus exploration, but learners still need definitions, examples and grounded knowledge.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

22. AI-era vocabulary

Large language models use contextual internal representations far richer than classical static embeddings. Understanding embeddings helps learners see how AI represents lexical similarity without assuming those vectors equal human meaning.

Embedding behaviour should be interpreted relative to model architecture, corpus, tokenization and task. Different models can place the same word in different neighbourhoods.

23. Embedding casebook — Cases 1–20

1. king

Illustrative neighbours: queen,prince,monarch,royal. Domain: status/gender/royalty. Lesson: static similarity.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

2. dog

Illustrative neighbours: cat,puppy,pet,animal. Domain: animals/pets. Lesson: category neighbourhood.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

3. doctor

Illustrative neighbours: nurse,physician,hospital,patient. Domain: healthcare. Lesson: topic + profession.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

4. car

Illustrative neighbours: vehicle,truck,automobile,road. Domain: transport. Lesson: category + context.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

5. apple

Illustrative neighbours: banana,fruit,pear,orange. Domain: fruit. Lesson: taxonomic.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

6. justice

Illustrative neighbours: fairness,law,equality,rights. Domain: abstract social. Lesson: conceptual.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

7. algorithm

Illustrative neighbours: model,data,compute,optimization. Domain: computing. Lesson: technical.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

8. teacher

Illustrative neighbours: student,classroom,school,lesson. Domain: education. Lesson: thematic.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

9. bank-finance

Illustrative neighbours: loan,money,account,interest. Domain: financial sense. Lesson: polysemy.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

10. bank-river

Illustrative neighbours: shore,river,water,stream. Domain: geographical sense. Lesson: contextual representation.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

11. cell-biology

Illustrative neighbours: nucleus,membrane,tissue,organism. Domain: biology. Lesson: technical sense.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

12. cell-prison

Illustrative neighbours: inmate,jail,guard,locked. Domain: prison. Lesson: contextual sense.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

13. light-brightness

Illustrative neighbours: dark,bright,lamp,shine. Domain: illumination. Lesson: sense cluster.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

14. light-weight

Illustrative neighbours: heavy,portable,thin,weight. Domain: mass/weight. Lesson: sense cluster.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

15. run-motion

Illustrative neighbours: walk,race,jog,sprint. Domain: motion. Lesson: verb sense.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

16. run-manage

Illustrative neighbours: operate,manage,business,company. Domain: management. Lesson: verb sense.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

17. happy

Illustrative neighbours: glad,joyful,pleased,cheerful. Domain: emotion. Lesson: near-synonyms.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

18. angry

Illustrative neighbours: furious,mad,annoyed,irritated. Domain: emotion. Lesson: degree/register.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

19. cold

Illustrative neighbours: cool,freezing,chilly,temperature. Domain: physical property. Lesson: literal.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

20. cold-personality

Illustrative neighbours: distant,unfriendly,aloof,emotionless. Domain: interpersonal. Lesson: figurative sense.

Neighbour audit: classify each neighbour as synonym, category member, topical associate, collocate or grammatical analogue. Vector proximity does not specify the relation automatically.

Cosine test: predict which neighbour should have the highest cosine similarity and explain what shared contexts would create that result.

Human comparison: ask whether people would judge the same pair as semantically similar. Differences reveal the gap between distributional geometry and conceptual judgement.

Teaching use: use the neighbourhood to generate contrasts, then verify each word’s actual meaning and register.

24. Embedding casebook — Cases 21–40

21. virus-biology

Illustrative neighbours: infection,disease,immune,pathogen. Domain: medicine. Lesson: literal domain.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

22. virus-computing

Illustrative neighbours: malware,computer,file,security. Domain: technology. Lesson: metaphorical extension.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

23. cloud-weather

Illustrative neighbours: rain,sky,storm,grey. Domain: weather. Lesson: literal.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

24. cloud-computing

Illustrative neighbours: server,storage,data,service. Domain: technology. Lesson: technical metaphor.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

25. root-plant

Illustrative neighbours: soil,tree,stem,grow. Domain: botany. Lesson: literal.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

26. root-linguistics

Illustrative neighbours: prefix,suffix,morpheme,word. Domain: linguistics. Lesson: technical metaphor.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

27. field-land

Illustrative neighbours: farm,grass,crop,soil. Domain: physical landscape. Lesson: literal.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

28. field-discipline

Illustrative neighbours: research,study,academic,domain. Domain: abstract domain. Lesson: metaphorical.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

29. model-fashion

Illustrative neighbours: runway,agency,photograph,pose. Domain: fashion. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

30. model-science

Illustrative neighbours: theory,data,predict,simulation. Domain: science. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

31. pitch-sound

Illustrative neighbours: tone,frequency,note,voice. Domain: acoustics. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

32. pitch-sport

Illustrative neighbours: throw,ball,field,baseball. Domain: sport. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

33. charge-money

Illustrative neighbours: fee,cost,pay,bill. Domain: commerce. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

34. charge-electric

Illustrative neighbours: voltage,current,electron,battery. Domain: physics. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

35. mouse-animal

Illustrative neighbours: rat,rodent,cheese,animal. Domain: biology/everyday. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

36. mouse-device

Illustrative neighbours: computer,click,pointer,keyboard. Domain: technology. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

37. spring-season

Illustrative neighbours: summer,winter,flower,warm. Domain: season. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

38. spring-water

Illustrative neighbours: source,water,river,flow. Domain: geography. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

39. capital-city

Illustrative neighbours: country,government,city,state. Domain: geopolitical. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

40. capital-finance

Illustrative neighbours: money,investment,asset,fund. Domain: finance. Lesson: sense.

Static-versus-contextual test: ask whether one vector could represent all senses adequately. If not, write two sentences that should produce different contextual representations.

Corpus test: predict how a specialist corpus would shift the vector neighbourhood. Domain data can pull technical senses closer together.

Bias test: identify a socially sensitive association the model might learn from corpus patterns and explain why statistical proximity should not be treated as normative truth.

Transfer: find another polysemous word and design contexts that should separate its embeddings.

25. Embedding casebook — Cases 41–60

41. head-body

Illustrative neighbours: face,neck,brain,hair. Domain: body. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

42. head-leader

Illustrative neighbours: chief,leader,director,manager. Domain: organisation. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

43. branch-tree

Illustrative neighbours: leaf,tree,root,wood. Domain: botany. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

44. branch-company

Illustrative neighbours: office,division,company,local. Domain: organisation. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

45. bright-light

Illustrative neighbours: shine,light,glow,luminous. Domain: perception. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

46. bright-intelligent

Illustrative neighbours: smart,clever,intelligent,student. Domain: cognitive metaphor. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

47. deep-water

Illustrative neighbours: shallow,water,ocean,depth. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

48. deep-thought

Illustrative neighbours: profound,thought,understanding,analysis. Domain: abstract. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

49. sharp-knife

Illustrative neighbours: blade,cut,point,edge. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

50. sharp-mind

Illustrative neighbours: clever,quick,intelligent,alert. Domain: cognitive. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

51. warm-temperature

Illustrative neighbours: hot,cold,heat,temperature. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

52. warm-social

Illustrative neighbours: friendly,welcome,kind,affectionate. Domain: interpersonal. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

53. platform-physical

Illustrative neighbours: stage,raised,surface,station. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

54. platform-digital

Illustrative neighbours: software,service,user,app. Domain: technology. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

55. thread-sewing

Illustrative neighbours: needle,cloth,string,sew. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

56. thread-online

Illustrative neighbours: discussion,post,message,forum. Domain: digital. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

57. stream-water

Illustrative neighbours: river,flow,water,bank. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

58. stream-media

Illustrative neighbours: video,live,audio,online. Domain: digital. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

59. frame-picture

Illustrative neighbours: photo,wall,border,wood. Domain: physical. Lesson: sense.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

60. frame-semantics

Illustrative neighbours: scene,role,meaning,background. Domain: linguistic/cognitive. Lesson: technical.

Vector-search test: imagine using this representation to retrieve documents. What relevant texts might appear even if they do not contain the exact keyword?

False-neighbour test: identify a word that might be close because of topic but is not semantically substitutable. This guards against treating similarity as synonymy.

Cross-lingual test: consider how a translation equivalent might be aligned in a shared space and where mismatched senses could create errors.

Grounding test: state which physical, sensory or cultural features are not guaranteed by textual vector proximity.

26. Reading embedding spaces responsibly

1. king

Step 1 — inspect: list nearest neighbours such as queen,prince,monarch,royal without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

2. dog

Step 1 — inspect: list nearest neighbours such as cat,puppy,pet,animal without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

3. doctor

Step 1 — inspect: list nearest neighbours such as nurse,physician,hospital,patient without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

4. car

Step 1 — inspect: list nearest neighbours such as vehicle,truck,automobile,road without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

5. apple

Step 1 — inspect: list nearest neighbours such as banana,fruit,pear,orange without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

6. justice

Step 1 — inspect: list nearest neighbours such as fairness,law,equality,rights without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

7. algorithm

Step 1 — inspect: list nearest neighbours such as model,data,compute,optimization without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

8. teacher

Step 1 — inspect: list nearest neighbours such as student,classroom,school,lesson without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

9. bank-finance

Step 1 — inspect: list nearest neighbours such as loan,money,account,interest without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

10. bank-river

Step 1 — inspect: list nearest neighbours such as shore,river,water,stream without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

11. cell-biology

Step 1 — inspect: list nearest neighbours such as nucleus,membrane,tissue,organism without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

12. cell-prison

Step 1 — inspect: list nearest neighbours such as inmate,jail,guard,locked without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

13. light-brightness

Step 1 — inspect: list nearest neighbours such as dark,bright,lamp,shine without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

14. light-weight

Step 1 — inspect: list nearest neighbours such as heavy,portable,thin,weight without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

15. run-motion

Step 1 — inspect: list nearest neighbours such as walk,race,jog,sprint without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

16. run-manage

Step 1 — inspect: list nearest neighbours such as operate,manage,business,company without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

17. happy

Step 1 — inspect: list nearest neighbours such as glad,joyful,pleased,cheerful without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

18. angry

Step 1 — inspect: list nearest neighbours such as furious,mad,annoyed,irritated without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

19. cold

Step 1 — inspect: list nearest neighbours such as cool,freezing,chilly,temperature without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

20. cold-personality

Step 1 — inspect: list nearest neighbours such as distant,unfriendly,aloof,emotionless without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

21. virus-biology

Step 1 — inspect: list nearest neighbours such as infection,disease,immune,pathogen without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

22. virus-computing

Step 1 — inspect: list nearest neighbours such as malware,computer,file,security without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

23. cloud-weather

Step 1 — inspect: list nearest neighbours such as rain,sky,storm,grey without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

24. cloud-computing

Step 1 — inspect: list nearest neighbours such as server,storage,data,service without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

25. root-plant

Step 1 — inspect: list nearest neighbours such as soil,tree,stem,grow without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

26. root-linguistics

Step 1 — inspect: list nearest neighbours such as prefix,suffix,morpheme,word without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

27. field-land

Step 1 — inspect: list nearest neighbours such as farm,grass,crop,soil without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

28. field-discipline

Step 1 — inspect: list nearest neighbours such as research,study,academic,domain without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

29. model-fashion

Step 1 — inspect: list nearest neighbours such as runway,agency,photograph,pose without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

30. model-science

Step 1 — inspect: list nearest neighbours such as theory,data,predict,simulation without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

31. pitch-sound

Step 1 — inspect: list nearest neighbours such as tone,frequency,note,voice without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

32. pitch-sport

Step 1 — inspect: list nearest neighbours such as throw,ball,field,baseball without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

33. charge-money

Step 1 — inspect: list nearest neighbours such as fee,cost,pay,bill without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

34. charge-electric

Step 1 — inspect: list nearest neighbours such as voltage,current,electron,battery without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

35. mouse-animal

Step 1 — inspect: list nearest neighbours such as rat,rodent,cheese,animal without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

36. mouse-device

Step 1 — inspect: list nearest neighbours such as computer,click,pointer,keyboard without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

37. spring-season

Step 1 — inspect: list nearest neighbours such as summer,winter,flower,warm without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

38. spring-water

Step 1 — inspect: list nearest neighbours such as source,water,river,flow without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

39. capital-city

Step 1 — inspect: list nearest neighbours such as country,government,city,state without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

40. capital-finance

Step 1 — inspect: list nearest neighbours such as money,investment,asset,fund without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

41. head-body

Step 1 — inspect: list nearest neighbours such as face,neck,brain,hair without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

42. head-leader

Step 1 — inspect: list nearest neighbours such as chief,leader,director,manager without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

43. branch-tree

Step 1 — inspect: list nearest neighbours such as leaf,tree,root,wood without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

44. branch-company

Step 1 — inspect: list nearest neighbours such as office,division,company,local without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

45. bright-light

Step 1 — inspect: list nearest neighbours such as shine,light,glow,luminous without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

46. bright-intelligent

Step 1 — inspect: list nearest neighbours such as smart,clever,intelligent,student without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

47. deep-water

Step 1 — inspect: list nearest neighbours such as shallow,water,ocean,depth without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

48. deep-thought

Step 1 — inspect: list nearest neighbours such as profound,thought,understanding,analysis without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

49. sharp-knife

Step 1 — inspect: list nearest neighbours such as blade,cut,point,edge without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

50. sharp-mind

Step 1 — inspect: list nearest neighbours such as clever,quick,intelligent,alert without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

51. warm-temperature

Step 1 — inspect: list nearest neighbours such as hot,cold,heat,temperature without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

52. warm-social

Step 1 — inspect: list nearest neighbours such as friendly,welcome,kind,affectionate without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

53. platform-physical

Step 1 — inspect: list nearest neighbours such as stage,raised,surface,station without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

54. platform-digital

Step 1 — inspect: list nearest neighbours such as software,service,user,app without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

55. thread-sewing

Step 1 — inspect: list nearest neighbours such as needle,cloth,string,sew without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

56. thread-online

Step 1 — inspect: list nearest neighbours such as discussion,post,message,forum without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

57. stream-water

Step 1 — inspect: list nearest neighbours such as river,flow,water,bank without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

58. stream-media

Step 1 — inspect: list nearest neighbours such as video,live,audio,online without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

59. frame-picture

Step 1 — inspect: list nearest neighbours such as photo,wall,border,wood without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

60. frame-semantics

Step 1 — inspect: list nearest neighbours such as scene,role,meaning,background without assuming they are synonyms.

Step 2 — label: classify the relationship behind each neighbour. This converts geometry into interpretable lexical structure.

Step 3 — contextualise: create at least two sentences for different senses or registers and compare expected contextual representations.

Step 4 — verify: check dictionaries, corpora and real-world knowledge before turning vector proximity into a vocabulary conclusion.

27. Frequently asked questions

What is a word embedding?

A numerical vector representation learned from patterns in language data.

What is a semantic vector?

A vector intended to encode contextual or semantic relationships numerically.

What is a static embedding?

One vector per vocabulary item regardless of context.

What is a contextual embedding?

A representation that changes with the surrounding sentence or discourse.

What is cosine similarity?

A measure commonly used to compare vector directions in semantic space.

Do close vectors mean synonyms?

Not necessarily. They may be topical associates, category neighbours or functionally similar words.

Why do static embeddings struggle with polysemy?

One vector averages several senses together.

What are subword embeddings?

Representations built from smaller units that help handle unseen or morphologically complex words.

Can embeddings contain bias?

Yes. They can reproduce statistical social biases from training data.

What is the simplest rule?

Treat vector proximity as evidence of shared usage, then interpret the relationship carefully.

28. Research grounding

Cambridge research describes word embeddings as a distributional-semantic approach in which neural models learn vector encodings from word co-occurrences in large corpora. Closer vectors correspond to more similar usage contexts and often more similar semantics. Cambridge work also distinguishes static prediction-based embeddings from contextualised representations such as BERT- and GPT-style states, which can encode richer context-sensitive patterns while remaining grounded in distributional learning.

• Cambridge — Word embeddings and semantic associations
• Cambridge — Static and contextual embeddings
• Cambridge — Word2vec and cosine similarity

29. eduKateSG routes

• Distributional Semantics
• Hybrid Retrieval
• Corpus Linguistics
• Vocabulary Master Router

30. Final model

Embeddings make lexical relations computable by turning context patterns into positions in a semantic space.

Their power lies in statistical structure; their limitation is the same. A vector can encode enormous regularity without becoming a complete human concept.

Vectors measure how words behave in language. Meaning still requires interpretation.

31. Final embedding interpretation lab

1. king

Neighbourhood stress test: remove the most obvious neighbour from queen,prince,monarch,royal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

2. dog

Neighbourhood stress test: remove the most obvious neighbour from cat,puppy,pet,animal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

3. doctor

Neighbourhood stress test: remove the most obvious neighbour from nurse,physician,hospital,patient and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

4. car

Neighbourhood stress test: remove the most obvious neighbour from vehicle,truck,automobile,road and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

5. apple

Neighbourhood stress test: remove the most obvious neighbour from banana,fruit,pear,orange and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

6. justice

Neighbourhood stress test: remove the most obvious neighbour from fairness,law,equality,rights and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

7. algorithm

Neighbourhood stress test: remove the most obvious neighbour from model,data,compute,optimization and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

8. teacher

Neighbourhood stress test: remove the most obvious neighbour from student,classroom,school,lesson and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

9. bank-finance

Neighbourhood stress test: remove the most obvious neighbour from loan,money,account,interest and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

10. bank-river

Neighbourhood stress test: remove the most obvious neighbour from shore,river,water,stream and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

11. cell-biology

Neighbourhood stress test: remove the most obvious neighbour from nucleus,membrane,tissue,organism and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

12. cell-prison

Neighbourhood stress test: remove the most obvious neighbour from inmate,jail,guard,locked and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

13. light-brightness

Neighbourhood stress test: remove the most obvious neighbour from dark,bright,lamp,shine and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

14. light-weight

Neighbourhood stress test: remove the most obvious neighbour from heavy,portable,thin,weight and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

15. run-motion

Neighbourhood stress test: remove the most obvious neighbour from walk,race,jog,sprint and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

16. run-manage

Neighbourhood stress test: remove the most obvious neighbour from operate,manage,business,company and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

17. happy

Neighbourhood stress test: remove the most obvious neighbour from glad,joyful,pleased,cheerful and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

18. angry

Neighbourhood stress test: remove the most obvious neighbour from furious,mad,annoyed,irritated and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

19. cold

Neighbourhood stress test: remove the most obvious neighbour from cool,freezing,chilly,temperature and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

20. cold-personality

Neighbourhood stress test: remove the most obvious neighbour from distant,unfriendly,aloof,emotionless and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

21. virus-biology

Neighbourhood stress test: remove the most obvious neighbour from infection,disease,immune,pathogen and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

22. virus-computing

Neighbourhood stress test: remove the most obvious neighbour from malware,computer,file,security and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

23. cloud-weather

Neighbourhood stress test: remove the most obvious neighbour from rain,sky,storm,grey and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

24. cloud-computing

Neighbourhood stress test: remove the most obvious neighbour from server,storage,data,service and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

25. root-plant

Neighbourhood stress test: remove the most obvious neighbour from soil,tree,stem,grow and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

26. root-linguistics

Neighbourhood stress test: remove the most obvious neighbour from prefix,suffix,morpheme,word and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

27. field-land

Neighbourhood stress test: remove the most obvious neighbour from farm,grass,crop,soil and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

28. field-discipline

Neighbourhood stress test: remove the most obvious neighbour from research,study,academic,domain and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

29. model-fashion

Neighbourhood stress test: remove the most obvious neighbour from runway,agency,photograph,pose and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

30. model-science

Neighbourhood stress test: remove the most obvious neighbour from theory,data,predict,simulation and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

31. pitch-sound

Neighbourhood stress test: remove the most obvious neighbour from tone,frequency,note,voice and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

32. pitch-sport

Neighbourhood stress test: remove the most obvious neighbour from throw,ball,field,baseball and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

33. charge-money

Neighbourhood stress test: remove the most obvious neighbour from fee,cost,pay,bill and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

34. charge-electric

Neighbourhood stress test: remove the most obvious neighbour from voltage,current,electron,battery and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

35. mouse-animal

Neighbourhood stress test: remove the most obvious neighbour from rat,rodent,cheese,animal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

36. mouse-device

Neighbourhood stress test: remove the most obvious neighbour from computer,click,pointer,keyboard and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

37. spring-season

Neighbourhood stress test: remove the most obvious neighbour from summer,winter,flower,warm and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

38. spring-water

Neighbourhood stress test: remove the most obvious neighbour from source,water,river,flow and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

39. capital-city

Neighbourhood stress test: remove the most obvious neighbour from country,government,city,state and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

40. capital-finance

Neighbourhood stress test: remove the most obvious neighbour from money,investment,asset,fund and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

41. head-body

Neighbourhood stress test: remove the most obvious neighbour from face,neck,brain,hair and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

42. head-leader

Neighbourhood stress test: remove the most obvious neighbour from chief,leader,director,manager and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

43. branch-tree

Neighbourhood stress test: remove the most obvious neighbour from leaf,tree,root,wood and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

44. branch-company

Neighbourhood stress test: remove the most obvious neighbour from office,division,company,local and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

45. bright-light

Neighbourhood stress test: remove the most obvious neighbour from shine,light,glow,luminous and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

46. bright-intelligent

Neighbourhood stress test: remove the most obvious neighbour from smart,clever,intelligent,student and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

47. deep-water

Neighbourhood stress test: remove the most obvious neighbour from shallow,water,ocean,depth and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

48. deep-thought

Neighbourhood stress test: remove the most obvious neighbour from profound,thought,understanding,analysis and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

49. sharp-knife

Neighbourhood stress test: remove the most obvious neighbour from blade,cut,point,edge and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

50. sharp-mind

Neighbourhood stress test: remove the most obvious neighbour from clever,quick,intelligent,alert and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

51. warm-temperature

Neighbourhood stress test: remove the most obvious neighbour from hot,cold,heat,temperature and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

52. warm-social

Neighbourhood stress test: remove the most obvious neighbour from friendly,welcome,kind,affectionate and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

53. platform-physical

Neighbourhood stress test: remove the most obvious neighbour from stage,raised,surface,station and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

54. platform-digital

Neighbourhood stress test: remove the most obvious neighbour from software,service,user,app and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

55. thread-sewing

Neighbourhood stress test: remove the most obvious neighbour from needle,cloth,string,sew and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

56. thread-online

Neighbourhood stress test: remove the most obvious neighbour from discussion,post,message,forum and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

57. stream-water

Neighbourhood stress test: remove the most obvious neighbour from river,flow,water,bank and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

58. stream-media

Neighbourhood stress test: remove the most obvious neighbour from video,live,audio,online and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

59. frame-picture

Neighbourhood stress test: remove the most obvious neighbour from photo,wall,border,wood and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

60. frame-semantics

Neighbourhood stress test: remove the most obvious neighbour from scene,role,meaning,background and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

61. king

Neighbourhood stress test: remove the most obvious neighbour from queen,prince,monarch,royal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

62. dog

Neighbourhood stress test: remove the most obvious neighbour from cat,puppy,pet,animal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

63. doctor

Neighbourhood stress test: remove the most obvious neighbour from nurse,physician,hospital,patient and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

64. car

Neighbourhood stress test: remove the most obvious neighbour from vehicle,truck,automobile,road and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

65. apple

Neighbourhood stress test: remove the most obvious neighbour from banana,fruit,pear,orange and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

66. justice

Neighbourhood stress test: remove the most obvious neighbour from fairness,law,equality,rights and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

67. algorithm

Neighbourhood stress test: remove the most obvious neighbour from model,data,compute,optimization and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

68. teacher

Neighbourhood stress test: remove the most obvious neighbour from student,classroom,school,lesson and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

69. bank-finance

Neighbourhood stress test: remove the most obvious neighbour from loan,money,account,interest and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

70. bank-river

Neighbourhood stress test: remove the most obvious neighbour from shore,river,water,stream and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

71. cell-biology

Neighbourhood stress test: remove the most obvious neighbour from nucleus,membrane,tissue,organism and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

72. cell-prison

Neighbourhood stress test: remove the most obvious neighbour from inmate,jail,guard,locked and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

73. light-brightness

Neighbourhood stress test: remove the most obvious neighbour from dark,bright,lamp,shine and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

74. light-weight

Neighbourhood stress test: remove the most obvious neighbour from heavy,portable,thin,weight and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

75. run-motion

Neighbourhood stress test: remove the most obvious neighbour from walk,race,jog,sprint and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

76. run-manage

Neighbourhood stress test: remove the most obvious neighbour from operate,manage,business,company and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

77. happy

Neighbourhood stress test: remove the most obvious neighbour from glad,joyful,pleased,cheerful and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

78. angry

Neighbourhood stress test: remove the most obvious neighbour from furious,mad,annoyed,irritated and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

79. cold

Neighbourhood stress test: remove the most obvious neighbour from cool,freezing,chilly,temperature and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

80. cold-personality

Neighbourhood stress test: remove the most obvious neighbour from distant,unfriendly,aloof,emotionless and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

81. virus-biology

Neighbourhood stress test: remove the most obvious neighbour from infection,disease,immune,pathogen and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

82. virus-computing

Neighbourhood stress test: remove the most obvious neighbour from malware,computer,file,security and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

83. cloud-weather

Neighbourhood stress test: remove the most obvious neighbour from rain,sky,storm,grey and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

84. cloud-computing

Neighbourhood stress test: remove the most obvious neighbour from server,storage,data,service and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

85. root-plant

Neighbourhood stress test: remove the most obvious neighbour from soil,tree,stem,grow and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

86. root-linguistics

Neighbourhood stress test: remove the most obvious neighbour from prefix,suffix,morpheme,word and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

87. field-land

Neighbourhood stress test: remove the most obvious neighbour from farm,grass,crop,soil and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

88. field-discipline

Neighbourhood stress test: remove the most obvious neighbour from research,study,academic,domain and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

89. model-fashion

Neighbourhood stress test: remove the most obvious neighbour from runway,agency,photograph,pose and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

90. model-science

Neighbourhood stress test: remove the most obvious neighbour from theory,data,predict,simulation and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

91. pitch-sound

Neighbourhood stress test: remove the most obvious neighbour from tone,frequency,note,voice and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

92. pitch-sport

Neighbourhood stress test: remove the most obvious neighbour from throw,ball,field,baseball and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

93. charge-money

Neighbourhood stress test: remove the most obvious neighbour from fee,cost,pay,bill and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

94. charge-electric

Neighbourhood stress test: remove the most obvious neighbour from voltage,current,electron,battery and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

95. mouse-animal

Neighbourhood stress test: remove the most obvious neighbour from rat,rodent,cheese,animal and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

96. mouse-device

Neighbourhood stress test: remove the most obvious neighbour from computer,click,pointer,keyboard and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

97. spring-season

Neighbourhood stress test: remove the most obvious neighbour from summer,winter,flower,warm and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

98. spring-water

Neighbourhood stress test: remove the most obvious neighbour from source,water,river,flow and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

99. capital-city

Neighbourhood stress test: remove the most obvious neighbour from country,government,city,state and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

100. capital-finance

Neighbourhood stress test: remove the most obvious neighbour from money,investment,asset,fund and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

101. head-body

Neighbourhood stress test: remove the most obvious neighbour from face,neck,brain,hair and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

102. head-leader

Neighbourhood stress test: remove the most obvious neighbour from chief,leader,director,manager and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

103. branch-tree

Neighbourhood stress test: remove the most obvious neighbour from leaf,tree,root,wood and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

104. branch-company

Neighbourhood stress test: remove the most obvious neighbour from office,division,company,local and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

105. bright-light

Neighbourhood stress test: remove the most obvious neighbour from shine,light,glow,luminous and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

106. bright-intelligent

Neighbourhood stress test: remove the most obvious neighbour from smart,clever,intelligent,student and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

107. deep-water

Neighbourhood stress test: remove the most obvious neighbour from shallow,water,ocean,depth and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

108. deep-thought

Neighbourhood stress test: remove the most obvious neighbour from profound,thought,understanding,analysis and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

109. sharp-knife

Neighbourhood stress test: remove the most obvious neighbour from blade,cut,point,edge and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

110. sharp-mind

Neighbourhood stress test: remove the most obvious neighbour from clever,quick,intelligent,alert and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

111. warm-temperature

Neighbourhood stress test: remove the most obvious neighbour from hot,cold,heat,temperature and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

112. warm-social

Neighbourhood stress test: remove the most obvious neighbour from friendly,welcome,kind,affectionate and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.

113. platform-physical

Neighbourhood stress test: remove the most obvious neighbour from stage,raised,surface,station and ask whether the remaining set still captures the same sense. If the neighbourhood collapses, the representation may be dominated by one topical association.

Contextualisation test: write two sentences that should move the token representation into different parts of semantic space. Explain which surrounding words drive the shift.

Pedagogical decision: choose one neighbour useful for synonym contrast, one for category learning and one that should be explicitly rejected as non-interchangeable.