A document–term matrix is a table that turns language into numerical features: each row represents a document, each column represents a word or phrase under defined processing rules, and each cell records presence, a count or a weight. This representation makes text collections searchable and comparable. It does not, by itself, preserve who said what, the order of words, negation or the truth of a claim.
Did you know that two sentences can have exactly the same word counts while saying opposite things? ‘The child chased the dog’ and ‘The dog chased the child’ contain the same individual words, but the subject and object have swapped. A basic document–term matrix can make search faster while losing that difference. Understanding this helps students see both the technological power of vocabulary and the limits of treating language as numbers.
This guide explains what document–term matrices are, how binary and count representations work, how stopwords, n-grams, TF–IDF and document frequency change the numbers, and why human context remains essential. It includes a complete original worked matrix and applications across Primary and Secondary English, Mathematics, Science, History, General Paper, translation and workplace writing, with primary references to Stanford’s information-retrieval text and scikit-learn documentation.
Editorial note: The mini-documents and calculations below are original teaching examples, not results from a corpus or data about actual students. Any algorithmic description depends on the documented implementation and its settings.
Contents: how text becomes a searchable numerical map
Open 43 chapters and worked investigations
- 1. A document–term matrix is a map of words across texts
- 2. Why vocabulary turns into data
- 3. Rows are documents under an explicit definition
- 4. Columns are defined features, not universal meanings
- 5. Binary presence versus occurrence counts
- 6. An original three-text demonstration
- 7. The same dataset as a count matrix
- 8. Document–term and term–document matrices
- 9. Tokenisation creates the starting vocabulary
- 10. Lowercasing can help and hurt
- 11. Stemming and lemmatisation change columns
- 12. Stopword filters alter the feature set
- 13. N-grams preserve short sequences
- 14. Sparse matrices store mostly zero entries
- 15. Why a matrix can support keyword search
- 16. Boolean AND and OR have precise meanings
- 17. Document frequency is a column-level question
- 18. Term frequency is a row-level feature
- 19. IDF changes the value of common terms
- 20. TF–IDF yields a weighted matrix
- 21. Different vectoriser settings change the result
- 22. A feature vocabulary can be fixed in advance
- 23. Training and evaluation must stay separate
- 24. A matrix usually loses original word order
- 25. Negation can disappear into the same feature set
- 26. Synonyms may create different columns
- 27. Homonyms may collapse into one column
- 28. A document is not always an independent source
- 29. Column totals do not establish public belief
- 30. A teacher can build a matrix by hand
- 31. Primary English: recognise topic words and function words
- 32. Secondary English: compare evidence and argument
- 33. Mathematics: cells need defined units
- 34. Science: a technical term can be rare but essential
- 35. History: named entities dominate some documents
- 36. General Paper: repeated language can be an echo
- 37. Translation: matching concepts across scripts is not automatic
- 38. School notices: formal accuracy exceeds keyword overlap
- 39. Privacy survives numerical transformation
- 40. A matrix does not know whether a word is understood
- 41. The best notebook records preprocessing
- 42. A five-minute matrix debugging routine
- 43. The final reader test: numbers return to sentences
The original three-document matrix
Fictional texts: A is ‘Readers compare evidence.’ B is ‘Evidence supports claims.’ C is ‘Readers question claims.’ The counting rule lowercases words and removes final punctuation. The example matrix uses binary presence, so repeated instances would still be represented as 1.
| Document | readers | evidence | claims | Meaning of each cell |
|---|---|---|---|---|
| A | 1 | 1 | 0 | A contains readers and evidence |
| B | 0 | 1 | 1 | B contains evidence and claims |
| C | 1 | 0 | 1 | C contains readers and claims |
| Document frequency | 2 | 2 | 2 | Each term occurs in two different documents |
A count-based matrix for these particular sentences happens to be identical, because no selected term is repeated within a single document. Add the extra sentence ‘Evidence matters’ to document B and the count for evidence in B becomes 2, while its binary presence remains 1.
1. A document–term matrix is a map of words across texts
A document–term matrix is a table or numerical array in which rows represent documents, columns represent selected terms and each cell records a defined feature of that term in that document. The simplest versions store zero or one for absence or presence; others store counts or weights.
The representation allows a computer to compare many documents efficiently. It can answer questions such as which texts mention evidence and which contain claim. But it is a compressed view of language: the numbers do not preserve all the words’ relationships, tone or truth.
2. Why vocabulary turns into data
People read whole sentences, but many search and classification systems first build simpler numerical representations. A document–term matrix converts selected features into columns that algorithms can count, compare or weight. This bridges language and computation.
For students, the important insight is that a term is not merely a dictionary item. It can also become a searchable signal. Understanding how the signal is created helps learners recognise what is visible to a machine and what might disappear when it is reduced to counts.
3. Rows are documents under an explicit definition
A document might be a short story, Science report, school notice or one paragraph. Whichever unit is chosen becomes a row in the matrix. If one essay is divided into five paragraphs, the matrix now contains five rows rather than one.
This boundary changes the values and comparisons. The researcher must decide whether the question concerns whole documents or smaller passages, then keep the same rule throughout the analysis. A row label needs to identify its source, not just a convenient number.
4. Columns are defined features, not universal meanings
A column might represent the exact word evidence, a lowercased spelling, a lemma or a multiword expression. The chosen analyser decides which terms belong to the vocabulary used by the matrix.
That means a column is a computational feature under stated processing rules. It does not automatically represent one concept. Bank can denote money or a riverbank; unless a method explicitly disambiguates senses, the same column may mix both.
5. Binary presence versus occurrence counts
In a binary matrix, a cell contains 1 if the term occurs at least once and 0 otherwise. A count matrix records how many times it occurs. One document mentioning a word twenty times has binary value 1 but count value 20.
Both can be useful. Binary representation answers whether a term is present; counts answer how often it appears. Confusing them can make a highly repetitive source seem like many independent documents or make two essays appear equally focused when their repetition differs.
6. An original three-text demonstration
Consider three invented mini-documents: A says ‘Readers compare evidence.’ B says ‘Evidence supports claims.’ C says ‘Readers question claims.’ We lower-case the words, remove end punctuation and treat each space-separated word as a token for this illustration.
Choose the three columns readers, evidence and claims. Text A receives 1,1,0; B receives 0,1,1; C receives 1,0,1. The numbers capture presence. They do not show whether a claim is sound or why a reader questioned it.
7. The same dataset as a count matrix
In the illustrative three-document collection, each selected term happens to occur no more than once per document, so binary and count values are identical. If we add another instance of evidence to document A, its count becomes 2 while its binary presence remains 1.
A student can extend this toy collection by repeating a word and observing which representation changes. The exercise clarifies the meaning of each cell before introducing larger datasets or mathematical transformations.
8. Document–term and term–document matrices
Some texts display terms as rows and documents as columns, calling the result a term–document matrix or incidence matrix. Others use documents as rows and terms as columns. These are transposes of one another when they encode the same data.
Always read the axis labels before computing sums. A row sum might mean words per document in one orientation and document frequency for a term in the other. A good spreadsheet gives unambiguous headings rather than relying on terminology alone.
9. Tokenisation creates the starting vocabulary
Before filling the matrix, a program must decide how to divide text into units. Apostrophes, hyphens, punctuation, digits and languages without routine word-separating spaces can complicate this step.
A matrix made from one tokeniser may differ from one produced by another. Even if both show the same document names, the columns could represent different units. For comparable results, document the preprocessing choices.
10. Lowercasing can help and hurt
Lowercasing can combine sentence-initial and ordinary spellings of the same word. It can also mix distinct uses such as May the month with may as a modal verb.
When the study involves permissions, months or names, a case-insensitive matrix may require additional filtering or interpretation. The same transformation that reduces superficial variation can erase a distinction needed for the question.
11. Stemming and lemmatisation change columns
A stemmer may reduce word forms to shortened character stems; a lemmatiser maps forms to dictionary-like base entries using linguistic analysis. Either operation can group multiple visible words under one feature.
This is useful when the learning objective concerns a word family, but it can combine forms whose grammatical roles differ or fail on ambiguous words. The original text should remain available so a reader can check what an aggregated feature actually represented.
12. Stopword filters alter the feature set
A vectoriser may ignore frequent function words such as the or and when building a topic-oriented representation. That can make subject terms more visible and reduce the number of columns.
The danger is losing logically important material. Not, only and unless can determine a sentence’s meaning. A filtered matrix should not be trusted to reproduce an instruction or legal condition word for word.
13. N-grams preserve short sequences
A unigram feature represents one token; a bigram feature represents a sequence of two, while an n-gram can represent a longer sequence. In scikit-learn, ngram_range controls the minimum and maximum n-gram sizes when using relevant analysers.
Features such as not approved may preserve a local distinction that separate-word counts overlook. Yet a few fixed-length phrases still cannot encode an entire argument, a distant negation or who asserted a claim. N-grams improve representation for some purposes without solving semantics completely.
14. Sparse matrices store mostly zero entries
If a corpus has thousands of documents and tens of thousands of distinct terms, most documents do not contain most terms. A dense matrix would store enormous numbers of zero values. A sparse representation records nonzero positions and associated values more efficiently.
Scikit-learn’s CountVectorizer produces a sparse document–term matrix. The efficiency matters for real systems, but students can understand the idea with a small table: keep the entries that contain information instead of writing every absence repeatedly.
15. Why a matrix can support keyword search
If the term evidence has a column, a system can find the rows in which the column is nonzero. This simple presence query is related to Boolean retrieval: documents can be selected based on the presence or absence of specified terms.
A matrix is a teaching representation of this principle, not a claim that modern search engines inspect a huge literal table on every query. Production systems commonly use inverted indexes and more sophisticated ranking methods to scale the task.
16. Boolean AND and OR have precise meanings
A query for documents containing evidence AND claims requires both features in the same document under the matching rule. OR admits documents containing either. NOT is a restriction on a defined set and must not be confused with sentence-level negation.
In the fictional three-document example, document B has both evidence and claims. That shows presence, not that B’s statement is true. A student should understand that search logic and argument logic are not interchangeable.
17. Document frequency is a column-level question
Document frequency counts how many documents contain a term. In a binary document–term matrix, it is the sum of the relevant term column when rows represent documents. For a count matrix, one first converts each nonzero entry to presence.
A document repeating a word ten times still contributes one to the document-frequency total. This is why a matrix makes the distinction between repetition and breadth visible, but only when the chosen cell meaning is clear.
18. Term frequency is a row-level feature
Term frequency concerns occurrences of a term in a particular document. In a count matrix it is the number stored at the document–term intersection. It may be transformed or normalised in later analyses.
A raw count is affected by document length and structure. A long Science report may repeat experiment far more often than a short story simply because it has more space and a specialist topic. The cell is an observation, not an immediate quality score.
19. IDF changes the value of common terms
Inverse document frequency assigns a form of retrieval weight based on how broadly a term occurs across a collection. Very common terms may offer less discrimination among documents than terms appearing in relatively few.
IDF does not measure how difficult a word is for a child or how important it is to the world. A specialist word can be rare in a general corpus yet common in a textbook. Each weighting scheme answers a retrieval question, not a complete educational one.
20. TF–IDF yields a weighted matrix
A TF–IDF vectoriser replaces or transforms raw counts with weights involving within-document use and cross-document spread. The exact details, including smoothing and normalisation, depend on the implementation.
It is tempting to call the highest-weight word the document’s ‘most important idea’. That claim is usually too strong. The feature may distinguish the text statistically while being a name, a narrow topic or a repeated term without central conceptual importance.
21. Different vectoriser settings change the result
CountVectorizer can vary whether it lowercases, uses stopwords, chooses n-gram lengths or includes terms above or below document-frequency thresholds. Each option changes what the output columns mean.
When two students get different matrices from the same texts, first compare the configurations and library versions. A discrepancy may be due to preprocessing rather than mathematical error. Reproducibility starts before the multiplication or search.
22. A feature vocabulary can be fixed in advance
Some workflows supply a predetermined set of terms, producing consistent columns across new documents. This helps compare texts against the same vocabulary or apply a model trained earlier.
The cost is that new words outside the fixed feature list may have no dedicated column. A system that appears indifferent to a new technical term may simply lack the feature. The absence of a column is not evidence that the term lacks meaning.
23. Training and evaluation must stay separate
If a model’s feature vocabulary is learned from all documents before an evaluation, information from test data may influence the representation. Depending on the learning setup, this can create leakage and overoptimistic results.
In a standard supervised-learning workflow, fit preprocessing and feature selection on the training portion and apply the learned transformation to evaluation data. This is a technical discipline with an educational parallel: do not inspect the answers before claiming that a method predicts them.
24. A matrix usually loses original word order
A simple bag-of-words matrix does not retain the order in which individual terms appeared. The two sentences ‘The dog chased the child’ and ‘The child chased the dog’ have the same word counts even though the event reverses.
This is a major limitation. A matrix may identify topic overlap without preserving who acted on whom. For comprehension, translation or safety instructions, return to the original sentences and check grammatical relationships.
25. Negation can disappear into the same feature set
A document containing approved and one containing not approved differ in meaning, but simple keyword-based matching may overemphasise the shared term. If not is removed as a stopword, the distinction may vanish in the representation.
Never use a bag-of-words matrix alone to decide whether permission was granted or a scientific claim proved. The authoritative text and complete conditions must be checked, especially when people could act on the result.
26. Synonyms may create different columns
Two texts can describe the same idea using purchase and buy. A surface-form matrix may create separate columns, making them appear unrelated at the feature level. A synonym ring or semantic model can add connections.
But synonym expansion also introduces risk because near-synonyms differ in tone, scope and grammar. A good retrieval system improves recall while preserving the intended meaning. Students should learn that a missing exact term does not necessarily imply a missing concept.
27. Homonyms may collapse into one column
The spelling bank can describe a financial institution or a river edge. A naive matrix may count both under one feature even though they refer to different concepts.
Sense-aware techniques, metadata and full-sentence reading can distinguish them. A learner’s job is not to memorise a matrix as truth, but to recognise when an apparently precise number is built on an ambiguous label.
28. A document is not always an independent source
Several rows may be copies of the same press release or template. A matrix can count them separately unless deduplication or source grouping is applied. That can make one phrase appear widespread across the document collection.
If the research claims independent corroboration, source provenance matters. Five copied reports are not necessarily five separate pieces of evidence. Word presence is not source credibility.
29. Column totals do not establish public belief
Suppose many documents mention the word fraud. A matrix can show that the term appears often in the stored texts; it cannot tell us whether authors accuse someone, deny an allegation or quote a disputed claim.
For History and media literacy, inspect quotation, evidence and speaker. Searchable language features point to passages for investigation, not verified facts about the world.
30. A teacher can build a matrix by hand
Choose four invented short passages and five target terms. Create a table with one row per passage and one column per term. Have pupils circle each appearance and enter 1 for presence or the actual count for a count matrix.
Then ask what information the numbers preserve and what disappears. The activity becomes much more meaningful when students read two sentences with the same feature counts but different meanings. Data literacy should lead back to language.
31. Primary English: recognise topic words and function words
A Primary child can compare a story about a storm with instructions for staying indoors. The two texts may share rain or door but organise the information differently.
After counting familiar vocabulary, the teacher asks the child to explain one instruction precisely. This prevents the table from becoming a list-learning exercise that neglects logical words such as before and until.
32. Secondary English: compare evidence and argument
Two persuasive paragraphs may contain the same terms evidence, claim and reason. A count matrix could suggest topic similarity without showing whether one paragraph offers valid support and the other merely repeats assertions.
Students should test each paragraph by identifying its main claim, evidence, qualifications and counterargument. The numerical representation is a map of vocabulary signals, not an essay-marking rubric.
33. Mathematics: cells need defined units
The table itself provides a useful Mathematics exercise in matrix organisation, counts and binary variables. But students should not treat each number as meaningful without naming its row, column and counting convention.
A column sum can calculate document frequency; a row sum of counts can calculate occurrences among the selected terms. Neither sum necessarily equals the total number of words in a passage because only selected features may appear.
34. Science: a technical term can be rare but essential
In a collection of Science and general reading, a word such as photosynthesis may occur in relatively few documents yet be indispensable to one subject. A weighted matrix might distinguish those specialist texts efficiently.
This is not a reason to teach the term according to its global frequency alone. Curriculum relevance, conceptual understanding and evidence from the actual lesson determine which vocabulary students need.
35. History: named entities dominate some documents
The names of people, places and events can become high-frequency features in a historical corpus. These help retrieve relevant documents but may not distinguish an author’s interpretation of the event.
A student must still examine verbs, claims and source context. Two histories mentioning the same leader can disagree sharply. A matrix establishes common vocabulary, not common historical judgement.
36. General Paper: repeated language can be an echo
A GP student may find the same phrase in many online articles and assume it reflects independent consensus. A document–term matrix would indeed show broad presence if the pages are stored separately.
Source examination may reveal they all copied one announcement. The lesson is to distinguish a feature appearing in many files from multiple independent confirmations. Good critical reading complements retrieval.
37. Translation: matching concepts across scripts is not automatic
A source and its translation may have different word counts, grammatical structures and term boundaries. A surface-level matrix in one language cannot be directly compared with another without appropriate tokenisation and mappings.
The goal is preserving who acts, what changes and what conditions apply. Cross-language retrieval can use more advanced techniques, but a simple term-count matrix is not proof of accurate translation.
38. School notices: formal accuracy exceeds keyword overlap
Two versions of an attendance notice may contain nearly identical words except may versus must. Their document–term vectors can look similar while the obligations differ.
For consequential wording, use complete-sentence comparison and the current authoritative policy. A spreadsheet of counts is not enough to approve a notice that families will rely upon.
39. Privacy survives numerical transformation
A matrix may still contain sensitive signals about individual students or rare personal experiences, especially when rows can be linked back to identifiable documents. Converting text to numbers is not automatically anonymisation.
Use authorised datasets, appropriate access controls and a minimal feature set. A public classroom demonstration can use fictional passages rather than uploading genuine pupil writing to unapproved services.
40. A matrix does not know whether a word is understood
A child may copy an unfamiliar word into an essay without understanding it, or may demonstrate deep knowledge through a common word used precisely. The matrix counts the appearance; it cannot measure the learner’s mental representation.
Vocabulary assessment should include contextual explanation, accurate use and transfer to a new situation. Computational counts can help identify patterns worth teaching but should not replace instructional judgement.
41. The best notebook records preprocessing
A responsible project log names the documents, tokeniser, case rule, vocabulary, stopword list, n-gram setting, weighting scheme and software version. It also records what interpretation the researcher was trying to support.
That transparency makes the result reproducible and prevents a chart from appearing more authoritative than its source. A well-documented small example is more educational than a huge matrix whose cells cannot be explained.
42. A five-minute matrix debugging routine
When a number looks wrong, check the original sentence, confirm how it was tokenised, inspect the exact column term and verify whether the matrix stores counts or binary values. Then check whether duplicates or filters changed the dataset.
This routine resembles debugging in coding or checking units in Mathematics. It turns surprise into a testable problem rather than a reason to distrust every computation. The outcome is a corrected observation and a better-defined question.
43. The final reader test: numbers return to sentences
Ask a learner to look at a small matrix and identify two texts that share a word. Then have the learner read both originals and explain how the word is used differently or similarly in context.
If the student understands that the matrix is a useful index into texts rather than a complete account of meaning, the lesson has succeeded. Vocabulary becomes technology when words can be found, compared and used without losing the human relationships they express.
Eight-step document–term matrix exercise
- Choose three or four original passages and specify the document boundary.
- Decide whether to count exact surface words, lowercased forms or lemmas.
- Select the terms to be used as column features.
- Build a binary presence table and explain each 0 and 1.
- Build a count table using the same documents and terms.
- Compare column sums for document frequency and individual cells for repeated occurrences.
- Read two texts with similar feature counts and identify meaning the matrix omitted.
- Write a fresh sentence and explain why a word choice must still be checked in context.
Frequently asked questions
1. What is a document–term matrix?
A numerical representation of documents and term features, usually with documents as rows and terms as columns.
2. Is a binary matrix a frequency matrix?
Not in the same sense. It records presence, whereas a count matrix records occurrences.
3. Is term–document matrix another name?
Yes, sometimes with the axes transposed. Always inspect row and column labels.
4. Why are most cells zero?
Most individual documents contain only a small fraction of the vocabulary in a large collection.
5. Can a matrix preserve grammar?
A simple bag-of-words matrix usually loses word order; n-grams add some local structure but not full understanding.
6. Does TF–IDF measure the truth of a statement?
No. It weights term features for retrieval or analysis under particular rules.
7. Can a school child build one?
Yes. A four-document presence table is a suitable introductory activity.
8. Can two opposite sentences have the same counts?
Yes, if they contain the same words in different orders or differ in relationships not represented by the features.
9. Does a matrix protect confidential data?
Not automatically. Rare features and document identifiers can still expose sensitive information.
10. What is the real vocabulary goal?
Use the table to locate language evidence, then return to full sentences and make an accurate communication decision.
Research sources and the wider vocabulary system
- Manning, Raghavan and Schütze: Introduction to Information Retrieval — official companion to the textbook explaining Boolean term–document incidence matrices, vocabulary and information retrieval.
- Scikit-learn: CountVectorizer — documentation for constructing sparse matrices of token counts and binary features.
- Scikit-learn: Feature Extraction — detailed treatment of text representations, tokenisation and weighting.
- NLTK Book: Processing Raw Text — why tokenisation choices affect the feature vocabulary.
- Vocabulary | What Is Document Frequency? — how to count the number of documents containing a term.
- Vocabulary | What Is a Stopword List? — decide which common words must be retained.
- Vocabulary | What Is a Word Token? — understand the units behind each matrix column.
- Vocabulary Learning Hub — explore vocabulary as a technology for school and life.
A document–term matrix is a powerful index into language, not a replacement for language. Once learners understand which information it keeps and discards, they can use counts to discover texts without surrendering meaning, evidence or judgement.