Digital Humanities uses computational methods to investigate questions about texts, history, languages, culture and evidence. It is not simply humanities students learning a coding language. Strong teaching shows how a meaningful research question determines the dataset, how technical transformations affect the evidence and how numerical patterns must return to their historical context.
Our original teaching case compares two fictional groups of community newsletters. The word ‘community’ appears fewer times in the later group, yet more often per thousand words because that group contains less text. Every count is invented, and no real historical trend is claimed. Students learn to distinguish corpus selection, transcription, tokenisation, normalisation and interpretation before drawing a conclusion.
The peer-reviewed Programming Historian lessons and DARIAH-Campus provide educational context. The DARIAH/CLARIN course registry shows the breadth of the field. In the eduKate Gold Standard approach, students should be able to demonstrate a defensible path from source to computation to scholarly explanation.
Find the digital-humanities method
Follow the invented corpus through to historical interpretation, or jump directly to a source, method or reporting challenge.
Chapter index
Question, corpus and source quality · Chapters 1–4
Count, contextualise and map · Chapters 5–10
Evaluate models and responsible use · Chapters 11–13
Reproduce, communicate and transfer · Chapters 14–16
1. Start with a humanities question, not a software menu
A fictional local-history research group wants to know whether the language used to describe a neighbourhood changed between two periods. A teacher could open a text-analysis website, upload documents and celebrate the most frequent word. A stronger Digital Humanities lesson begins by asking what the documents are, who created them, which audiences they addressed and what a change in language would mean in historical context. The computer is useful because it can reveal patterns across material; it does not decide which patterns deserve interpretation.
Our classroom collection is invented: short community newsletters from an earlier period and a later period, with different document lengths and possible digitisation errors. No real community’s views are represented by the example. Students learn to move between close reading of individual passages and computational comparison of a small corpus. The first standard is a defensible research question. Without it, a chart of frequent words can become a highly polished answer to no meaningful historical problem.
Back to chapters · Next chapter
2. Teach corpus construction as historical source selection
A corpus is a defined collection of texts, not simply whatever files happened to be easy to download. In the fictional study, one period contains long printed newsletters, while the later period contains shorter digital notices. If the earlier materials were written mainly by committee chairs and the later ones by many participants, a difference in vocabulary may reflect creator roles as much as social change. Students should list inclusion criteria before counting.
Give learners a source table with creator, approximate date, document type, preservation status and known gaps. Ask what the collection represents and who is missing. A word-frequency result describes the selected documents under a particular processing method; it is not a direct measure of how every resident spoke. This teaches the humanistic meaning of sampling: digital abundance does not automatically create representative history.
Back to chapters · Next chapter
3. Preserve provenance from document to dataset
Turning a page into a row of text changes the form in which scholars encounter it. A printed newsletter includes layout, typography, advertisements, marginal notes and other features that may matter. A transcription should therefore link back to the source image or descriptive record and note whether the text was manually typed or automatically recognised. Without provenance, an unexpected phrase cannot be checked against the underlying document.
Ask students to keep a fictional document ID on every text and create a correspondence between file, date and source note. If one page was missing when scanned, the dataset should reflect that limitation rather than silently present a complete newsletter. This bridges Digital Humanities with Archival Studies: computation becomes intellectually responsible when transformations remain traceable to sources and the researcher can return to the material evidence.
Back to chapters · Next chapter
4. Explain text recognition errors as research errors
Automatic text recognition can mistake characters, especially in difficult fonts, damaged pages or poor scans. In the invented corpus, the word ‘river’ appears once in an image but is transcribed as ‘r1ver’. A naive count may therefore underestimate the term. For a very small corpus, the class can compare a source image and transcript; for larger corpora, sampling and error assessment become important.
Teach learners to distinguish scanning, recognition and subsequent correction. Do not ask them to treat machine output as authoritative because it looks like ordinary text. A correction log should state what was altered and why, while preserving access to the earlier version. DARIAH-Campus provides introductory resources on automatic text recognition and digital source criticism. The skill is not merely fixing spelling: it is understanding how a technical transformation changes the evidence used to make a historical claim.
Back to chapters · Next chapter
5. Define tokens and words before counting
A tokenisation rule determines how text is divided into units for analysis. Is ‘neighbourhood-based’ one token or two? Are possessives separated? Are uppercase and lowercase words considered identical? Choices that look minor can change reported counts, especially in multilingual material. A student who simply presses a Count button may not know what was counted.
Give the class three fictional sentences and ask them to write an explicit procedure: convert to lowercase, separate common punctuation, retain or split hyphenated forms according to a stated rule, and count only the chosen token class. Discuss how that rule can disadvantage some languages or text types. The Programming Historian’s word-frequency lesson provides technical examples, while this classroom exercise emphasises understanding the preprocessing decisions before relying on any software.
Back to chapters · Next chapter
6. Normalise counts before comparing unequal documents
In our fictional records, the term ‘community’ appears twenty times in an earlier group containing 1,000 counted words and fifteen times in a later group containing 500 words. Looking only at counts suggests the term became less common. Yet the earlier rate is 20 per 1,000 words, while the later rate is 30 per 1,000 words. The direction reverses after normalising by corpus size.
Students should state the denominator and what their chosen rate represents. The later corpus is small, and different document genres or creators may explain the pattern. No exact trend in real language has been established. This one calculation shows why Digital Humanities needs Statistics and historical interpretation together. Without normalisation, a persuasive graph can mislead simply because one period has more surviving pages.
Back to chapters · Next chapter
7. Move from frequency to usage in context
Even a correctly normalised frequency does not tell students what a term means in every occurrence. ‘Community’ might refer to a neighbourhood, a school group, online participants or a government programme. The same word can be used positively, critically or neutrally. A count alone cannot establish the author’s stance. Teachers should return to sample passages and examine the surrounding sentences.
A keyword-in-context display can show a short window around each occurrence, allowing learners to compare usage rather than relying on a bare total. Programming Historian’s keywords-in-context lesson offers a technical method. In a small classroom, index cards with source ID, quotation and immediate context accomplish the same conceptual task. The standard is knowing when numerical patterns require human reading to become meaningful.
Back to chapters · Next chapter
8. Teach metadata as part of the argument
A table of documents may include title, creator, time period, language, document type and the rights under which it may be analysed. These are not mere housekeeping fields. If later notices are mostly shorter or use a different writing style, metadata helps a researcher test whether apparent language change follows a genre shift. If dates are missing, pupils must decide whether to exclude the source or report an uncertain range.
An introductory class can group the invented records by genre and calculate separate rates. Does ‘community’ change within newsletters as well as across all documents? If not, the broad result may reflect changing collection composition. Such checks bring source criticism into computational work. A tidy spreadsheet without trustworthy metadata cannot support a confident historical explanation.
Back to chapters · Next chapter
9. Introduce mapping with uncertainty and geography
Digital Humanities can connect texts and places through maps, but historic place names, changing boundaries and vague references require care. A newsletter may say ‘near the old market’ without coordinates. Placing a dot at the modern shopping centre could be a speculative geolocation, not a sourced fact. Map interfaces often make approximate points look exact.
Give students a fictional set of place references labelled ‘precise address’, ‘approximate neighbourhood’ and ‘unresolved’. They may map the first and represent the second as a broader area while keeping the third unplaced. Ask how a missing or uncertain location affects the interpretation. A more sophisticated map should make uncertainty visible rather than bury it behind precise-looking pins.
Back to chapters · Next chapter
10. Teach network graphs without inventing social closeness
A network visualisation might connect people, organisations or publications based on documented correspondence. An edge needs a clear meaning: two people exchanged letters, co-authored a text or attended the same meeting. These are different relationships. A colourful network diagram can imply friendship or influence even when it shows only co-occurrence in a document.
Students can build a five-node fictional network from a small correspondence log and label exactly what qualifies as a link. Then consider missing archives: an unpreserved letter can remove a visible connection without proving the relationship never existed. The teacher should ask what the diagram explains and what remains uncertain. The value of computation lies in inspectable patterns, not in pretending every line is a complete social history.
Back to chapters · Next chapter
11. Use automated classification as a hypothesis, not a verdict
Machine-learning methods can classify texts by topic, sentiment or style, but their outputs depend on training data, preprocessing and categories chosen by people. A programme labelling a historical complaint as ‘negative sentiment’ has not discovered the full social meaning or moral legitimacy of the complaint. It has assigned a category under a model. Older spelling, irony and multilingual usage may reduce reliability.
Show pupils a handful of fictional passages and hypothetical automated labels, including an obvious misclassification. Ask what the error reveals about the task definition and whether the output should be used to support a historical claim. Students need not train a large model to learn responsible algorithmic criticism. They should be able to describe what the tool predicts, where its evidence comes from, and what human reading and evaluation remain essential.
Back to chapters · Next chapter
12. Check patterns against alternative explanations
An increase in a word’s relative frequency may reflect a genuine thematic shift, changes in available sources, copying errors, genre, audience or the definitions used for counting. Before writing a conclusion, ask students to propose at least two competing explanations and identify which source checks would separate them.
For the fictional corpus, a researcher might compare only newsletters written for the same audience, sample the recognition errors and read examples of ‘community’ in context. If the pattern disappears within matched genres, the original historical claim needs revision. That does not mean the whole project failed. It means computation successfully raised a question and humanistic reasoning improved the interpretation. The gold standard values explicit uncertainty over spectacular unsupported certainty.
Back to chapters · Next chapter
13. Teach copyright, personal data and ethical reuse
Digitised material may remain protected by copyright, donor terms or privacy requirements. Availability on a website does not automatically authorise wholesale downloading, redistribution or publication of personal details. The classroom should use public-domain, appropriately licensed or clearly invented documents for practice, following institutional policies. Sensitive records need special protection even when technically accessible.
Ask learners to document source URLs, access conditions, creator credit and any restrictions on derivative datasets. A word-frequency table may be acceptable to publish in some contexts where the full underlying texts are not; the precise answer depends on rights and policy, not a general declaration that academic work is exempt. Responsible digital-humanities teaching includes the legal and ethical history of the material as part of the research design.
Back to chapters · Next chapter
14. Make computational work reproducible without worshipping code
A reproducible digital project records input sources, preprocessing rules, tool or script versions, analysis choices and output files. Another reader should be able to understand how the chart was produced even if they do not use the same software. A classroom can accomplish this with a clearly labelled spreadsheet and a plain-language method before moving to programming.
Students may write a tiny script to count words, but typing code is not itself evidence of a sound historical argument. A well-documented manual calculation can be more convincing than an opaque model trained on unknown data. The lesson should distinguish reproducibility of the computational result from truth of the historical interpretation. Re-running an error consistently does not make the original source choice appropriate.
Back to chapters · Next chapter
15. Design visualisations that carry context
A line chart showing two time periods may imply a steady trend across the years between them, even when no intervening documents were examined. A two-bar comparison with clearly labelled dates and corpus sizes may be more honest. A cloud of prominent words can attract attention while concealing denominators and context. The teacher should ask which visual form represents the available evidence rather than which is most eye-catching.
In the fictional example, show both raw counts and normalised rates. Label the corpus periods, word totals and chosen tokenisation procedure. Include a note that these are invented practice texts. Students can compare a misleading chart with a corrected one and explain the change. Visualisation becomes a form of scholarly argument, and its design choices belong under the same standards of evidence as the analysis itself.
Back to chapters · Next chapter
16. Prove transfer through a different historical source
The final task changes from newsletters to a fictional collection of theatre programmes. The researcher asks whether performances increasingly featured local-language titles across two periods. Students must define the question, build a defensible corpus, identify missing programmes, standardise metadata and specify how multilingual titles will be counted. Simply copying the word-frequency method would not be enough without adapting it to this new material.
An effective response ends with a cautious interpretive statement and an audit trail: which records were included, how terms were classified, what alternative explanations remain and which findings require close reading. The gold standard of teaching Digital Humanities is the ability to combine computational leverage with the traditions of source criticism, contextual interpretation and responsible publication. Technology makes patterns visible; scholarship decides what they can mean.
Worked text-analysis problem: when raw counts reverse the story
The two corpora below are fictional. Counts represent the exact tokenisation procedure chosen for the example, not measured use in a real community. The word totals differ, so direct raw-count comparison is misleading.
| Measure | Earlier invented newsletter group | Later invented notice group |
|---|---|---|
| Counted words | 1,000 | 500 |
| Raw occurrences of “community” | 20 | 15 |
| Occurrences per 1,000 words | 20 ÷ 1,000 × 1,000 = 20 | 15 ÷ 500 × 1,000 = 30 |
| Genre and authorship | Long printed newsletters, limited writers | Shorter digital notices, broader contributors |
| Permitted conclusion | The normalised count is higher in these later selected texts; historical cause remains unresolved | |
A student who reports ‘community became less important’ from the raw counts has ignored differing corpus sizes. Another who reports ‘community consciousness rose by 50%’ from the normalised numbers has overinterpreted usage and failed to account for genre and context. A stronger report explains that the relative term frequency changes from 20 to 30 per thousand counted words, then asks how source selection and meaning should be checked.
A four-lesson source-to-method sequence
- Lesson 1 — Source design: frame the humanities question and document corpus inclusion, missing material and provenance.
- Lesson 2 — Process: identify transcription errors, define tokenisation and produce verified counts.
- Lesson 3 — Interpret: normalise rates, read words in context and compare metadata or alternative explanations.
- Lesson 4 — Transfer: analyse a different fictional cultural dataset, choose an appropriate digital method and write a bounded conclusion.
Frequently asked Digital Humanities questions
Is Digital Humanities just using computers to make charts?
No. It connects computational methods to historical, linguistic, literary and cultural questions through careful source selection and interpretation.
Does more text automatically mean stronger evidence?
No. Coverage, provenance, genre, missingness, transcription quality and the relationship between the corpus and the question matter.
Why might OCR change the result?
Recognition errors can change or omit searchable words. The nature and frequency of errors should be assessed before relying on counts.
Does a frequency change prove a historical change in belief?
Not by itself. The same word can have different meanings, while corpus composition and context may explain the observed difference.
Must every learner be an expert programmer?
No. The goal is to understand the transformation and interpretation of evidence; the appropriate technical depth depends on the course and learning objective.
Verified Digital Humanities learning sources
- Programming Historian: Lesson Directory
- Programming Historian: Counting Word Frequencies
- Programming Historian: Keywords in Context
- DARIAH-Campus: Open Training Resources
- DARIAH/CLARIN Digital Humanities Course Registry
Connect with eduKate’s teaching ecosystem
- Teaching History | Sources and Explanation
- Teaching Linguistics | Structure and Evidence
- Teaching Statistics | Variation and Inference
- Teaching Archival Studies | Provenance and Digital Records
- Teaching Library and Information Science | Metadata and Access
- The Gold Standard master hub
The gold standard of teaching Digital Humanities is a learner who can make a computational finding understandable and auditable, then explain why it matters within the human history that produced the evidence. The computer can count, sort and visualise; it cannot absolve the researcher of asking whether the result is true, representative or ethically usable.