NLP in the humanities: language processing in textual research

A corpus of 50,000 letters from the 19th century. A dictionary of 40 different spellings of the same surname. Three terabytes of digitized manuscripts without structured metadata. These are the real scales historians and literary scholars face today in archives. Manual review of such collections takes years. NLP tools can perform preliminary selection, annotation, and classification in hours—but it’s the word “preliminary” that captures the essence of this technology: it opens the door to the material, not walks through it for the researcher.

What NLP actually does in humanities research#

Natural language processing is a set of computational techniques that allow a computer to “read” text in a statistical sense: recognizing proper names, classifying sentiments, detecting topics, and mapping relationships between concepts.

In humanities practice, several techniques are most commonly used:

Named Entity Recognition (NER). The model identifies people, places, dates, and institutions in raw text. For a historian processing thousands of court records, this is the first step toward building an index of persons and events without manual coding.

Topic modeling. Algorithms like LDA or its newer computational counterparts group documents by co-occurring words. In literary studies, this allows mapping how discourse topics shift between decades or genres in a corpus too large for close reading.

Sentiment analysis. Models classify the emotional tone of text. Useful for comparing the tone of press from different political orientations in the same historical period—but unreliable with irony, euphemisms, and archaic language.

Embeddings and semantic search. Language models convert text fragments into numerical vectors, enabling the discovery of semantically similar passages even without shared keywords. This is particularly valuable for multilingual corpora or texts with variant historical spelling.

How this works in research practice#

At Cashcrown, we work with organizations that have large textual resources and want to make them searchable or quantitatively analyzable. The workflow is repeatable:

First comes corpus preparation: OCR of digitized documents, encoding normalization, removal of scanning artifacts. This stage determines the quality of everything that follows. Garbage in, garbage out.

Next, preliminary classification by the model. A RAG system or classifier identifies candidates for further analysis, groups documents thematically, and extracts entities. Results take the form of a list with assigned confidence weights.

Then, verification by an expert. The researcher reviews a sample of results, assesses whether the classification makes substantive sense, and corrects labels where the model erred. This is the point where domain knowledge is irreplaceable. The model doesn’t know that the word “wiktor” in a 19th-century Polish text could be a first name or an award title. The historian does.

Finally, iteration and quantitative analysis: the corrected model processes the rest of the corpus, and the researcher analyzes aggregated statistical results. Interpretation remains human.

Where the model regularly fails#

This section is more important than the list of capabilities because it determines when to trust results and when not to.

Irony and negation. Sentiment analysis fails with texts that say the opposite of their surface meaning. A sermon that “praises” dishonest merchants to shame them will be classified by the model as positive toward dishonesty.

Historical context and lexical shifts. Words change meaning over time. “Liberal” in 1850 means something different than in 2020. A model trained on contemporary texts will carry modern connotations into historical usage, potentially distorting topic modeling or sentiment classification results.

Hallucinations in fact extraction. When asking an LLM to summarize or extract data from historical text, the model may fill gaps with its own “knowledge” instead of admitting the information is missing. In humanities research, every claim must be verifiable by reference to the source. Automatic extraction without human verification is unacceptable here.

Low-representation languages and dialects. Models trained primarily on English perform significantly worse with texts in regional languages, dialects, and historical scripts. A researcher working with Old Polish or Gothic script must check the model’s quality on their own sample before scaling.

Ethics and research integrity#

Using NLP in the humanities raises questions that methodology must address directly.

Representation bias. Digitized archives are incomplete. Texts written by women, lower-class individuals, linguistic or religious minorities are historically underrepresented. A model trained or applied to such a corpus will reproduce this skewed representation as a result. The researcher must understand this and interpret quantitative results with this awareness.

Methodological transparency. An increasing number of academic journals require disclosure of which analysis stages were assisted by AI systems and what tools were used. Hiding this fact violates reproducibility standards. Good practice is to treat prompt logs and model outputs as part of the methodological documentation.

Human oversight as a principle, not an exception. A model’s output is a hypothesis for verification, not a conclusion to cite. Every claim that makes it into a publication must be confirmed by an expert based on primary sources. This requirement isn’t bureaucratic red tape—it’s a guarantee that science remains falsifiable.

Practical entry point#

The most common mistake in the first NLP implementation in a research project: applying the model to the entire corpus without validation on a sample.

Recommended workflow:

1. Randomly select 100-200 documents from the corpus for manual annotation by an expert.
2. Run the model on the same sample and compare results.
3. Measure precision and recall for categories relevant to the project.
4. Researcher’s decision: Is the quality sufficient for application to the entire corpus? If not, what needs correction (fine-tuning, model change, instruction refinement)?

Only after this calibration does scaling make sense. Skipping this step leads to discovering months later that the model made systematic errors invalidating a significant portion of results.

FAQ#

Can NLP be used for texts in historical Polish, e.g., from the 16th-18th centuries?#

Yes, but with caution. Models trained on contemporary Polish have significant limitations with archaic texts: different inflectional forms, Latin insertions, lack of orthographic normalization, and phonetic spelling instead of modern. The practical approach is to use multilingual models with manual verification on a sample and, where possible, leverage historical corpora as fine-tuning data. Results should be treated as preliminary indications, not certain data.

How large must a corpus be for NLP to make sense?#

There’s no hard threshold. With a few dozen documents, manual analysis is faster and more accurate. NLP becomes cost-effective when the corpus numbers in the hundreds or thousands, and the research question concerns large-scale repeatable patterns. For small collections, using a model to assist close reading is more valuable than quantitative analysis.

Can NLP analysis results be cited directly in a scientific article?#

No, not as standalone claims. A model’s output is a statistical observation that requires interpretation and verification by an expert based on sources. In the methodology section, you can describe that a model was used for preliminary classification or extraction, providing parameters and tools. The scientific claim itself must be based on verifiable evidence, not algorithmic output.

What are the risks of using general language models instead of specialized ones?#

General models (trained on broad internet corpora) don’t know domain-specific vocabulary, genre conventions, or historical context. They may assign semantically plausible but substantively inadequate labels. Specialized models or models with additional domain context (via RAG architecture) perform better with specialized texts. Calibration on your own sample is the only way to check which model is good enough for a specific project.

Does NLP violate privacy when analyzing texts containing personal data?#

It depends on the nature of the texts and processing method. Analyzing published historical materials rarely raises legal issues. Processing documents containing personal data of living individuals falls under GDPR and requires a legal basis assessment, often with pseudonymization before transfer to an external AI system. Confidential data should only be processed through local systems or with a full data processing agreement.

Related topics: scientists with AI achieve more, limits of AI autonomy in science, the role of humans in the loop, how to limit hallucinations, the black box problem.