What Is Text Analysis?
What you’ll learn: How Text Analysis is defined, what distinguishes it from related fields, and why it is becoming central to humanities and social science research
Text Analysis (TA) refers to the process of examining, processing, and interpreting unstructured data — texts — to uncover patterns, relationships, and insights using computational methods. The texts in question can be extraordinarily diverse: emails, literary novels, historical letters, newspaper articles, social media posts, court transcripts, parliamentary debates, product reviews, academic papers, or any other written record of human language and thought.
The word unstructured is used to contrast textual data with structured (tabular) data, such as a spreadsheet where each column has a predefined meaning and data type. Texts do not arrive pre-organised into rows and columns; extracting structure from them is precisely the challenge that TA addresses.
Actionable knowledge — the goal of TA — refers to insights that can be used to classify documents, track trends, extract information, map relationships, test hypotheses, or support decision-making.
The terms Text Analysis and Text Analytics are sometimes distinguished: Text Analysis is used for qualitative, interpretative, close-reading approaches, while Text Analytics refers to quantitative, computational approaches. In this tutorial series, we treat them as synonymous, encompassing any computer-assisted approach — qualitative or quantitative — to analysing text. The important distinction is not between the two labels, but between human-in-the-loop approaches (where computational methods assist human interpretation) and fully automated approaches (where algorithms process text with minimal human supervision).
Why Text Analysis Now?
The growth of interest in computational text analysis is not accidental — it reflects structural changes in how textual data is produced and stored:
Digitisation of existing records: Libraries, archives, and cultural institutions worldwide have digitised millions of historical documents, literary texts, newspapers, and government records, making them computationally accessible for the first time.
Born-digital text production: The internet has produced an unprecedented volume of natively digital text — social media, forums, news sites, emails, and more — which is already in a form amenable to computational analysis.
Methodological advances: Improvements in natural language processing, machine learning, and statistical methods have made it possible to extract meaningful information from text at scales previously unimaginable.
Computational accessibility: Tools and programming environments (including R and Python) have lowered the technical barrier to text analysis considerably. A researcher with modest programming skills can now perform analyses that previously required specialist expertise.
The Text Analysis Landscape
What you’ll learn: How Text Analysis relates to — and differs from — Corpus Linguistics, Natural Language Processing, Text Mining, and Distant Reading; and why understanding these distinctions matters for research design
Text Analysis (TA) is an umbrella term that encompasses several related fields and approaches. Understanding how these relate helps researchers choose the most appropriate framing and methods for their work.
Distant Reading
Distant Reading (DR) is an approach to analysing literary texts, pioneered by Franco Moretti (Moretti 2005, 2013). The central idea is that by analysing large corpora computationally — looking at statistical patterns across hundreds or thousands of texts — we can perceive literary and cultural trends that are invisible to any reader of individual texts. Distant reading explicitly positions itself as a complement to (not a replacement for) close reading, the traditional practice of interpreting individual texts in detail.
Moretti’s key provocation was that literature scholars could only ever closely read a small fraction of all published literature, and that this selection bias shapes what we think we know about literary history. Distant reading aims to correct this bias through systematic quantitative analysis. Representative applications include mapping genre evolution over centuries, tracking the rise and fall of specific narrative conventions, or analysing the representation of genders across thousands of novels.
Distant reading is sometimes used synonymously with Computational Literary Studies (CLS), which applies computational methods specifically to literary research questions.
Corpus Linguistics
Corpus Linguistics (CL) is a branch of linguistics that uses large, principled collections of authentic language data (corpora) to study language structure, use, variation, and change (Gries 2009; Stefanowitsch 2020; Amador Moreno 2010). CL is distinguished from armchair linguistics (constructing example sentences to illustrate grammatical points) by its commitment to empirical evidence drawn from actual language use.
Key features of corpus linguistics include:
- Frequency-based reasoning: CL emphasises that what is common matters; grammatical and lexical choices that are frequent reveal what speakers actually do, not just what they can do
- Probabilistic patterns: CL looks for tendencies and gradients, not categorical rules
- Large-scale generalisation: findings about language are grounded in thousands or millions of examples, not individual judgements
- Cross-corpus comparison: different corpora representing different varieties, genres, or time periods can be compared systematically
CL is generally more methodologically rigorous and statistically sophisticated than most TA work in the humanities — a heritage that comes from decades of developing and refining quantitative methods for language data.
Natural Language Processing
Natural Language Processing (NLP) is a field of computer science and artificial intelligence concerned with enabling computers to understand, interpret, and generate human language (Chowdhary and Chowdhary 2020; Eisenstein 2019; Indurkhya and Damerau 2010; Mitkov 2022). NLP is primarily technology development: its goal is to build systems and tools that process language automatically.
Key NLP applications include machine translation (Google Translate), speech recognition (Siri, Alexa), question answering (ChatGPT, search engines), information extraction, summarisation, and sentiment analysis. Modern NLP is dominated by large neural language models (transformer architectures such as BERT, GPT, and their successors), which are trained on enormous quantities of text and have achieved remarkable performance on many language tasks.
The relationship between NLP and TA is that of producer and consumer: NLP develops the tools (taggers, parsers, classifiers, embedding models) that TA researchers apply to their data. Many TA analyses rely on NLP-developed methods — part-of-speech tagging, named entity recognition, dependency parsing — without requiring expertise in the underlying machine learning.
Text Mining
Text Mining (TM) is a data science field focused on extracting information from large volumes of unstructured text using automated methods from NLP, machine learning, and statistics. Text Mining is decidedly data-driven and typically operates with minimal human supervision at scale. It is particularly associated with commercial applications — analysing customer reviews, social media monitoring, processing large document repositories — and with Big Data contexts where manual analysis is simply not feasible.
TM prioritises scalability and automation. The assumption is that the volume of data is so large that humans cannot be in the loop for individual documents; algorithms must classify, cluster, and extract information entirely automatically. This distinguishes TM from most TA and CL work, where human interpretation of results remains central.
How They Relate
Warning: package 'flextable' was built under R version 4.4.3
Field | Primary Domain | Data Types | Key Goal | Human in Loop? |
|---|
Text Analysis (TA) | Humanities, social science | Any text | Understand texts and phenomena | Usually |
Distant Reading (DR) | Literary / cultural studies | Literary corpora | Perceive literary/cultural patterns | Usually |
Corpus Linguistics (CL) | Linguistics | Language corpora | Understand language use | Always |
Natural Language Processing (NLP) | Computer science / AI | Any language data | Build language technology | Rarely (in production) |
Text Mining (TM) | Data science / industry | Large-scale text (web, reviews) | Extract info at scale | Rarely |
TA is broader than CL and DR — it is not limited to language study or literary texts, and it encompasses both qualitative and quantitative approaches. Compared to NLP, TA is more concerned with answering research questions than with building tools. Compared to TM, TA typically involves more interpretive engagement with individual texts and results.
Q1. A researcher wants to study whether the way newspapers represent climate change has shifted between 2000 and 2020, using a collection of 50,000 articles. Which label best describes this work?
Q2. What is the key distinction between Text Mining and most Text Analysis work in the humanities?
The Text Analysis Pipeline
What you’ll learn: The typical sequence of steps from raw text to publishable finding, and the decisions that must be made at each stage
Why it matters: Treating text analysis as a linear pipeline makes it possible to identify where methodological choices have the greatest impact on results — and where errors are most likely to occur
Text analysis projects rarely proceed in a straight line, but it is useful to think in terms of a pipeline with distinct stages:
[1] Research Question
↓
[2] Data Collection and Corpus Construction
↓
[3] Pre-processing
↓
[4] Annotation (optional)
↓
[5] Analysis
↓
[6] Interpretation and Validation
↓
[7] Reporting and Sharing
Stage 1: Research Question
Every text analysis project should begin with a clearly formulated research question. Computational methods are tools in service of research questions; they should not be applied blindly to data in the hope that something interesting emerges. Good research questions are specific enough to be testable, open enough to allow surprising findings, and tied to a body of existing literature that contextualises what findings would mean.
Common types of TA research questions include:
- Descriptive: What words, topics, or entities are most prominent in this corpus?
- Comparative: How does the language used in text collection A differ from text collection B?
- Diachronic: How has the representation of X changed over time?
- Relational: What entities co-occur or are associated in this corpus?
- Classificatory: Which documents belong to category A versus category B?
Stage 2: Data Collection and Corpus Construction
For most TA projects, the data is a corpus — a principled collection of texts assembled to address a specific research question. Corpus construction involves decisions about:
- Representativeness: Does the corpus cover the phenomenon you want to study?
- Balance: Are different varieties, genres, time periods, or authors proportionally represented?
- Size: Is the corpus large enough to support statistical inferences? (A common misconception is that bigger is always better; a small but well-designed corpus often outperforms a large but poorly designed one.)
- Sampling: How were texts selected? Is the selection principled and replicable?
- Metadata: What contextual information about each text (author, date, genre, source) needs to be recorded?
- Rights and licensing: Are the texts available for the intended use? (See the section on ethics below.)
Data can come from many sources: digitised book archives (Project Gutenberg, HathiTrust), news databases, social media APIs, government data portals, institutional repositories, or purpose-built linguistic corpora.
Stage 3: Pre-processing
Raw text is almost never in a form suitable for direct analysis. Pre-processing transforms raw text into a structured representation:
- Encoding normalisation: Ensuring consistent character encoding (UTF-8 is standard)
- Tokenisation: Splitting text into units (tokens) — words, sentences, or sub-word units
- Lowercasing: Converting all text to lowercase to treat The and the as the same word
- Punctuation and noise removal: Removing characters that are not relevant to the analysis
- Stopword removal: Removing high-frequency function words (the, and, of) for frequency-based analyses where they would dominate
- Normalisation: Stemming or lemmatisation to group inflected forms of the same word
Pre-processing decisions have a significant impact on results. Removing stopwords before a topic model, for example, will produce very different topics than keeping them. There are no universally correct choices — the right pre-processing depends on the research question and analysis method.
Stage 4: Annotation
Many analyses require adding linguistic or semantic information to the text. Annotation is the process of associating tags or labels with textual units:
- Part-of-speech (PoS) tagging: labelling each token with its grammatical category (noun, verb, adjective…)
- Named entity recognition (NER): identifying and categorising references to people, places, organisations, dates, and other real-world entities
- Dependency parsing: labelling grammatical relationships between words (subject, object, modifier…)
- Semantic tagging: assigning semantic category labels based on word meaning
- Sentiment annotation: labelling text units with polarity (positive/negative) or emotion
Annotation can be manual (expensive, slow, high quality) or automatic (fast, scalable, error-prone). Most large-scale TA relies on automated annotation tools trained on manually annotated data.
Stage 5: Analysis
The analysis stage applies the methods described in the Glossary section below — frequency analysis, concordancing, collocation analysis, keyword analysis, topic modelling, sentiment analysis, and so on — to the pre-processed (and optionally annotated) text. Method choice should be driven by the research question, not by availability or familiarity.
Stage 6: Interpretation and Validation
Computational results require interpretation. A topic model produces lists of words; it does not produce topics with labels and meanings — that is the researcher’s job. A sentiment analysis produces polarity scores; it does not explain why texts have those scores. A keyword analysis produces statistically significant words; it does not explain what those words mean in context.
Validation — checking that results are meaningful and not artefacts of pre-processing or analytical choices — is often the most time-consuming and undervalued part of text analysis. Good practice includes inspecting samples of the data at each stage, checking whether different analytical choices produce consistent results (robustness checks), and comparing computational findings against expert judgement or existing knowledge.
Stage 7: Reporting and Sharing
Reproducibility is a core value of scientific research. For TA, this means sharing code, data (where licensing permits), and detailed descriptions of pre-processing and analytical choices so that other researchers can replicate, verify, and build upon findings. R Markdown and Quarto documents are ideal vehicles for this: they combine code, data, analysis, and narrative in a single reproducible document.
Q1. A researcher collects all tweets mentioning “climate change” from 2015–2023. Before analysis, they remove all non-English tweets, convert text to lowercase, and remove URLs and (mentions?). These steps belong to which pipeline stage?
Q2. Why is interpretation (stage 6) indispensable even when computational methods are fully automated?
Core Concepts and Terminology
What you’ll learn: The key concepts, units, and methods of text analysis — explained with sufficient depth to understand not just what they are but why they matter and when to use them
Words: Types, Tokens, and Lemmas
Understanding what counts as a “word” is the foundation of text analysis, and the answer is more nuanced than it first appears.
Consider the sentence: The cat sat on the mat.
One answer is that there are six words — six groups of characters separated by spaces and punctuation. These are tokens: individual instances of word forms in the text, counted with repetition. Tokens are the raw material of frequency analysis.
A different answer is five words — five distinct character sequences, one of which (the) appears twice. These are types: the unique vocabulary items in the text, counted without repetition. The ratio of types to tokens is the type-token ratio (TTR), a measure of lexical diversity. A higher TTR indicates more varied vocabulary.
A third consideration: are cat and cats the same word? They are distinct types, but intuitively they are variants of the same underlying word. This underlying form — the base form — is the lemma. In English, the lemma of running, ran, and runs is RUN. Lemmatisation is the process of mapping word forms to their lemmas, a common pre-processing step for analyses that should treat inflected variants as instances of the same word.
A related process is stemming, which truncates word forms to a common root using rule-based algorithms (e.g., running → run, runner → runner — but also university → univers). Stemming is faster than lemmatisation but less accurate; it does not consult a dictionary or consider semantic differences.
Concept | Definition | Example |
|---|
Token | Individual word instance in text (counting with repetition) | 'the' appears twice → 2 tokens |
Type | Unique word form (counting without repetition) | 6 tokens but 5 types in 'the cat sat on the mat' |
Lemma | Base/dictionary form of a word | WALK is the lemma of walk, walks, walked, walking |
Type-Token Ratio | Types ÷ Tokens — measure of lexical diversity | 5 types / 6 tokens = 0.83 in the example sentence |
Stem | Reduced root form produced by rule-based truncation | walk, walks, walked → 'walk' (via stemming) |
Corpus (pl. Corpora)
A corpus is a machine-readable, electronically stored collection of natural language texts representing writing or speech, assembled according to explicit design criteria (Sinclair 1991). The design criteria — what texts to include, from whom, when, in what proportions — are what distinguish a corpus from a mere collection of texts.
Corpora serve as the empirical foundation of text analysis: they provide the data from which frequencies are computed, patterns extracted, and comparisons drawn. The quality of a corpus — its representativeness, balance, and documentation — directly determines the validity of analyses based on it.
There are four main types of corpora:
Monitor corpora are large, balanced collections representing a language or variety in its full breadth, continuously updated to track language change. Examples include the Corpus of Contemporary American English (COCA, 1 billion+ words) and the International Corpus of English (ICE). These are used for studying language use, testing linguistic hypotheses, and extracting authentic examples.
Learner corpora contain language produced by first or second language learners and are used to study acquisition, development, and learner-native speaker differences. Examples include the Child Language Data Exchange System (CHILDES) for L1 acquisition and the International Corpus of Learner English (ICLE) for L2 writing.
Historical or diachronic corpora contain texts from different time periods and enable the study of language change. Examples include the Penn Parsed Corpora of Historical English and The Helsinki Corpus of English Texts, both widely used in historical linguistics research.
Specialised corpora are narrow in scope, representing a specific genre, register, domain, or community. Examples include the British Academic Written English Corpus (BAWE) for academic writing and various specialised corpora of medical, legal, or scientific language.
Concordancing
Concordancing is the extraction and display of all occurrences of a word or phrase in a text, each shown with its surrounding context (Lindquist 2009). The most common display format is the Key Word In Context (KWIC) display, in which the search term is centred with a window of preceding and following words on either side — typically five to ten words in each direction. Aligning all instances of the keyword vertically makes it easy to scan for patterns in the surrounding context at a glance.
Concordancing serves multiple purposes: inspecting how a term is used (disambiguation, pragmatic function), extracting authentic examples for illustration, identifying collocational patterns, and providing the qualitative evidence that quantitative frequency analysis cannot supply. It is often the first analytical step in a text analysis workflow — a sanity check before statistical analysis, and a source of examples to explain statistical results.
Collocations
Collocations are word pairs (or groups) that co-occur together more frequently than would be expected by chance, reflecting a genuine linguistic attraction between the words (Sinclair 1991; Evert et al. 2008). Classic English collocations include Merry Christmas (not Happy Christmas), strong tea (not powerful tea), and make a decision (not do a decision).
Collocations matter because they reveal the hidden patterning of language: every word has a characteristic set of collocates that defines its contextual behaviour, and these patterns are often invisible to introspection but emerge clearly from corpus data. They are central to lexicographic work, second-language teaching, and computational models of word meaning.
Collocation strength is measured by association measures derived from contingency tables — cross-tabulations of how often two words co-occur versus how often they occur independently. The most widely used measures are:
- Mutual Information (MI): measures how much more often two words co-occur than expected by chance; sensitive to rare words
- Log-likelihood (G²): a significance test for the co-occurrence; less sensitive to rare words than MI
- phi (φ): an effect size measure derived from chi-square; captures the strength of the association
- Delta P (ΔP): a directional measure that captures the asymmetric attraction between two words (how strongly does word A predict word B, and vice versa?)
Document-Term Matrix
A Document-Term Matrix (DTM) — also called a Term-Document Matrix (TDM) — is a mathematical representation of a collection of texts in which each row represents a document, each column represents a term (word), and each cell contains the frequency of that term in that document. The DTM is the central data structure for most quantitative text analysis methods: frequency analysis, keyword analysis, topic modelling, and text classification all operate on DTMs.
DTMs are typically very sparse: most words appear in only a small fraction of documents, so the matrix contains many zeros. This sparsity is both a computational challenge and an analytical opportunity — terms that are non-zero in a document are the ones that characterise it.
A key variant is the tf-idf weighted DTM (see below), in which raw frequencies are replaced by scores that reflect how characteristic each term is of each document relative to the collection as a whole.
Frequency Analysis
Frequency Analysis is the most fundamental text analysis method: counting how often words (or other units) occur in a text or corpus. Despite its simplicity, frequency analysis is enormously informative. The most frequent words in a text reveal its vocabulary profile; comparing frequency profiles across texts reveals vocabulary differences; tracking frequency changes over time reveals diachronic trends.
Raw frequency counts are often less useful than relative frequencies (counts per 1,000 or per million words), which make it possible to compare texts of different lengths. Frequency distributions in natural language follow Zipf’s Law: a small number of words (mostly function words) account for a large proportion of all tokens, while the vast majority of types occur very rarely. This heavily skewed distribution has significant implications for statistical analysis.
Keyword Analysis
Keyword Analysis identifies words that are statistically over-represented in one text or corpus (the target corpus) compared to another (the reference corpus). A keyword is not simply a frequent word — it is a word that is disproportionately frequent, whose elevated frequency cannot be explained by chance. Keywords characterise what makes one text or collection of texts different from another.
The concept of keyness is operationalised through various statistical measures, including log-likelihood, chi-square, and tf-idf (see below). The choice of reference corpus is critical: the same text can produce very different keywords depending on whether it is compared to a general language corpus, a genre-specific corpus, or another text in the same collection.
Term Frequency–Inverse Document Frequency (tf-idf)
tf-idf is a weighting scheme that reflects how characteristic a word is of a specific document within a collection. It combines two components:
- Term frequency (tf): how often the word appears in the document (higher = more relevant to this document)
- Inverse document frequency (idf): the inverse of how many documents contain the word (higher = rarer across the collection = more distinctive)
Words that are very common across all documents (like the, and, is) get a low tf-idf score even if they appear frequently in a document, because their presence does not distinguish the document from others. Words that appear frequently in one document but rarely in others get a high tf-idf score, marking them as characteristic of that document. Tf-idf is widely used in information retrieval (search engines) and as a feature weighting method for text classification.
N-Grams
N-grams are contiguous sequences of n tokens from a text. Unigrams are single words; bigrams are two-word sequences; trigrams are three-word sequences, and so on. The sentence I really like pizza contains the bigrams I really, really like, and like pizza, and the trigrams I really like and really like pizza.
N-gram analysis extends frequency analysis from single words to multi-word sequences, capturing phraseological patterns that word-level analysis misses. N-grams are foundational in language modelling (predicting the next word given the preceding n-1 words), in stylometry (authorship attribution), in collocation analysis, and in building phrase-level features for text classification.
Part-of-Speech Tagging
Part-of-Speech (PoS) Tagging is an annotation process that assigns grammatical category labels (noun, verb, adjective, adverb, preposition, etc.) to each token in a text. PoS tags enable grammatical filtering (extract all nouns; find all adjective-noun sequences), disambiguation (distinguish bank as a noun from bank as a verb), and more sophisticated analyses that take grammatical structure into account.
The most widely used tagset for English is the Penn Treebank tagset (45 tags), though the Universal Dependencies (UD) tagset (17 coarse tags) is increasingly standard for cross-linguistic work. Modern PoS taggers achieve accuracy above 97% on standard English text; accuracy drops on social media, historical texts, and other non-standard varieties.
Named Entity Recognition
Named Entity Recognition (NER) identifies and classifies references to real-world entities in text — people, organisations, locations, dates, monetary values, and other named items. NER is foundational for information extraction: automatically populating databases from text, building knowledge graphs, answering questions about specific entities, and mapping networks of relationships.
NER systems typically use sequence labelling models (CRF, neural networks) trained on annotated data. Performance is high for well-resourced languages and standard entity categories (person, location, organisation) but degrades on specialised domains, historical text, and languages with limited training data.
Dependency Parsing
Dependency Parsing is the analysis of grammatical structure in terms of binary relationships between tokens — a head word and a dependent word, connected by a labelled arc indicating the grammatical relationship (subject, object, modifier, etc.). The result is a dependency tree rooted at the main verb of each sentence.
Dependency structures are useful for extracting subject-verb-object triples (who does what to whom), identifying nominal modification patterns (what adjectives modify which nouns), and building structured representations of sentence meaning for downstream tasks. In conjunction with NER, dependency parsing enables the extraction of relational facts: [Amazon] acquired [MGM] for [8.45 billion dollars].
Sentiment Analysis
Sentiment Analysis is a suite of computational methods for determining the polarity (positive/negative) or emotional tone (joy, anger, fear, sadness, etc.) of a word, sentence, or text. Sentiment analysis is one of the most widely applied TA methods, particularly for analysing social media, product reviews, and customer feedback.
Two main approaches dominate:
Lexicon-based methods use dictionaries in which words are annotated with polarity or emotion labels. The sentiment of a text is computed by summing or averaging the scores of its words. Common lexicons include AFINN, BING, NRC, and SentiWordNet. Lexicon-based methods are transparent and require no training data, but they cannot handle irony, negation, or domain-specific language well.
Machine learning methods train a classifier on labelled examples (texts with known polarity). These methods can learn complex patterns including negation and domain-specific cues, but require labelled training data and are less transparent.
Semantic Analysis
Semantic Analysis encompasses methods that go beyond word form to analyse word meaning. Common approaches include:
- Semantic field analysis: Grouping words by meaning category using annotated tagsets such as the UCREL Semantic Analysis System (USAS)
- Word embeddings: Representing words as dense vectors in high-dimensional space (Word2Vec, GloVe, fastText) such that semantically similar words have similar vectors; enables computation of semantic similarity and analogy
- Contextualised representations: Large language models (BERT, GPT) produce word representations that vary with context, capturing polysemy and nuance
Topic Modelling
Topic Modelling is a family of unsupervised machine learning methods that discover latent thematic structure in a collection of documents by identifying groups of words that tend to co-occur (Blei, Ng, and Jordan 2003). The most widely used algorithm is Latent Dirichlet Allocation (LDA), which assumes that each document is a mixture of topics and each topic is a mixture of words, and estimates both simultaneously from the data.
The output of a topic model is a set of topics — each represented as a probability distribution over the vocabulary — and a set of document-topic distributions showing how much each topic contributes to each document. Interpreting topics requires human judgement: the researcher must examine the top words for each topic and decide what theme they represent.
Topic models are exploratory rather than confirmatory: they are most useful for generating hypotheses about the thematic structure of a corpus rather than for testing specific hypotheses. There are two main flavours:
- Unsupervised (unseeded) topic models: The researcher specifies the number of topics k; the model finds k topics that best explain the corpus. The researcher must then interpret and label each topic.
- Supervised (seeded) topic models: The researcher provides seed terms that anchor specific topics, guiding the model toward themes of interest. More appropriate when the researcher has specific thematic hypotheses.
Network Analysis
Network Analysis visualises and quantifies relationships between entities as graphs of nodes (entities) and edges (relationships). In text analysis, networks are used to represent co-occurrence relationships between words (collocation networks), interactions between characters in literary texts (character networks), citation relationships between academic papers, and communication networks derived from emails or social media.
Key network metrics include degree (how many connections a node has), betweenness centrality (how often a node lies on the shortest path between other nodes — a measure of broker/bridge status), and clustering coefficient (how interconnected a node’s neighbours are). Visualising networks reveals structural properties of relationships that tabular data cannot capture.
Regular Expressions
Regular expressions (regex) are patterns of characters used to search, match, and manipulate text. A regular expression like \b[Ww]alk\w*\b matches any word beginning with walk or Walk — walk, walks, walked, walking, walker, Walk, Walks, etc. — regardless of where in the text it appears.
Regular expressions are indispensable for text pre-processing (finding and replacing patterns), corpus searching (extracting examples matching complex criteria), and data cleaning. They are used in virtually every text analysis pipeline. A full introduction to regular expressions in R is available here.
Document Classification
Document Classification (also called text categorisation) is the process of assigning documents to predefined categories — genres, authors, languages, sentiment classes, topics, or any other labelling scheme. Classification can be:
- Rule-based: Categories defined by explicit rules (e.g., documents containing the word invoice belong to the financial category)
- Feature-based machine learning: Documents are represented as feature vectors (word frequencies, tf-idf weights, n-gram counts) and a classifier (logistic regression, SVM, random forest) is trained on labelled examples
- Deep learning: Neural networks (LSTM, transformer models) that learn representations and classifications jointly from raw text
Applications include authorship attribution, genre classification, spam filtering, language identification, and sentiment classification.
Q1. A text contains 200 tokens but only 80 unique word forms. What is the type-token ratio, and what does it indicate?
Q2. What is the key conceptual difference between a COLLOCATION and a KEYWORD?
Q3. A topic model of 500 news articles produces 10 topics. Topic 3 has the top words: bank, loan, rate, interest, federal, mortgage, credit, reserve. What is the appropriate next step?
Epistemological and Ethical Considerations
What you’ll learn: The limitations of computational text analysis, the epistemological claims it can and cannot support, and the ethical issues raised by working with text data
Why it matters: Understanding what text analysis cannot do is as important as understanding what it can do
What Text Analysis Can and Cannot Do
Text analysis is a powerful set of methods, but it is not a magic box that produces truth from data. Several limitations deserve explicit acknowledgement:
Texts are not behaviour: Most text data is not a random sample of linguistic behaviour — it is a selection shaped by who wrote, what they wrote about, who published it, and who preserved it. Literary corpora overrepresent canonical (and often Western, male, and elite) texts. Social media data overrepresents young, English-speaking, and technology-using populations. Historical corpora preserve official and literate production and lose oral, informal, and marginalised voices.
Frequencies are not meanings: The fact that a word is frequent, or that two words co-occur, or that a topic appears in a corpus, does not tell us what these patterns mean. Meaning is always contextual, cultural, and interpretive — and assigning it to computational outputs requires the same careful scholarly judgement as any other form of textual interpretation.
Models embody assumptions: Every text analysis method embeds assumptions about language, meaning, and text. A bag-of-words model assumes that word order does not matter. A topic model assumes that documents are mixtures of topics. A sentiment lexicon encodes one community’s judgements about word valence. When these assumptions do not match the data or the research question, results can be misleading.
Criticism of Digital Humanities: Text Analysis and computational literary studies have attracted criticism for producing results that are either “banal or, if interesting, not statistically robust” — findings that are either obvious or methodologically questionable. This criticism is not without merit: many early text analysis studies applied methods uncritically, without adequate statistical rigour or connection to theoretical frameworks. The field is maturing, and researchers coming from quantitative traditions (statistics, corpus linguistics) are increasingly contributing methodological rigour. But the obligation to evaluate whether computational findings are actually new, meaningful, and methodologically sound remains a live concern.
Ethical Considerations
Copyright and intellectual property: Most texts produced after the early 20th century are under copyright, which restricts how they can be copied, distributed, and used. Research use of copyright material is governed by fair use / fair dealing provisions in different jurisdictions, but these are not unlimited. Researchers should understand the licensing terms of any data source they use.
Privacy and informed consent: Texts produced in public contexts (newspaper articles, published books) are generally available for research. Texts produced in semi-public or private contexts — social media posts, forum contributions, emails — raise more complex questions. Individuals may not expect their words to be collected, analysed, and published, even if those words appeared in a publicly accessible space.
Bias and representation: As noted above, text corpora are never neutral or representative. Computational analyses can amplify existing biases in the data, producing findings that reflect the composition of the corpus rather than anything true about the phenomena of interest.
Dual use: Some text analysis methods — surveillance of online communication, automated profiling of individuals, large-scale opinion monitoring — can be used in ways that raise serious concerns about privacy, civil liberties, and the instrumentalisation of language research.
Tools versus Scripts: The Case for Reproducible Research
What you’ll learn: Why code-based, script-driven text analysis is preferable to point-and-click tools for research purposes
Key principle: Reproducibility is not just a technical nicety — it is a scientific and ethical obligation
It is entirely possible to perform many text analysis tasks using graphical tools like AntConc (Anthony 2024) or SketchEngine (Kilgarriff et al. 2014). These tools are user-friendly, powerful, and appropriate for many purposes. So why does LADAL recommend script-based analysis in R?
Reproducibility: When analysis is performed by clicking through a graphical interface, there is no automatic record of what was done. Replicating the analysis requires redoing every step manually. When analysis is performed in an R script, every step is recorded, can be rerun with one command, and can be shared with other researchers.
Transparency: Scripts make the full analytical workflow visible — every pre-processing decision, every parameter setting, every statistical test. Point-and-click tools often obscure these decisions, making it difficult for readers to evaluate the methodological choices.
Flexibility and power: Graphical tools implement a fixed set of analyses. R provides access to hundreds of packages covering the full range of text analysis methods, which can be combined in arbitrary ways to address novel research questions.
Scalability: Running the same analysis on 100 files in R takes the same code as running it on 1 file. Doing the same in a GUI requires 100 repetitions of the same manual steps.
No vendor lock-in: Commercial tools (SketchEngine, NVivo) can change their pricing, APIs, or feature sets. Open source R packages remain available indefinitely and can be archived along with data and analysis code to ensure long-term reproducibility.
Community and error correction: R is one of the most widely used languages for statistical computing globally. A large, active community means that errors in packages are typically discovered and corrected quickly, documentation is extensive, and help is readily available.
None of this means that GUI tools have no place in research. They are valuable for exploratory work, teaching, and contexts where reproducibility requirements are lower. But for published research that claims to support generalisable conclusions, script-based analysis is the appropriate standard.
Q1. A researcher analyses 10,000 tweets from 2020 about the US presidential election and concludes that “American voters were more excited about Biden than Trump.” What is the most significant methodological problem with this conclusion?
Q2. A PhD student performs keyword analysis by clicking through a commercial text analysis tool, finds significant keywords, and writes up the results. A reviewer asks: ‘What stopword list did you use? How was the reference corpus constructed?’ Why can the student not easily answer these questions?
