Six-minute primer
Text as Data in Six Moves
Computational text analysis is a chain of research decisions. Each decision makes some evidence easier to see and pushes other evidence into the background.
Imagine a small collection of speeches. You want to know how different speakers describe economic responsibility. A computer cannot begin with that substantive question. It needs a sequence of more concrete instructions.
1. Ask a bounded question
“What do these speeches mean?” is too broad. “Which economic terms are especially characteristic of each speaker, and how are those terms used in context?” is answerable with a small corpus and a limited method. A good question identifies the texts, the comparison, and the kind of evidence needed.
2. Define the corpus
A corpus is a selection of texts. Who produced them? When and under what conditions? What is missing? A pattern may describe the archive or collection procedure instead of the wider world. Before analysis, record the source and dates. Also record the inclusion rules, language, and known gaps.
3. Turn text into units
Software usually divides text into units called tokens. These are often words or punctuation marks. Decisions about lowercase text, stopwords, spelling, and Korean morphology change which tokens remain. Punctuation and multi-word phrases also matter. Save the settings and compare the processed text with the original.
4. Represent and compare
Raw counts show what occurs often. TF-IDF gives more weight to words that are common in one document but less common across the corpus. Embeddings place documents in a numerical space based on patterns learned from language data. Each representation answers a different comparison question. Each also leaves out some features of meaning.
5. Find a pattern and return to the text
A cluster, sentiment score, prediction, or topic is an invitation to inspect the documents. Open surprising cases. Read passages that support and challenge the apparent pattern. Ask whether the output follows the research question, a preprocessing decision, an unusual document, or a limitation of the model.
6. Report a claim with boundaries
A useful result states what was compared, how it was produced, what textual evidence supports it, and what it cannot establish. Keep the code or Orange workflow, source information, outputs, and short explanation together so another person can understand and reproduce the path.
Compare the computational overview with the underlying text. The overview helps you decide where to read. Close reading helps you decide whether the overview is credible.
If you want more
- Analyzing Documents with TF-IDF explains why TF-IDF can be more useful than raw frequency. The Python exercises are optional.
- Text Mining with R provides an open, code-based reference for students working in R.
- Orange’s Text Mining widget catalog documents the current behavior of each Orange component.