Text as Data Approach
The Text as Data approach treats the written record of politics as evidence. The lab does conventional text analysis of topics, sentiment and frames in news coverage, laws, policies and official documents. It is even more interested in what texts reveal about who holds power. Using named entity recognition with large language models, the lab extracts people, positions and organizations along several dimensions at once, including their dates, places and relationships, and builds network, organizational and statistical datasets from them. Each step is validated so that results can be checked and reproduced.
Research Streams
Named Entity Recognition with LLM
Using large language models to extract precisely defined people, organizations, events, attributes, and relationships from text.
Frame Analysis with LLM
Using large language models to identify how texts define problems, causes, values, actors, and possible responses.
Sentiment Analysis with LLM
Moving from document-level sentiment scores to contextual evaluations expressed by particular subjects toward objects.
Semantic Network Analysis of Latent Topics
Representing latent topics and their co-occurrence as semantic networks for comparative discourse analysis.
Targets of Text Analysis
Text analysis can examine topics, sentiment, frames, conversation, discourse, named entities, and the relationships expressed among them. The target determines the unit of analysis, coding instructions, and validation strategy.
From Unigrams to LLMs
Computational text analysis has moved from dictionaries and keywords to statistically generated features such as TF-IDF, bags of words and latent topics, machine-learning classifiers, embeddings and transformers, and now large language models. A structured comparison table may be added after its dimensions are approved.
Large language models support more specific and relational coding of topics, sentiment, frames, conversation, discourse, and named entities. Instead of scoring only an entire document or paragraph, researchers can ask how a particular subject evaluates a particular object within a defined context.
Generating Data with LLM-NER
Named entity recognition once focused mainly on extracting broad categories such as people, organizations, places, events, and themes. LLM-assisted NER can target more precise combinations—for example, politicians who contested particular offices in specified elections—and can extract relationships alongside entities.
Challenges of LLM in Text Analysis
LLM-assisted coding raises provisional concerns about reliability across repeated runs, sensitivity to prompts and model versions, inconsistent handling of ambiguous cases, bias, missing or fabricated outputs, reproducibility, cost, and protection of source data. This filler text will be revised later.
Emerging Solutions
The lab is exploring randomized validation subsets; AI-AI, AI-human, and human-human intercoder reliability; repeated coding and adjudication of disagreements; versioned prompts and model settings; structured output validation; stratified error analysis; and LLM extraction of relevant excerpts for human interpretation alongside quantitative results. This provisional list is editable in Django Admin.
List of Resources
This section will link to a crowd-sourced Google Doc and other practical resources for collecting, coding, validating, and interpreting text as data.
Lab Connections
- Publications
- 8 publications →
- Research Streams
- 4 research streams →
- Training
- 6 training modules →
- Relevant Literature
- See literature for this analytic approach in Zotero.