A multilingual LDA topic modeling platform for humanities researchers.
Lemmata is a browser-based tool that lets humanities researchers perform LDA (Latent Dirichlet Allocation) topic modeling on literary and historical texts — without writing a single line of code.
Upload your texts, choose your language, adjust parameters, and download reproducible results. Everything runs in the browser.
Supported languages: Italian, English, German, French, Spanish
No installation required. Open the platform in your browser:
- Select your language and POS filter in the sidebar.
- Upload one or more files (.txt, .pdf, .odt, .docx, .epub) or paste text directly.
- Click Run Analysis.
- Explore the results across seven tabs (Overview, Topics, Topic Map, Heatmap, Distribution, Preprocessing, Export).
- Download all outputs as a ZIP file or generate a PDF report.
| Feature | Description |
|---|---|
| Multilingual NLP | Five languages with dedicated spaCy pipelines and language-specific stopwords |
| POS filtering | Presets (content words, content + verbs, all open classes) or custom selection |
| Custom stopwords | Add domain-specific stopwords on top of the built-in lists |
| Coherence scoring | C_v coherence metric with interpretive guidance (good / fair / weak) |
| Interactive charts | Altair-based interactive visualizations with hover details |
| Topic Map | pyLDAvis visualization with graceful fallback |
| Preprocessing trace | Token-level table showing every step (original, lemma, kept/removed, reason) |
| Deterministic results | random_state=42 and batch learning ensure identical results on repeated runs |
| PDF report | Auto-generated analysis report suitable for course assignments or publications |
| ZIP export | All outputs (CSV, JSON, PNG, SVG, environment report) in one download |
git clone https://github.com/oguzkoran-max/lemmata.git
cd lemmatapython -m venv .venv
source .venv/bin/activate # Linux/macOS
.venv\Scripts\activate # Windowspip install -r requirements.txtpython -m spacy download it_core_news_sm
python -m spacy download en_core_web_sm
python -m spacy download de_core_news_sm
python -m spacy download fr_core_news_sm
python -m spacy download es_core_news_smstreamlit run src/lemmata/app.pyThe application opens at http://localhost:8501.
Lemmata follows a modular design with strict separation of concerns. The UI layer (app.py) contains no business logic.
lemmata/
├── app.py # Streamlit UI only
├── config.py # Constants, language configs, slider ranges
├── preprocessing.py # spaCy NLP pipeline, POS filtering, lemmatization, trace
├── modelling.py # LDA (scikit-learn), coherence (Gensim), corpus stats
├── visualisation.py # Charts, wordclouds, pyLDAvis (zero st.* calls)
├── file_io.py # File readers, ZIP/PDF export, environment report
├── requirements.txt # Pinned dependencies
├── ARCHITECTURE.md # 200 design decisions
├── prompts/ # Vibe coding development logs
└── tests/ # Pytest test suite
Design decisions:
- scikit-learn LDA over Gensim: deterministic output with random_state=42.
- Gensim CoherenceModel used independently for C_v evaluation.
- Language configurations reviewed and approved by a corpus linguistics specialist (Doc. Dr. Hakan Cangir).
- Fixed random state: random_state=42 across all stochastic processes.
- Batch learning: learning_method='batch' eliminates document-order effects.
- Environment report: Every analysis exports Python version, package versions, and all parameters.
- Same data + same parameters = same results. Guaranteed and verifiable.
Lemmata was developed entirely through vibe coding — a researcher with no programming expertise communicated requirements to a large language model exclusively in natural language, and the LLM generated all code. Development was carried out using Claude Code (Anthropic).
Three safeguards ensured methodological rigor:
-
Full prompt-response documentation. Every prompt and response is archived in the
prompts/directory and published as supplementary material. -
Expert validation at every decision point. Domain-specific decisions followed a human-in-the-loop model: corpus linguistics choices were reviewed by a specialist; literary interpretations were provided by a scholar. Technical code generation operated under a human-on-the-loop model: the LLM produced code, the principal investigator tested outputs.
-
Automated testing. A 120-test suite with determinism verification ensures correctness and reproducibility.
An accompanying article is in preparation for submission to Digital Scholarship in the Humanities (Oxford University Press).
@software{koran_lemmata_2026,
author = {Koran, Oğuz and Cangır, Hakan and Yücesan, Barış},
title = {Lemmata: A Multilingual LDA Topic Modeling Platform for Digital Humanities},
year = {2026},
url = {https://lemmata.app}
}Click "Cite this repository" on the GitHub page for auto-generated citation via CITATION.cff.
Contributions are welcome. See CONTRIBUTING.md.
MIT License. See LICENSE.
Lemmata is developed at Ankara University, School of Foreign Languages (Italian Language and Literature) and Faculty of Languages, History and Geography — DTCF (Italian Language and Literature).
The entire codebase was generated through LLM-assisted development using Claude and Claude Code by Anthropic. Built with spaCy, scikit-learn, Gensim, Altair, and Streamlit.