Nothing happens on Start Analysis
Check that a corpus is loaded. The counters in the header should be above zero and that at least one analysis is ticked.
Web application
TALOS Text Analyzer:
AI Corpus Analysis
A browser workbench for analysing corpora of Ancient and Modern Greek: frequency and collocation statistics, co-occurrence networks, and LLM-backed entity recognition, part-of-speech tagging, sentiment and topic modelling.
Overview
The Text Analyzer is a corpus workbench that runs entirely in the browser. You give it a typed text, uploaded or fetched from a URL, choose which analyses to run, and it returns statistics, networks and annotations over the whole corpus.
It is built for Greek. The interface asks for a text in Ancient or Modern Greek, and the models behind it are selected with that in mind, which is what separates it from a general-purpose NLP demo.
What it does
- Corpus statistics: word and sentence counts, unique words, reading time, file size
- Frequency, collocation and keyword analysis
- Co-occurrence networks between terms
- Named entity recognition and part-of-speech tagging, run through a large language model
- Sentiment analysis and topic modelling
- A pattern builder for extracting sequences by part-of-speech
Getting started
Nothing to install. Open the tool, give it a text, and press Start Analysis.
- Provide a corpus through Text Input, File Upload or URL Import or press Load Sample to try it with a supplied text.
- Under Analysis Parameters, tick the analyses you want.
- Press Start Analysis. Progress is reported as items processed, and Stop Analysis halts a run in progress.
- Read the results, then Export Results when you are satisfied.
The header keeps a live count of words, sentences, unique words, estimated reading time and file size, so you can sanity-check that the text loaded as expected before spending time on an analysis.
Providing a corpus
Text Input
Paste directly into the editor. Best for short passages and for testing a pattern before running it over a full corpus.
File Upload
Drag files onto the drop zone or click to browse. Accepted formats:
| Plain text | TXT |
|---|---|
| Documents | PDF, DOC, DOCX |
| Structured | RDF, XML |
RDF and XML support means a corpus can come straight from the published datasets rather than being flattened to plain text first.
URL Import
Give a web address and press Fetch. The Extract text content only option strips the markup and keeps the prose, which is what you want when pointing at an ordinary web page rather than a data file.
Basic analysis
Four statistical analyses that run without a language model. They are fast and deterministic, and are the right starting point for an unfamiliar corpus.
- Word Frequency: how often each form occurs.
- Co-occurrence Network: which terms appear near one another, drawn as a network.
- Collocations: word pairs that occur together more often than chance would predict.
- Keyword List: the terms most characteristic of this corpus.
Advanced analysis
Four analyses that call a large language model. They are slower and, unlike the basic analyses, may vary between runs.
- Named Entity Recognition: people, places and organisations.
- Part-of-Speech Tagging: grammatical category per token, and the basis for pattern extraction.
- Sentiment Analysis.
- Topic Modeling: recurring themes across the corpus.
Check the output: these are model predictions, not deterministic parses. For Ancient Greek in particular, verify a sample by hand before building an argument on the results.
Pattern extraction
The pattern builder finds sequences by grammatical category rather than by word. You constrain up to five positions, choosing a part-of-speech tag for each and leaving the rest free, then press Extract Pattern.
ADJ + NOUN | “beautiful house” |
|---|---|
DET + ADJ + NOUN | “the red car” |
NOUN + ADP + NOUN | “cup of coffee” |
ADJ + ADJ + NOUN | “big red balloon” |
VERB + DET + NOUN | “read the book” |
Pattern extraction depends on part-of-speech tagging, so run that first. This is the feature that turns the tool from a summariser into something you can ask a specific question with.
Model settings
Under Advanced Settings you choose which model runs the advanced analyses, and how it behaves.
| LLaMA 3.3 70B | Largest and slowest; best quality |
|---|---|
| LLaMA 3.1 8B | Smaller and faster; good for iterating |
| Kimi K2 | Alternative model |
| Max Tokens | Caps the length of each model response |
| Temperature | Defaults to 0.2: low, so output stays consistent |
Leave the temperature low for analysis work. Raising it makes the model more inventive, which is the opposite of what annotation needs. Reset Defaults restores the original settings.
Exporting results
Export Results in the header downloads the analysis once a run has completed.
Troubleshooting
An advanced analysis takes a long time
The 70B model is substantially slower than the 8B. Switch models while you are still exploring, or reduce the corpus size.
Results differ between runs
Expected for the advanced analyses. Lower the temperature for more consistent output; the basic analyses are deterministic and will not vary.
A file will not load
Confirm the format is one of PDF, TXT, DOC, DOCX, RDF or XML. Scanned PDFs contain no text layer and need OCR first.
Citation & licence
Developed by Maria Schoinaki, within the Horizon Europe project grant agreement 101087269.
Licence
Released under the PolyForm Noncommercial 1.0.0. You may use, modify and share it for any purpose that is not primarily intended for commercial advantage or monetary compensation: research, teaching and personal use all qualify. Copies and modified versions must carry the licence and the notices that came with them. The software is provided as is, without warranty.
PolyForm Noncommercial is a source-available licence rather than an open source one: the noncommercial restriction is not permitted under the Open Source Definition.
Citation
Schoinaki, M. (2026). TALOS Text Analyzer: AI Corpus Analysis (version 1.0).
TALOS AI4SSH Laboratory, University of Crete.
https://tools.talos-lab.eu/text_analyzer/frontend/.