How to use the Greek NLP Swiss Knife platform — ten research applications and approximately 200,000 lines of code for computational analysis of Greek texts from antiquity to the present.
The Greek NLP Swiss Knife is a suite of ten web applications for Greek text analysis, organized into four tiers based on their underlying technology. All applications run in your browser — no installation required. Some applications require API keys for the LLM providers they use.
API Keys: Applications in the LLM-powered and Neurosymbolic tiers require you to provide your own API keys from one or more of: Google Gemini, OpenAI, or Anthropic. These keys are entered directly in each application's interface and are never stored on the server. Classical NLP apps and the Dialect Generator do not require any API key.
No LLM required — deterministic NLP pipelines for text analysis. No API key needed.
NATS provides named entity recognition with full Greek language support, document embeddings, interactive network visualizations, and community detection for textual analysis. It works entirely with classical NLP models, so no API key is required.
Identify persons, places, organizations in Greek text.
Interactive graphs showing entity relationships.
Automatic clustering of related entities.
Vectorize texts for similarity comparison.
A comprehensive text analysis workbench inspired by Voyant Tools, enhanced with NLP capabilities specifically tuned for Greek. Upload one or more texts and explore them through multiple analytical lenses.
Frequency distributions with stopword filtering.
Keyword in context search across your corpus.
Train embeddings and explore word analogies.
Discover thematic clusters in your texts.
Part-of-speech annotation for Greek.
Positive/negative polarity detection.
Multi-provider applications using Gemini, OpenAI, and Anthropic. Bring your own API key.
MEDEA is the most comprehensive application in the suite, offering six integrated AI capabilities for classical studies: cross-lingual translation across 18 ancient languages, knowledge graph extraction, emotional landscape analysis, and Prolog-based symbolic reasoning via the Zeugma module.
Analyze how nature is discussed in any text. The application segments text by paragraph, sentence, or token and uses LLMs to detect and classify nature-related discourse, producing statistical breakdowns and exportable data.
Measure linguistic distance across seven independent dimensions for eleven language pairs. This tool quantifies how languages have changed over time — from Ancient Greek to Modern Greek, Latin to Romance languages, and more — using established typological databases.
Core vocabulary comparison across time.
Automated phonological similarity.
Syntactic feature divergence.
Typological feature comparison.
LLM generation verified by Prolog reasoning, phonological engines, or rule-based parsers. API key required.
Identify and generate rhyme patterns in Modern Greek poetry using a neurosymbolic pipeline that combines multi-model LLM analysis with a deterministic phonological engine. The system implements a complete taxonomy of Greek rhyme types and uses RAG-enhanced corpus retrieval for verification.
Analyze the plot structure of any narrative text using five major narrative theories. PlotAnalyzer uses LLMs to identify structural elements and then verifies them against formal definitions of each theory's components.
Extract terms from domain-specific corpora, generate ISO 1087-compliant definitions using multi-LLM comparison, and validate each definition with a symbolic parser that checks for proper genus, differentia, circularity, and conciseness.
Locally hosted models fine-tuned on the GRDD+ dialectal corpus. No API key needed.
Generate text in four Greek regional dialects — Pontic, Cretan, Northern Greek, and Cypriot — using open-source language models fine-tuned with LoRA adapters on the GRDD+ dialectal corpus. Unlike the other applications, the Dialect Generator runs models locally on the server and requires no external API key.
Pontic, Cretan, Northern Greek, Cypriot.
Llama 3.1, Llama 3, and Krikri 8B.
Trained on 20k+ GRDD+ dialectal examples.
Tokens appear as they are generated.
Do I need to create an account?
No. All applications are freely accessible without registration. You only need API keys for the LLM-powered and Neurosymbolic tier applications.
Which LLM provider should I use?
All supported providers (Google Gemini, OpenAI, Anthropic) work well. For Greek-specific tasks, we recommend trying multiple providers using the multi-model comparison feature available in several applications. Each provider has different strengths with Greek text.
Are my texts stored on the server?
No. Texts are processed in memory during your session and are not stored. API keys are sent directly to the respective provider and are not retained.
Can I use these tools for Ancient Greek?
MEDEA-NEUMOUSA's Necromancer module supports Ancient Greek translation. NATS and Voyant-NLP can process any Greek text including Ancient Greek, though their NER models are primarily trained on Modern Greek. The Linguistic Distance tool includes Ancient Greek as a language pair.
The Dialect Generator is slow — is something wrong?
No. The dialect models run on CPU, which is inherently slower than GPU for neural text generation. Expect roughly 5–10 seconds per token. We are working on GPU acceleration for faster inference.
I get "Could not find a replica" or the app takes a long time to load.
Some applications — particularly the Dialect Generator — use scale-to-zero to manage hosting costs. When nobody has used the app for a while, Azure shuts down the container to save resources. The first request after that triggers a cold start: the container boots up and loads the model into memory, which can take a couple of minutes. Just wait and retry. Once the app is running, subsequent requests will be much faster (no restart needed).
Who built this platform?
The Greek NLP Swiss Knife was built by Stergios Chatzikyriakidis and collaborators. The dialect generation models are based on the GRDD+ dataset research. The full platform is described in the forthcoming DSH paper.