A Model Context Protocol (MCP) server for Latin Natural Language Processing (NLP), reported speech detection, and LiLa Knowledge Base querying.

Features

• Latin tokenization with enclitic -que splitting
• UDPipe-based lemmatization, POS tagging, and morphological analysis
• Non-finite verb identification
• Reported speech detection using a fine-tuned LaBERTa transformer latin model
• LiLa Knowledge Base SPARQL querying and exporting results
• MCP-compatible tool interface for Claude Desktop, VS Code, and MCP Inspector

---

System Overview

The server provides a pipeline of interoperable MCP tools for Latin NLP and Digital Humanities workflows.

The tools are designed to be used sequentially, but may also be used independently.

---

Tools

| Tool | Description | --------------------------------------------------------------------------------------------------------| | tokenize_latin_text | Tokenize Latin text with sentence splitting and enclitic handling | | parser | Morphological analysis and preprocessing using UDPipe | | detect_reported_speech_from_text | Transformer-based reported speech detection | | get_lila_lemma_info | Query the LiLa Knowledge Base for lemma information | | get_lila_lemma_tokens_dataframe | Retrieve LiLa corpus token occurrences and count attestations per work | | export_lila_lemma_tokens_csv | Export LiLa corpus token occurrences as a CSV file |

1. Latin NLP Parsing Pipeline

The parser tool performs:

• tokenization
• sentence segmentation
• lemmatization
• POS tagging
• morphological analysis
• non-finite verb identification

The tool calls the UDPipe API and uses the model:

latin-evalatin24-240520

The model was evaluated on EvaLatin campaign in 2024 and trained with on Latin Dependency Treebanks.

References

• EvaLatin 2024 overview: https://aclanthology.org/2024.lt4hala-1.21/

• UDPipe model repository: https://github.com/ufal/evalatin2024-latinpipe

---

2. Reported Speech Detection

A. Preparation Tool

The parser tool also prepares the linguistic input required by the reported speech detection model.

This preparation stage:

• aligns UDPipe tokenization with the original tokens
• prepares aligned linguistic features
• formats the input for transformer inference

---

B. Reported Speech Detector

The detect_reported_speech tool performs token-level reported speech prediction.

It takes the output of the parsing/preparation stage as input and returns:

• token-level predictions
• confidence scores

The model is:

• the first experimental Latin reported speech detection model at token level
• a fine-tuned LaBERTa model for token classification

References

• Hugging Face model repository: https://huggingface.co/agudei/latin-reported-speech-laberta

• Paper describing the experiment: https://aclanthology.org/2026.latechclfl-1.24/

---

3. LiLa Knowledge Base Querying

The get_lila_lemma_info tool provides simplified access to the LiLa Knowledge Base.

The tool:

• accepts a Latin lemma
• performs SPARQL queries automatically
• retrieves lexical and linguistic information
• simplifies access to Linked Open Data resources

The tool is designed to help users interact with LiLa without manually writing SPARQL queries.

LiLa

• LiLa Knowledge Base: https://lila-erc.eu/sparql/

4. LiLa Corpus Attestation Retrieval

The get_lila_lemma_tokens_dataframe tool retrieves corpus attestations linked to a Latin lemma in the LiLa Knowledge Base.

The tool:

• retrieves token occurrences associated with a lemma
• retrieves token URIs
• retrieves work titles
• computes occurrence frequencies per work
• structures results as a dataframe-like output

This enables corpus-based lexical exploration and quantitative analysis of lemma attestations across Latin works.

---

5. LiLa CSV Export

The export_lila_lemma_tokens_csv tool exports LiLa corpus attestation results as a CSV file.

The exported CSV includes:

• token forms
• token URIs
• work titles

The tool is designed for:

• corpus analysis
• spreadsheet analysis
• downstream NLP workflows
• Digital Humanities research pipelines

---

Installation

Install dependencies with:

uv sync

---

Running the Server

Recommended

uv run mcp-latin

Alternative

uv run python -m mcp_latin -vv

The MCP server will run at:

http://localhost:8001/mcp

---

MCP Inspector

You can test the server locally with:

npx @modelcontextprotocol/inspector

Then connect Inspector to:

http://localhost:8001/mcp

---

Example Prompts

Tokenization

Use the MCP tool tokenize_latin_text on:
"Senatus populusque Romanus."

---

Parsing

Use the MCP tool parser on:
"Non potui, inquit, sustinere illud durum spectaculum."

---

Reported Speech Detection

Use the Latin MCP tools only.

1. Parse:
"HISPO ROMANIUS alio colore dixit illam non amore adulescentis sed odio patris sui secutam"

2. Detect reported speech.

---

LiLa Query

Use the MCP tool get Lila information on the  "probabilis".

LiLa Query

Use the MCP tool get occurrences of the on the lemma "probabilis" and export the results.

---

Development Container

A reproducible VS Code devcontainer is included in:

.devcontainer/

See:

.devcontainer/README.md

for details.

---

Notes

• UDPipe requests require internet access.
• Hugging Face model weights are downloaded automatically.
• The server is designed for MCP-compatible clients such as Claude Desktop and VS Code MCP integration.