Llm Tool Greek Lexicon

This repository contains the code, processing scripts, and resources for the project "Tool-Augmented Language Models for Precision Analysis of Ancient/Koine Greek Texts Using a Graph-based Lexicon".

The project introduces a novel methodology that enhances Large Language Models (LLMs) with a dedicated, real-time lexical tool to achieve high-precision morphosyntactic and functional grammar analysis of Ancient Greek texts.

An interactive website to explore the parsed results is available here: https://wmotte.github.io/llm_tool_greek_lexicon/docs/

---

The Problem

Computational analysis of Ancient Greek has traditionally been hampered by the limitations of existing methods.

• Rule-based morphological parsers often fail to handle the language's dialectical complexity and the frequent occurrence of ambiguous word forms.
• Modern Retrieval-Augmented Generation (RAG) systems, while powerful, retrieve information based on semantic similarity (approximate matching). This lacks the precision required for a morphologically rich language like Ancient Greek, where a single character can fundamentally alter a word's meaning (e.g., λόγος vs. λύγος).

This project bridges this methodological gap by developing a system capable of exact lemma identification and subsequent linguistic analysis, moving beyond purely rule-based and similarity-based approaches.

Our Approach: Tool-Augmented AI

We present a tool-augmented modeling approach that integrates LLMs with real-time, dynamic access to a structured lexical database. This architecture moves beyond simple context-stuffing (RAG) and allows the LLM to actively query a specialized tool for precise information when needed.

The core components of our system are:

1. A Neo4j Knowledge Graph: We constructed a graph database containing a complete, scholarly Greek-Dutch lexicon. Graphs are well-suited for structuring relational databases.
2. The Model Context Protocol (MCP): The knowledge graph is made available to the LLM as a "tool" via an MCP server. This protocol allows any tool-calling LLM to dynamically query the lexicon for exact lemma matching, overcoming the context window limitations of traditional methods.

The Lexicon: Knowledge Graph Construction

The basis of our tool is a machine-readable knowledge graph built from an up-to-date Greek-Dutch Lexicon (SvBKR), a scholarly dictionary covering Greek vocabulary from Homer to the second century CE.

The construction process involved several key steps:

• Data Retrieval: The complete lexicon, comprising >40k lemma entries, was retrieved by web scraping the publicly available HTML interface.
• Expansion of Forms: Many dictionary entries were in a compressed format (e.g., providing only the nominative and genitive singular forms). These were manually expanded into their full inflected forms using standard paradigms to facilitate precise matching.
• Diacritic-Insensitive Indexing: To allow for flexible matching against unaccented Greek input, each lexical form was duplicated into a variant without diacritics (e.g., ἀβάκιον → αβακιον). This ensures robust lookup while preserving the integrity of the original, accented forms.
• Graph Modeling: All structured data was imported into a Neo4j graph database. Each lemma is a LexicalEntry node with properties for grammar, glosses, and source info. This graph structure enables efficient querying of lemma frequency, contextual usage, and intertextual relationships.

The processing scripts used for this process are available in this repository to ensure replication and adaptation.

Example Database Queries

The LLM interacts with the knowledge graph using Cypher queries. This allows for precise, efficient retrieval of lexical data.

1. Single Lemma Lookup (Exact Match) This query finds a lemma by matching its normalized, accent-free form.

// This query locates the exact lemma λόγος by matching its normalized form ‘logos’
MATCH (l:Lemma)-[:HAS_ENTRY]->(e:Entry)-[:BELONGS_TO]->(d:Dictionary)
WHERE l.text_no_accents = "logos"
AND d.name = "SvBKR"
RETURN l.text, e.text

2. Batch Processing for Multiple Lemmas This demonstrates the efficiency of retrieving multiple entries in a single call.

// This batch query simultaneously retrieves lexical entries for multiple lemmas
MATCH (l:Lemma)-[:HAS_ENTRY]->(e:Entry)-[:BELONGS_TO]->(d:Dictionary)
WHERE l.text_no_accents IN ["eimi", "logos", "theos", "anthropos"]
AND d.name = "SvBKR"
RETURN l.text, l.text_no_accents, e.text

Evaluation and Test Cases

The system's performance was systematically evaluated using three distinct and challenging Greek texts, each chosen for a specific purpose.

1. Hebrews 2:9 (New Testament): Selected for its degree of grammatical and syntactical complexity. We also tested a minority textual variant (χωρὶς θεοῦ vs. χάριτι θεοῦ) to evaluate the model's ability to handle subtle but critical input changes, a misguided attention test against the "Einstellung effect".
2. Job 2:9 (Septuagint): This passage features a relative long, 93-word diatribe that is an addition in the Greek LXX version compared to the Hebrew Masoretic Text. It serves as a test case for morphosyntactic performance on a lengthy, less common text that is not well-covered by online biblical data.
3. The Gospel according to the Egyptians: This fragmentary, apocryphal text is not openly available online and is restricted to licensed academic databases. This minimizes the chance it was in the LLM's pre-training data, providing a good benchmark for evaluating the model's parsing capabilities on "unseen" text with complex syntactic structures.

The results, which can be explored on the interactive website, demonstrate that the tool-augmented approach provides good performance in both morphosyntactic parsing and functional grammar analysis across all test cases.

Citation

If you use this work in your research, please cite the following manuscript:

Otte, W. M., van Wieringen, A. L. H. M., & Koet, B. J. (in preparation). Tool-Augmented Language Models for Precision Analysis of Ancient/Koine Greek Texts Using a Graph-based Lexicon.

License

This project is licensed under the CC0 1.0 Universal License. See the LICENSE file for details.