Inspicio: Open-Vocabulary, LLM-Based Sense Retrieval for Historical Languages
Inspicio is an open-vocabulary retrieval pipeline that links tokens in context to synsets of the Open English WordNet without requiring any source-language inventory or mapping. For each occurrence, an instruction-tuned LLM produces two English translations of the surrounding sentence, a small set of candidate dictionary-style definitions, and a few candidate English lemmas. These outputs drive a hybrid retrieval step combining dense definition-synset similarity, sparse lemma matching, and Maximal Marginal Relevance re-ranking. The pipeline is evaluated across a 6x6 grid of LLMs and sentence-embedding models on a new bilingual set of manually annotated Latin and Ancient Greek perception verbs, on a subset of the PREMOVE dataset, and on a diachronic sample.
The approach removes the need for a source-language sense inventory by using an instruction-tuned LLM to generate English translations, definitions, and lemmas, then combining dense and sparse retrieval with MMR re-ranking. Evaluation across a 6x6 grid of LLMs and sentence-embedding models on Latin and Ancient Greek perception verbs, PREMOVE subset, and diachronic sample provides a systematic comparison of component choices.
This work addresses a gap in NLP for historical and low-resource languages, where dedicated WordNets are incomplete or under construction. By leveraging English resources and LLM generation, it offers a scalable method for sense disambiguation without language-specific annotation, potentially lowering barriers for digital humanities and linguistic research.
The pipeline could enable new tools for historical text analysis, translation, and lexicography, reducing the cost of building sense inventories for low-resource languages. It may be valuable for academic research, digital humanities platforms, and language technology companies targeting multilingual or historical content.
Next observable signals include publication of the full evaluation results, release of the manually annotated Latin and Ancient Greek dataset, and potential application to other low-resource or historical languages. Further work may explore integration with existing diachronic corpora or adaptation to additional LLM architectures.