Identifying Logical Patterns in Text for Reasoning

Translating unstructured text into logical format is a key challenge for building ontologies automatically and addressing deductive inference. Most of the approaches have tackled the identification of concepts and relations in text, but few of them have addressed the most complex axioms like class expression subsumption. This work proposes DeLIR, a neuro-symbolic approach to identify complex logical patterns in text by combining a grammatical translation of dependency parsing trees and a fine-tuned Large language Model (LLM). DeLIR combines the strength of the parsing accuracy provided by a grammatical approach and pattern flexibility provided by a finetuned LLM. We evaluated our approach on FOLIO dataset for both translation capacity and inference capability. Our grammatical approach has a perfect parsing accuracy and combining the grammatical approach with LLMs improves the LLMS translation capacity: tinyLlama, T5-small-text2logic, Llama-7B and Mistral-7B. We also evaluate the inference capacity of the different LLMs. Mistral-7B, while being smaller than the state-of-the-art approach using GPT-4, presents similar results to predict the correct inference labels.