A Co-Attentive Infusion of Structural Knowledge into Language Models towards Legal Textual Inference

In this paper, we explore the potential and impact of semantic structures in legal textual inference and propose a structural knowledge infusion architecture, AIKLR, which integrates abstract meaning representation (AMR) with pre-trained language models (PLMs). We hypothesize that AMR's structural representations can effectively capture complex logical relationships in sentences, complementing their semantic representations and thereby enhancing the legal reasoning and inference capabilities of language models. Our proposed architecture demonstrates significant accuracy improvements of 16% to 18% over baseline models. Furthermore, AIKLR achieves competitive results against recent large language models (LLMs) on the COLIEE datasets, with a remarkable 32% difference in predicted outputs. These results highlight AIKLR's ability to leverage structural representations for improved logical reasoning, making it a valuable reinforcement for LLMs in legal text processing.