SRLM: Attentive fusion of structural representation into language model for legal textual entailment

Legal textual entailment (LTE) remains a major challenge in legal natural language processing due to the difficulty in capturing complex logical relations and the scarcity of annotated legal data. To address these limitations, this paper introduces SRLM (Structural Representation into Language Model), a method that attentively fuses semantic structures derived from Abstract Meaning Representation (AMR) of legal texts with contextual embeddings from pre-trained language models (PLMs). Although prior studies have explored combining AMR graphs with PLMs via simple concatenation or continued pre-training, these approaches failed to outperform text-only baselines, leaving the question of how to effectively leverage AMR for language understanding tasks open. The underlying idea of SRLM is a co-attentive fusion mechanism that establishes mutual dependencies and complementarity between structural and textual information, enabling the model to capture complex logical and semantic relations in juridical texts. Experiments on well-known benchmarks demonstrate that SRLM achieves up to 7.87 absolute F1-point improvements over text-only baselines and achieves performance competitive with or superior to large language models (LLMs) in certain settings while keeping the model compact. In particular, SRLM is highly effective in low-resource conditions, achieving up to 13 F1-point gains with only 10% of the training data. Comprehensive ablation studies further validate the efficacy of the proposed co-attentive mechanism compared to alternative fusion methods.