Cross-Document Cross-Lingual Natural Language Inference via RST-enhanced Graph Fusion and Interpretability Prediction

Natural Language Inference (NLI) is a fundamental task in both natural language processing and information retrieval. While NLI has developed many sub-directions such as sentence-level NLI, document-level NLI and cross-lingual NLI, Cross-Document Cross-Lingual NLI (CDCL-NLI) remains largely unexplored. In this paper, we propose a novel paradigm for CDCL-NLI that extends traditional NLI capabilities to multi-document, multilingual scenarios. To support this task, we construct a high-quality CDCL-NLI dataset including 1,110 instances and spanning 26 languages. To build a baseline for this task, we also propose an innovative method that integrates RST-enhanced graph fusion and interpretability prediction. Our method employs RST (Rhetorical Structure Theory) on RGAT (Relation-aware Graph Attention Network) for cross-document context modeling, coupled with a structure-aware semantic alignment mechanism based on lexical chains for cross-lingual understanding. For NLI interpretability, we develop an EDU-level attribution framework that generates extractive explanations. Extensive experiments demonstrate our approach's superior performance, achieving significant improvements over both traditional NLI models such as DocNLI and R2F, as well as LLMs like Llama3 and GPT-4o. Our work sheds light on the study of NLI and will bring research interest on cross-document cross-lingual context understanding, semantic retrieval and interpretability inference. Our dataset and code are available at \href{https://anonymous.4open.science/r/CDCL-NLI-637E/}{CDCL-NLI-Link for peer review}.