Cross-Document Cross-Lingual NLI via RST-Enhanced Graph Fusion and Interpretability Prediction

TL;DR

This work defines Cross-Document Cross-Lingual NLI (CDCL-NLI) and releases a 25,410-sample dataset spanning 26 languages to probe multi-document, multilingual inference. It introduces a three-component framework that fuses RST-based discourse graphs across documents via lexical chains, encodes them with RGAT to capture cross-document structure, and provides EDU-level extractive explanations to enhance interpretability. Across extensive experiments, the proposed method outperforms conventional NLI models and several large language models, while delivering transparent rationale through EDU-level attributions. The dataset and code at https://github.com/Leonardo123-ui/CDCL_NLI enable further exploration of cross-document cross-lingual reasoning, hallucination mitigation, and interpretable inference.

Abstract

Natural Language Inference (NLI) is a fundamental task in natural language processing. 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: CDCL-NLI, which extends traditional NLI capabilities to multi-document, multilingual scenarios. To support this task, we construct a high-quality CDCL-NLI dataset including 25,410 instances and spanning 26 languages. To address the limitations of previous methods on CDCL-NLI task, we further propose an innovative method that integrates RST-enhanced graph fusion with interpretability-aware prediction. Our approach leverages RST (Rhetorical Structure Theory) within heterogeneous graph neural networks for cross-document context modeling, and employs a structure-aware semantic alignment based on lexical chains for cross-lingual understanding. For NLI interpretability, we develop an EDU (Elementary Discourse Unit)-level attribution framework that produces extractive explanations. Extensive experiments demonstrate our approach's superior performance, achieving significant improvements over both conventional NLI models as well as large language models. Our work sheds light on the study of NLI and will bring research interest on cross-document cross-lingual context understanding, hallucination elimination and interpretability inference. Our code and datasets are available at "https://github.com/Leonardo123-ui/CDCL_NLI" for peer review.

Figures (14)

• Figure 1: A CDCL-NLI example. Premise in English and French. The Entailment label requires combining information from both documents in premise.
• Figure 2: Overview of the CDCL-NLI dataset construction process and a data example. Premise contains $D_1$ and $D_2$. Explanation is extracted from premise to enhance interpretability. Human annotation is based on language translated into English.
• Figure 3: Statistic visualization of token length, EDU numbers, label distribution and language composition.
• Figure 4: Our CDCL-NLI framework processes premise documents (${D_1,D_2}$) and a hypothesis through a multi-stage process: 1) RST Graph Construction, where an RST parser generates initial discourse structures (${G_{D_1}}$ and ${G_{D_2}}$) which are then fused into a single premise graph (${G_P}$) using semantic edges derived from XLM-RoBERTa embeddings; 2) Graph Representation, where the fused graph is processed by RST-GAT layers; and 3) Interpretability and Classification, which extracts node-level explanations and uses the graph representations ($\bm{h}_{G_p}$) and hypothesis representation ($\bm{h}_{hypo}$) to predict the final NLI label.
• Figure 5: RST graph construction. The RST module first segments text into EDUs(EDU$_1$-EDU$_6$), with boundaries in blue, and then organizes an RST tree $\mathcal{T}$ showing discourse relations.