HERO：結合異質圖建模與大型語言模型相對因果評分之對話情緒因果識別模型

Abstract

對話情緒因果識別任務（Causal Emotion Entailment, CEE）旨在從對話歷史中識別出引發目標情緒的成因語句。現有基於圖神經網路的方法雖能有效建模對話的顯式結構（如時序順序、說話者關係與對話互動），但在需要深層語意推理的隱含因果識別上仍有不足。同時，直接以大型語言模型進行情緒成因判斷雖能提供有用的推理訊號，卻缺乏結構化對話建模能力，且可能引入不穩定或高成本的推論問題。 針對上述限制，本論文提出 HERO（HEterogeneous graph modeling with LLM-based Relative causal scOring for causal emotion entailment），一套結合大型語言模型相對因果評分與異質圖推理的整合框架。其中，大型語言模型相對因果評分模組（LLM-Relative Causal Scoring Module, LRCSM）估計每個候選語句相對於目標情緒的因果相關性強度。LRCSM 先以直接評分取得候選語句的初始因果相關分數，再透過比較式評分校正候選間的相對關係，使所得分數具備跨語句的可比性。上述由大型語言模型導出的因果相關分數隨後與異質圖建模模組（Heterogeneous Graph Modeling Module, HGMM）的結構建模結果整合，構成最終的成對因果分類器輸入。HGMM 整合語句上下文、情緒資訊與相對目標位置等多類型關係，並透過關係圖卷積網路進行訊息聚合。大型語言模型的因果相關分數則在 logit 層與圖結構輸出融合，使兩類訊號能互補協同。 在 RECCON-DD 與 RECCON-IE 上的實驗結果顯示，本方法在主要評估指標上優於本文納入比較的基準方法，且進一步的消融分析驗證了大型語言模型引導訊號與異質圖建模各自的貢獻。

Causal Emotion Entailment (CEE) aims to identify which utterances in a dialogue history cause the non-neutral emotion expressed by a target utterance. Existing graph-based methods are effective in modeling explicit dialogue structures, such as temporal order, speaker relations, and conversational interactions, but they may still struggle with implicit causes that require semantic reasoning beyond observable graph topology. Meanwhile, directly using Large Language Models (LLMs) to identify emotion causes can provide useful reasoning signals, but may lack structured dialogue modeling and introduce unstable or costly inference. To address these limitations, we propose HERO (HEterogeneous graph modeling with LLM-based Relative causal scOring for causal emotion entailment), an LLM-enhanced heterogeneous graph framework for CEE. HERO uses the LLM-Relative Causal Scoring Module (LRCSM) to estimate causal-relevance scores between candidate utterances and the target emotion. LRCSM first obtains direct causal-relevance scores and then applies comparative scoring to model relative relations among candidates. These LLM-derived scores are integrated with the structural representations produced by the Heterogeneous Graph Modeling Module (HGMM), which models utterance context, emotion information, and target-relative positions through relational graph convolution. The final prediction is made by a trainable pairwise cause classifier that combines structural representations with LLM-derived causal-relevance signals at the logit level. Experiments on RECCON-DD and RECCON-IE show that HERO outperforms the baselines considered in this thesis on the main evaluation metrics, with further analyses demonstrating the effectiveness of LLM guidance and heterogeneous graph modeling.