[Paper Review] Neutral Utterances are Also Causes: Enhancing Conversational Causal Emotion Entailment with Social Commonsense Knowledge

Jiangnan Li, Fandong Meng+6 more|arXiv (Cornell University)|May 2, 2022

Sentiment Analysis and Opinion Mining4 citations

TL;DR

This paper proposes a Knowledge Enhanced Conversation graph (KEC) to improve Conversational Causal Emotion Entailment by integrating social commonsense knowledge (CSK) and a sentiment-aware knowledge selection strategy, significantly boosting detection of causal utterances with different emotions—especially neutral causes—beyond baseline models that rely solely on emotion embeddings.

ABSTRACT

Conversational Causal Emotion Entailment aims to detect causal utterances for a non-neutral targeted utterance from a conversation. In this work, we build conversations as graphs to overcome implicit contextual modelling of the original entailment style. Following the previous work, we further introduce the emotion information into graphs. Emotion information can markedly promote the detection of causal utterances whose emotion is the same as the targeted utterance. However, it is still hard to detect causal utterances with different emotions, especially neutral ones. The reason is that models are limited in reasoning causal clues and passing them between utterances. To alleviate this problem, we introduce social commonsense knowledge (CSK) and propose a Knowledge Enhanced Conversation graph (KEC). KEC propagates the CSK between two utterances. As not all CSK is emotionally suitable for utterances, we therefore propose a sentiment-realized knowledge selecting strategy to filter CSK. To process KEC, we further construct the Knowledge Enhanced Directed Acyclic Graph networks. Experimental results show that our method outperforms baselines and infers more causes with different emotions from the targeted utterance.

Motivation & Objective

• To address the limitation of existing models in detecting causal utterances with emotions different from the targeted utterance, especially neutral causes.
• To enhance causal reasoning in conversational context by incorporating social commonsense knowledge (CSK) into graph-based representations.
• To design a sentiment-realized knowledge selection strategy that filters CSK for emotional and contextual relevance.
• To improve the performance of causal emotion entailment models beyond those relying only on emotion embeddings.
• To develop a Knowledge Enhanced Directed Acyclic Graph (KE-DAG) network for effective message passing over KEC graphs.

Proposed method

• Construct a Knowledge Enhanced Conversation graph (KEC) where nodes represent utterances and edges carry both utterance representations and selected CSK.
• Introduce a sentiment-realized knowledge selecting strategy that filters CSK based on speaker identity and emotional alignment with the utterance.
• Design a Knowledge Enhanced DAG network with separate GRU units for contextual knowledge (GRU_k) and self-loop knowledge (GRU_s) to model knowledge propagation.
• Use gated mechanisms to control knowledge flow, ensuring only relevant CSK influences node representations during message passing.
• Train the model using a contrastive prediction head that predicts causal pairs between utterances in the conversation graph.
• Apply a window-based context aggregation mechanism to focus on locally relevant utterances, with optimal performance at small window sizes.

Experimental results

Research questions

• RQ1 Can social commonsense knowledge improve the detection of causal utterances with emotions different from the targeted utterance?
• RQ2 How does sentiment-aware knowledge selection affect model performance in causal emotion entailment?
• RQ3 Does integrating CSK into a graph-based architecture outperform models relying only on emotion embeddings?
• RQ4 What is the impact of window size on causal reasoning performance in conversational context?
• RQ5 Can the model effectively detect neutral causal utterances, which are prevalent but hard to identify?

Key findings

• The proposed method achieves a macro F1 score of 83.42 on the RECCON dataset, significantly outperforming the DAG-ERC baseline.
• The model improves recall by 12% on causal pairs with different emotions (DE pairs), especially boosting detection of neutral causal utterances by 12%.
• Removal of the contextual knowledge GRU unit leads to a performance drop, confirming its necessity for effective knowledge integration.
• The sentiment-realized knowledge selection strategy enables the model to maintain strong performance even without explicit emotion embeddings, due to effective CSK filtering.
• The optimal window size for context aggregation is small (e.g., 2 steps), as larger windows reduce performance due to overemphasis on nearby utterances.
• The variant model that directly adds knowledge to node representations performs worse (macro F1: 80.38) than the proposed gated knowledge mechanism, proving the value of controlled knowledge flow.