Causal Discovery Inspired Unsupervised Domain Adaptation for Emotion-Cause Pair Extraction

TL;DR

This work tackles emotion-cause pair extraction under unsupervised domain shift (UDA-ECPE) by marrying causal discovery with a variational autoencoder. It introduces CaRel-VAE, a two-latent VAE modeling emotions ($Z^e$) and events ($Z^c$) with a discriminative, task-focused ELBO and a variational-posterior regularizer (Bhattacharyya distance and MMD) to disentangle the latent spaces. It further proposes CD-SelfTrain, a target-domain self-training regime that updates pseudo-labeled data each epoch to uncover domain-specific causal relations. Empirically, the approach achieves notable gains over strong baselines on CH-ECPE and EN-ECPE, demonstrating effective cross-domain ECPE and enhanced causal discovery capabilities.

Abstract

This paper tackles the task of emotion-cause pair extraction in the unsupervised domain adaptation setting. The problem is challenging as the distributions of the events causing emotions in target domains are dramatically different than those in source domains, despite the distributions of emotional expressions between domains are overlapped. Inspired by causal discovery, we propose a novel deep latent model in the variational autoencoder (VAE) framework, which not only captures the underlying latent structures of data but also utilizes the easily transferable knowledge of emotions as the bridge to link the distributions of events in different domains. To facilitate knowledge transfer across domains, we also propose a novel variational posterior regularization technique to disentangle the latent representations of emotions from those of events in order to mitigate the damage caused by the spurious correlations related to the events in source domains. Through extensive experiments, we demonstrate that our model outperforms the strongest baseline by approximately 11.05\% on a Chinese benchmark and 2.45\% on a English benchmark in terms of weighted-average F1 score. We have released our source code and the generated dataset publicly at: https://github.com/tk1363704/CAREL-VAE.

Figures (6)

• Figure 1: An illustrative example of the UDA-ECPE task. Orange and green highlights respectively denote emotion and cause clauses.
• Figure 2: The t-SNE visualizations of the sentence embeddings from Amazon Reviews multi-domain sentiment corpus and the clause embeddings from the Chinese UDA-ECPE corpora.
• Figure 3: The architecture of our model CaRel-VAE.
• Figure 4: The t-SNE visualizations of the clause embeddings from the English UDA-ECPE corpora
• Figure 5: Experimental results of CaRel-VAE w/o MMD and CaRel-VAE for normal and self-chain cases. The normal case refers to an emotion-cause pair composed of two different clauses, while for the self-chain case a pair are mentioned in the same clause.