Adversarial Training Based Cross-Lingual Emotion Cause Extraction

Emotion cause extraction (ECA) aims to identify the reasons behind a certain emotion expression in a text. It is a key topic in natural language processing. Existing methods relies on high-quality emotion resources and focuses on only one language. However, the public annotated corpora is fairly rare. Therefore, we propose an adversarial training based cross-lingual emotion cause extraction approach to leverage the semantic and emotion knowledge in a resource-abundant language (source language) for ECA in a resource-scarce language (target language). Instead of large-scale parallel corpora, we capture task-related but language-irrelevant features only on a small-scale Chinese corpora and an English corpora. In addition, an attention mechanism based on position and emotion expression information is designed to obtain the key parts of the clause devoting to ECA. Our proposed approach could capture rich semantic and emotion information in ECA learning process. It is demonstrated that our method can achieve better performance than the state-of-the-art results.