Seq2Path: Generating Sentiment Tuples as Paths of a Tree

ACL Conference

Annual Meeting of the Association for Computational Linguistics is the most prestigious top academic conference in the field of natural language processing, sponsored by the International Association for Computational Linguistics. It is held once a year, and the accepted papers cover various research directions in natural language processing, including dialogue systems, semantic analysis, summary generation, information extraction, question answering systems, language models, machine translation, sentiment computing, and knowledge graphs.

Seq2Path: Generating Sentiment Tuples as Paths of a Tree

Generative information extraction has become a new paradigm for many information extraction tasks due to its good performance and end-to-end characteristics. The mainstream Seq2Seq information extraction framework faces the problem of introducing sequence constraints between tuples during the decoding stage when dealing with multiple tuples. This paper proposes a Seq2Path method to address this issue. Specifically, each tuple to be identified is regarded as an independent path in a tree structure, and the average loss over paths is fitted during the training stage. In the prediction stage, beam search with constrained decoding is used for decoding, and a discriminative token is introduced to automatically select the correct path. This paper takes the aspect-based sentiment analysis (ABSA) task as an example and conducts experiments on five public datasets corresponding to ABSA subtasks. The results fully verify the effectiveness of the proposed method. The current method has been applied in sentiment computing related scenarios of Alibaba Entertainment AI Brain (Beidou Star).