IF: An Interactive Fusion Network for Targeted Aspect-Based Sentiment Analysis

• Published in: 2022 3rd International Conference on Electronic Communication and Artificial Intelligence (IWECAI) pp. 134 - 140
• Main Authors: Li, Jiale, Zhou, Yajian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.01.2022
• Subjects: Analytical models; attention mechanism; Benchmark testing; Coal; Encoding; Fuses; LSTM; Pipelines; Sentiment analysis
• DOI: 10.1109/IWECAI55315.2022.00035

Targeted aspect-based sentiment analysis (TABSA)is a very challenging task in sentiment analysis field, which aims to determine the sentiment polarity for an aspect of a certain target. TABSA task is usually divided into two subtasks: aspect category detection and aspect sentiment classification. One simple thought for them is to use a pipeline method, but it is difficult to use in application and will produce cascading errors. Therefore, some researchers proposed joint methods by sharing the encoding of inputs. However, they did not explicitly model the relevance between the two subtasks or fuse the acquired information deeply. To address this problem, we propose a novel Interactive Fusion (IF) model, which can enable direct interaction between two subtasks and generate aspect categories and their sentiments simultaneously. Specifically, the sentence representations are obtained from the encoding layer, and then a fusion layer is introduced to make the two subtasks interact multiple rounds to fuse the representations better. Experimental results on two benchmark datasets show that our approach yields the state-of-the-art performance in TABSA task.