Open AccessAspect-Based Sentiment Analysis Based on Multi-Channel and Dynamic Weight
Author(s) -
Jin Cheng,
Huixiong Yi
Publication year - 2021
Publication title -
iop conference series. earth and environmental science
Language(s) - English
Resource type - Journals
eISSN - 1755-1307
pISSN - 1755-1315
DOI - 10.1088/1755-1315/693/1/012065
Subject(s) - computer science , concatenation (mathematics) , dependency (uml) , sentiment analysis , interaction information , benchmark (surveying) , artificial intelligence , graph , channel (broadcasting) , representation (politics) , mechanism (biology) , tree (set theory) , information loss , natural language processing , theoretical computer science , computer network , mathematical analysis , philosophy , statistics , mathematics , geodesy , epistemology , combinatorics , politics , political science , law , geography
With the wide use of graph convolution network (GCN) in various NLP tasks, researchers introduce GCN to aspect-based sentiment analysis. They utilize GCN over dependency tree to capture the syntactic information and obtain the final representation by modeling the interaction between the semantic and syntactic information. However, due to the unreliability of the construction of dependency tree, these interaction-based models may suffer from the side effect of unreliable syntactic information. To tackle the problem, we propose a multi-channel aspect-specific attention mechanism to ease the possible side effect by retaining and exploiting the original semantic information, which is not affected by the syntactic information. We also propose a dynamic weighted concatenation mechanism to sufficiently utilize the multi-channel information by dynamically assigning weights to the multi-channel information. The extensive experiments on five benchmark datasets indicate the effectiveness of the proposed mechanisms.