Double embedding and bidirectional sentiment dependence detector for aspect sentiment triplet extraction

Abstract

Aspect sentiment triplet extraction (ASTE) is a popular subtask related to aspect-based sentiment analysis (ABSA). It extracts aspects and their associated opinion expressions and sentiment polarities from comment sentences. Previous studies have proposed a multitask learning framework that jointly extracts aspect and opinion terms and treats the sentiment analysis task as a table-filling problem. Although the multitask learning framework solves the problem of identifying overlapping opinion triples, the entire model cannot explicitly simulate interactions between aspects and opinions. Therefore, we propose a sentiment-dependence detector based on a dual-table structure that starts from two directions, aspect-to-opinion and opinion-to-aspect, to generate two sentiment-dependence tables dominated by two types of information. These complementary directions allow our framework to explicitly consider interactions between aspects and opinions and better identify triples. Moreover, we use a double-embedding mechanism—character-level and word-vector embeddings—in the model for triplet extraction that enables it to represent contexts at different granularity levels and explore high-level semantic features. To the best of our knowledge, this study presents the first bidirectional long short-term memory (BiLSTM) model based on double embedding used to perform ASTE tasks. Finally, our analysis shows that our proposed bidirectional sentiment-dependence detector and double-embedding BiLSTM model achieve more significant results than the baseline model for triples with multiple identical aspects or opinions.