Research on End-to-End Aspect-Level Sentiment Analysis Based on the BBGRU Model
YuJia Su · 2024
Aiming at the inadequacy of existing research on end-to-end aspect-based sentiment analysis (E2E_ABSA) methods that don't make full utilize textual information, we propose a bidirectional encoding representation of encoder-decoder - bidirectional gated recurrent unit, i.e., the BBGRU model. The model adopts the uniform sequence annotation method in E2E_ABSA, which uniformly annotates aspectual item positions and sentiment polarity. Additionally, instead of using traditional Word2vec and Glove models, a pre-trained language model BERT is utilized as the embedding layer to represent word vectors. A BIGRU layer is added behind the BERT model to capture short-term and long-term dependencies by employing reset gate and update gate, respectively. At the same time, it applies a bidirectional gating mechanism aiming to fully consider both forward and backward input information to obtain richer feature representations, thus improving the accuracy of aspect term extraction and sentiment polarity prediction. Comparative experiments were conducted by using the proposed BBGRU model with BERT+other models. The experimental results subtly reveal that the model presented in this paper improves the precision (P), recall (R) and micro-averaging (Micro_F1) on three customary benchmark datasets, Laptop14, Rest14 and Rest16, by 1.25%, 1.22%, 1.50%; 1.41%, 0.64%, 1.36% and 1.68%, 1.08%, 1.38%, all of which performed better than outperform the other models.