Using BERT and Word Definitions for Implicit Sentiment Analysis

Implicit sentiment frequently appears in texts. To accurately classify the implicit sentiments of sentences, this paper uses search engines to retrieve definitions of words and incorporates knowledge into language models for implicit sentiment analysis. We propose a model using BERT and word definitions for implicit sentiment analysis (BDISA). Specifically, we use BERT and Word2Vec to represent tokens and word definitions, respectively. Then, by using a fusing layer, we fuse these two representations. Finally, the model outputs the predicted sentiment labels. The experimental results show that the macro-F1 value of the BDISA model is 9.56% and 0.46% higher than KG-MPOA (random and BERT) on the SMP2019-ECISA dataset. The ablation experiments illustrate that word definitions and the fusion layer are critical for the model.