Abstract: Aiming at the problems of   less obvious or fewer sentiment words and euphemistic expressions in current implicit sentiment statements, we proposed an implicit sentiment analysis method based on semantic feature extraction. The method  introduced factual information related to implicit sentiment statements as auxiliary features, and used RoBERTa pre-training model to perform deep semantic interaction between the text and its auxiliary features in order to obtain global features. At the same time, a bidirectional gated recurrent unit (BiGRU) was used to capture local features, and finally, the sentiment weight was calculated by combining with attention pooling technique, so as to identify and understand the implicit sentiment information more accurately. The simulation experiments were conducted on  Snopes and PolitiFact datasets, and the results show  that the method has excellent performance  in implicit sentiment analysis. It not only surpasses existing methods in multiple evaluation metrics, but also significantly improves the overall performance, providing an effective solution for a wider range of sentiment analysis application scenarios, especially when dealing with complex and indirectly expressed sentiment content, it has important application value and significance.

Key words: semantic feature, implicit sentiment analysis, bidirectional gated recurrent unit, attention pooling