A spect-level Slentiment Analysis based on BERT Fusion Multi-module

Aspect-level sentiment analysis is a fine-grained sentiment analysis task. Its task is to calculate people's opinions, evaluations, attitudes and emotions expressed by entities. In recent years, many deep language models have made great progress in this regard, including BERT. The initial layer to the middle layer of BERT can extract grammatical information, but the semantic information of higher layers is often easily ignored. Because extracting sentence sentiment is based on semantics, this article adds two modules of parallel aggregation and hierarchical aggregation on the basis of BERT. Parallel aggregation is used for aspect extraction, and hierarchical aggregation is used for aspect sentiment classification tasks, and the conditional random field is used as Sequence marking tasks to extract more semantic information. From the experimental results on the SemEval 2014 and SemEval 2016 data sets, it can be seen that the accuracy and F1 value of the model proposed in this paper are better than the comparison model, confirming the effectiveness of the model.