Methods Study of a Unified ABSA Generation Framework Based on Pre-Trained Model Induced Dependency Tree

Aspect-based Sentiment analysis (ABSA) is a fine-grained sentiment analysis designed to extract the sentimental polarity and sentimental words related to aspect words. Sentiment analysis plays an important role in extracting users' sentiments. Dependency analysis, also called syntactic analysis, is used to study the grammatical structure of sentences and provide information about the connections between sentence words. In order to improve the accuracy of ABSA, researchers integrated syntactic information into ABSA to obtain the correct dependencies between words in the sentence, making ABSA achieve unprecedented good results on syntactic information. However, now the syntactic information of ABSA is provided by syntactic analysis tools. It pays too much attention to the correct sentence structure, which is easy to confuse the colloquial sentences, and lacks the flexibility of downstream task application. Therefore, the pre-trained model induced syntax tree with rich language knowledge was introduced, and the AE, SC and AESC tasks were performed combined with the unified ABSA generation framework. The results show that the syntactic tree induced by the pre-trained model combined with the unified ABSA generation framework enables the three subtasks, AE, ALSC and AESC, to achieve superior results on the Twitter dataset.