| With the increasing frequency of Internet commodity transactions,merchants hope to understand users’ preferences from user reviews,so as to improve their products.The predicament of traditional sentiment analysis is that it cannot accurately judge the user’s sentiment preference for various aspects of the product in the reviews,while fine-grained methods such as aspect-level sentiment analysis can effectively identify different aspects of sentiment,so it can better meet the needs of businesses.Dependency syntax-based methods can capture the long-distance information between aspect words and opinion words,so they are widely used in aspect-level sentiment analysis tasks,but these methods still suffer from insufficient syntactic information learning,and the separation of syntactic information learning and context information learning.To conduct in-depth analysis and research on this,the specific contents are as follows:1.Aiming at the problem of separation of syntactic information learning and contextual information learning,the syntactic information of sentences is obtained by using dependency syntax analysis as prior knowledge,the prior knowledge and sentence splicing are used as input,and syntactic information and contextual information are learned at the same time.Aiming at the problem of insufficient syntactic information learning,dynamic weight matching is carried out through dependency distance,and self-attention mechanism is used to learn dependency label information.Experimental results show that the learning of syntactic information and the interaction with contextual information are more effective than other methods.2.Aiming at the unstable acquisition of prior knowledge in the above methods and the problem that dynamic weight matching does not distinguish the same dependency distance,it is corrected.In view of the problem that it is difficult to fully judge emotions based on the current sentence information in multiple aspects of a sentence,an aspect-level sentiment analysis method based on multi-granularity fusion is proposed.By introducing contextual sentence information,the emotional judgment of aspects in the current sentence is enhanced,and the output of the enhanced sentence is adjusted through the gating mechanism to alleviate the excessive influence of the contextual sentence information.The experimental results show that the multi-granularity information fusion and gating mechanism can further optimize and improve the model compared to the model that only fuses dependent syntactic information and word context information.3.Applying the above algorithm,a product review analysis system is designed and implemented.The results of the aspect sentiment analysis are displayed in a structured way,and the sentiment distribution of each aspect is counted,so as to understand the user’s satisfaction with all aspects of the product,and prompt the merchants to formulate better marketing strategies. |
