Research On Aspect-level Sentiment Analysis Method Based On Dependency Syntax Pruning

With the deepening of the concept of digital life and the rapid popularization of social media and e-commerce,a large amount of information in people’s lives is generated in the form of text.Generate a wide range of application value.Aspect-level sentiment analysis obtains people’s fine-grained emotional expression of opinion targets by discovering various aspects of text comments and determining the polarity of expressed emotions for each aspect.According to the aspect term,finding the most relevant emotional part from the context is the key to the aspect-target sentiment classification task.Some works achieve the goal by using the syntactic structure,but there is information irrelevant to the aspect term in the complete syntactic structure,which will bring some noise during context encoding,which will affect the model’s performance.Secondly,the ordinary graph convolutional network only uses the topological for information fusion,and cannot encode the category of dependency edges,resulting in the loss of some emotional information and effective text representation.In order to solve the above problems and effectively improve the classification accuracy,the main research contents of this paper are as follows:(1)When the graph structure encoder encodes the context according to the traditional dependency tree structure,irrelevant information will generate noise,and wrong emotional features will be learned,which will affect the performance of the model.This paper proposes a static pruning strategy named SRPP(Shortest Root Path Pruning).Taking the shortest dependency path from the root node to the aspect word as the center,the nodes directly related to the aspect term are reserved as the syntactic structure after pruning.At the same time,an edge-enhanced graph attention convolutional network(Edge-enhanced Graph Attention Networks,EGAT)is used to encode the dependency structure of the dependency syntax as well as the category of the edge so that more emotional information and more effective text representation can be obtained.Experiments were conducted on two public datasets to prove the effectiveness of the EGAT model and the SRPP strategy.(2)When faced with some complex sentence patterns,the static pruning method has problems such as the long distance of aspect word description dependency,the loss of corresponding description fragments and the parsing errors of syntax analysis tools,etc.,making it difficult to formulate targeted pruning strategies.Therefore,an end-to-end dynamic syntactic pruning strategt SDSP(Sematic-enhanced Dynamic Syntactic Pruning)is proposed.SDSP sets a selection gate for each node through a selection module.This module selects nodes related to aspect term,and uses fully connected semantic graphs to ensure the connectivity of the graph structure during dynamic pruning.In addition,this part also introduces a rethinking mechanism to provide more accurate semantic and syntactic information for the dynamic pruning process by introducing high-level representations into the low-level process.And the effectiveness of the proposed dynamic pruning scheme and reflection mechanism is proved by experimental results.