Research On Target-based Sentiment Analysis Algorithm Based On Self-attention Mechanism

Target-based sentiment analysis is a kind of fine-grained sentiment analysis research,which is a popular direction in the field of natural language processing.The research aims to mining fine-grained elements,products,services,events and other targets and their opinions and sentiments from massive Internet texts,including three subtasks of target extraction,opinion extraction,and target-specific sentiment analysis.Research that addresses only one of these subtasks requires the assumption that the results of the other tasks are already given,however,there is no perfect model given results in practical applications,thus there is error propagation and limitations.The targets in real scenarios are usually not given and need to be automatically extracted not only for the targets but also for predicting their sentiment polarity and even for extracting the corresponding viewpoint opinions.Therefore,solving two or three of these subtasks at the same time can obtain a more application-worthy subtlety element.In this thesis,we focus on two tasks,target extraction and sentiment prediction,and target sentiment triplets extraction,to improve the accuracy and completeness of target-based sentiment analysis algorithms.Target extraction and sentiment prediction,which aims to simultaneously handle two subtasks of target extraction and target-specific sentiment analysis,identifies the target objects in the text and their sentiment polarity.Target sentiment triplets extraction,which completely solves three subtasks,extracts the target,opinion,and sentiment triplets.The work on target extraction and sentiment prediction,and target sentiment triplets extraction in target-based sentiment analysis studies mainly suffers from two deficiencies: on the one hand,subtasks have dependencies on each other,models do not make full use of subtask information,and there are difficulties in learning different types of subtasks at the same time.On the other hand,most studies use techniques such as long and short-term memory networks and single-layer self-attention mechanisms to extract feature information,which cannot capture the complex internal relationships of the input text.To address the above difficulties,this thesis conducts research on target extraction and sentiment prediction,and target sentiment triplets extraction,respectively,with the following work.1)Solving the task of target extraction and sentiment prediction: Existing research efforts are unable to fully utilize subtask information to assist in the simultaneous generation of goals and sentiments,and in addition,are generally based on long-and short-term memory networks that cannot capture the internal relationships of input sentences for sentiment analysis prediction.In order to solve the above problems,this thesis proposes a Dual-assist Network based model for Target extraction and Sentiment prediction.The dual-assist network is designed to enhance target extraction and sentiment recognition,and to alleviate the difficulty of the model to learn two different subtasks simultaneously.Further,by introducing a direction-aware Transformer as a feature extractor to effectively align the intrinsic connections of multiple target objects and sentiment words.The model is experimented on three publicly available datasets and shows a significant improvement over the benchmark model.2)Solving the target sentiment triplets extraction task: Existing mainstream research approaches focus on sequential labeling design,and the models do not fully utilize subtask information when generating goal,opinion,and sentiment polarity simultaneously.In addition,they commonly use a single-layer self-attentive mechanism,and their ability to learn the connections between words in sentences is weak,and they cannot learn the complex multifaceted relationships between words in input sentences due to multi-task complexity.To solve the above problems,this thesis proposes a Two-stage Learning-based model for target sentiment Triplets Extraction.The two-stage learning mechanism enhances the model's ability to extract target objects and opinion terms,provides guidance information for predicting sentiment polarity,and contributes to the simultaneous generation of targets,opinions and sentiments.Further,the complex connections between words in the input sentences are learned by introducing the Transformer network.The model shows some improvement in effectiveness when comparing the benchmark model on four publicly available datasets.