Design And Implementation Of Aspect Based Sentiment Analysis System Based On Deep Learning

Aspect based sentiment analysis is a key research in natural language processing.This task aims to mine the evaluation aspect from the subjective review text and analyze the sentiment polarity of the evaluation aspect.Due to the complexity of syntactic structure and semantics and the limited ability of existing models to process complex sentences,the following problems and challenges still exist in aspect based sentiment analysis now:1)Capturing the context information of target aspect mixed with noise information leads to a decrease in task accuracy,and focusing on one single subtask leads to a general dilemma when the model transfers to other domain because of the poor performance in the remaining subtasks;2)Insufficient understanding of complex triplet and the lack of exploration of the interaction between subtasks bring models poor interpretability.These relationships between subtasks encodes the collaborative signals between subtasks,but they have not been fully utilized.Secondly,ignorance on opinion term extraction makes it difficult to fully depict the emotional state expressed by the entire sentence;3)There is still a lack of fine-grained sentiment analysis systems and algorithm research platforms in specific fields,due to the complexity of tasks and the diversity of scenes,there are few related platforms that can meet the demands of fine-grained sentiment analysis tasks.Based on the current problems and challenges faced aspect based sentiment analysis,the research work carried out in this paper is as follows:1)Proposed and implemented a multi-task joint learning algorithm based on multi-head self-attention mechanism.Under a unified framework,aspect term extraction and sentiment polarity classification are training jointly,and local context and non-local context are processed based on semantic relative distance.In the comparative experiments with different baseline models,the effectiveness of the proposed method is proved.2)Proposed and implemented a triplet joint extraction algorithm based on relation enhancement.The opinion term extraction is added to the joint learning framework for triplet extraction,and the joint extraction is constrained by the interactive information between the subtasks.Experimental results on benchmark datasets demonstrate the effectiveness of the method,which can provide a complete solution for fine-grained sentiment analysis tasks.3)Designed and implemented a fine-grained sentiment analysis system.The system provides services such as model management,dataset management,dataset annotation,and model evaluation for algorithm developers.