Research On Aspect-Level Sentiment Analysis Based On LSTM And Multi-Task Neural Network

Aspect-level sentiment analysis is a hot topic in the field of natural language processing.Aspect level sentiment analysis is a fine-grained sentiment analysis,which mainly includes two important sub-tasks: aspect term extraction and aspect sentiment classification.Some methods solve only one subtask individually,and some solve two subtasks in the form of a pipeline.These methods are easy to implement,but the connection between the two subtasks cannot be fully considered,nor can the interaction between the two subtasks be explicitly modeled.The aspect term and the comment's emotional information about the aspect term are both important.Therefore,this paper proposes two neural network models that simultaneously solve aspect term extraction and aspect sentiment classification.The main work of this article is as follows:(1)This paper proposes a joint model of fusion attention mechanism,which can capture important emotional information in the context,and can simultaneously solve the problem of aspect term extraction and aspect sentiment classification.Based on the bidirectional long-term and short-term memory network,combining two attention mechanisms,not only considering content information,but also adding location information,calculating the relevance of each context word and aspect term.The attention mechanism can model the internal structure of the sentence and fuse the emotional information in the sentence.In order to make the model better fit and optimize the parameters in the training process,improve the model's effect,this paper optimizes the loss function of the neural network model.(2)This paper proposes a guided attention and interactive multitasking network model.This model is based on a gated recurrent unit network implementation and can simultaneously solve both aspect term extraction and aspect sentiment classification.The interactive component is implemented through a gating mechanism,explicitly modeling the interactive relationship between aspect term extraction and aspect sentiment classification,and passing messages between two tasks.The guided attention mechanism can introduce the prediction results of opinion word extraction into aspect sentiment classification tasks to guide aspect sentiment classification task pays more attention to opinion words with emotional polarity.The data set of aspect-level sentiment analysis is small,and the advantages of neural network models cannot be fully utilized.Therefore,this paper adds document-level auxiliary tasks to migrate semantic and emotional knowledge in documents.(3)This article conducted extensive experiments on the two proposed models.In this paper,the F1 score evaluation model is used to solve the effect of aspect term extraction and aspect sentiment classification problem,and the F1 score and accuracy are respectively used to evaluate the effect of the model to solve the aspect term extraction and aspect sentiment classification respectively.The experimental results show that,for the F1 score of solving aspect term extraction and aspect sentiment classification problems on two data sets at the same time,the joint model of the fusion attention mechanism is improved by 0.97 and 0.79 percentage points compared with the optimal baseline model,guided attention and interactive multitasking network model is 1.14 and 1.97 percentage points higher than the optimal baseline model,respectively.Moreover,the effect of the model alone in solving aspect term extraction or aspect sentiment classification is somewhat improved compared to the baseline model.