Research On Fine-grained Sentiment Analysis Based On Deep Learning

Fine-grained(or aspect-based)sentiment analysis takes an important role in the field of natural language processing.Compared with classic sentence-level or document-level sentiment classification,it can identify the sentiment polarity(e.g.,positive,negative,neutral)of the given aspect words.The core of this task is to calculate the correlation between the aspect words and other words in the sentence and obtain the aspect-specific vector.Most of the classic fine-grained sentiment analysis solutions are based on long short term memory networks or convolutional neural networks,and at the same time introduce the attention mechanism to generate corresponding text representations.Although a large number of researches have demonstrated the effectiveness of these methods,there are still some shortcomings,such as inability to efficiently learn syntactic knowledge,failure to distinguish the differences between semantics and syntactic relations,failure to deal with the particularity of aspect words and the need of a large amount of training data.The main work of this paper is as follows:(1)The baseline model based on LSTM and the graph convolutional neural network model based on dependency trees are proposed.The comparison experiments with some classic models show that the introduction of the graph convolutional neural networks can efficiently learn syntactic information.It makes up for the shortcomings of only using grammatical information,and effectively improves the performance on the benchmark datasets.(2)Aiming at the different relationships between aspect words and other words,a neural network model based on specific relationships is proposed,including a gating module and a dynamic dependency tree module.From the perspective of semantic relevance and syntactic relevance,these two modules improve the calculation of the relevance of aspect words and other words in the sentence.Comparison experiments and ablation experiments show the important contribution of this method to performance.(3)Aiming at the problem of insufficient emphasis on aspect words in classic methods,a neural network model based on aspect word enhancement is proposed,including aspect word enhancement module and kernel pooling module.These two modules respectively solve the problem of not paying enough attention to the particularity of aspect words in the process of graph convolution and calculating the semantic relevance of aspect words and other words with single granularity.Comparison experiments and ablation experiments show the important contribution of this method to performance.(4)Aiming at the problem of the scarcity of labeled data,Zero shot and Few shot methods based on pre-trained language models are proposed.This method solves the problem of insufficient static word embedding expression ability by introducing a pre-trained language model,and at the same time introduces few-shot learning and zero-shot learning to effectively solve the problem of large data requirements for classic models.The comparative experiment shows the important contribution of this method to performance.