Aspect-level Sentiment Analysis Research Based On Deep Learning

Language and writing are the foundation of the inheritance of human civilization.It enables humans to have more efficient learning and cognitive abilities,and sentiment analysis is the key technology for extracting and mining text information.How to efficiently use information technology to mine the sentiment polarity of text data,and combine actual problems to analyze and predict is of great significance to the business and academic fields.According to the granularity of the processed text,sentiment analysis can be divided into three levels:text level,sentence level,and aspect level.Among them,the text-level and sentence-level sentiment analysis mainly analyzes a kind of sentiment polarity in a piece of text or a single sentence,and the aspect-level sentiment analysis can mine the sentiment polarity of a specific aspect for different aspect words in the text.Aspect level sentiment analysis is a fine-grained sentiment classification task,and it is one of the important issues currently concerned in the field of sentiment analysis.In recent years,with the development of natural language processing technology,deep learning has become the mainstream method of sentiment analysis research.The deep learning framework that combines pre-training models and attention mechanisms has been favored by many scholars due to its excellent performance.However,most of the current research methods have problems such as single feature perception or insufficient introduction of emotional attention.Therefore,this article focuses on the introduction of multi-feature perception methods and attention mechanisms for sentiment analysis tasks.The main work of this paper is as follows:(1)Design a lightweight pre-training model.The model uses an improved MASK scheme to mask a small number of English letters and whole words that appear in the Chinese Wikipedia corpus,making the model more suitable for Chinese text processing tasks.In addition,special semantic substitute characters such as numbers are added to the dictionary,which reduces the external interference of the model in processing sentiment classification tasks.At the same time,the model reduces the depth of the pre-training network and improves the convergence speed of model training.Finally,after comparing with the benchmark model experiment,the effectiveness of the pre-training model in this article is verified.(2)Propose a sentiment analysis model of multi-feature perception.Firstly,based on the lightweight pre-training model,multiple local features of emotional text are extracted from different angles for combination,and two feature vectors of emotional corpus are obtained by pooling method in the hidden layer as the global features of the text.Secondly,the local features and global features are fused to enhance the text information association.Finally,the experimental results on public data sets show that the design of this paper is feasible The proposed method is effective.(3)Aiming at the problem of insufficient aspect word vector weights in current aspect-level sentiment analysis tasks,this paper proposes an aspect-level sentiment analysis model that integrates multi-layer attention.The model is based on a lightweight pre-training model.It first extracts the aspect word attention of multiple hidden layers in the pre-training model,and then fuses the encoded aspect information with the BERT hidden layer representation vector to form a multi-layer aspect attention,and finally The multi-layer aspect attention is cascaded with the encoded output text,thereby enhancing the long dependency relationship between emotional text and aspect words.This paper conducts experimental verification on the SemEva12014 Task4 and AI challenger 2018 data sets.The experimental results show that the use of context to interact with the target aspect weight is effective for aspect sentiment classification.