【作者】 刘妍；

【导师】 侯晓荣；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 情感分析作为自然语言处理的一项经典的任务,在舆情监控、书评影评统计、商品和服务用户反馈分析等方面有着广泛的应用和重要的作用。从任务关注的目标粒度大小,情感分析可分为粗粒度的句子级/文档级的情感分析和细粒度的方面级的情感分析,两种情感分析都有着重要的研究价值。近年来随着自注意力机制的提出,两种情感分析都有了性能上的突破。但由于情感分析是特殊的自然语言理解任务,自然语言中的情感表达通常不仅有直接简单的形式,还有很多委婉曲折或夹杂着无关信息的表达形式。对于包含着多重否定、对比转折等这类复杂情感表述的文本,挖掘文本中所包含的情感是一项有难度的任务。此外,在方面级情感分析任务中,对于包含多个方面词的文本,找到与方面词相关的情感表述,并理解其情感也是一个值得研究的问题。本文分别针对以上两个问题,在句子级/文档级情感分析任务上和方面级情感分析任务上,提出了有效的模型,具体如下:1.在句子级/文档级的文本情感分析任务上,现有的主流模型仅对文本进行上下文相关的偏词义的计算与推理,不能准确理解受语法影响的复杂情感。针对这类问题本文提出了从语法信息和词义信息结合的角度构建情感分析模型,从而改善模型对于复杂情感的理解。实验证明,提出的模型能够更好地理解复杂情感,在IMDB、SST-2和Yelp-2数据集上的准确率分别达到了96.5%、96.5%和98.8%。2.在方面级的情感分析任务中,现有的方法根据句法依赖关系构建图模型,通过图卷积定位与方面相关的情感表述,从而完成对方面的情感分析。有些依赖关系对识别与方面相关的情感表述明显有利,但现有模型对所有依赖关系都赋予相同权重,没有充分利用句法依赖关系中的信息。此外,现有模型存在图卷积层过多会对方面级情感判别引入干扰信息,过少则不能全面捕获相关的情感描述造成理解错误的问题。针对这些问题,本文提出了区别权重图卷积方法,帮助方面词与其相关情感的表述快速建立稳固的联系,并利用成分树提供的知识泛化情感表述视野,捕获到全面而不冗余的与方面相关的情感表述,并结合复杂情感分析的经验,实现方面级情感的分析。实验表明本文提出的方法在多个数据集上表现突出,在TWITTER数据集以及Semeval 2014-2016年的细粒度情感分析数据集LAP14、REST14、REST15、REST16中,准确率分别达到了78.1%、80.2%、87.0%、84.4%和92.5%,且消融实验验证提出的改进方法是有效的。更多还原

【Abstract】 Sentiment analysis,as a classic task of natural language processing,applied to a wide range of fields.Sentiment analysis plays an important role in public opinion monitoring,such as book review and film review statistics,product and service user feedback analysis,etc.From the granularity of the target concerned by the task,sentiment analysis can be divided into sentence level or document level sentiment analysis and aspect level sentiment analysis,both of which have important research value.In recent years,with the introduction of the self-attention mechanism,both kinds of sentiment analysis have made breakthroughs in performance.However,as a special task of natural language understanding,sentiment analysis not only requires the model to understand direct and simple expressions,but also requires the model to understand euphemistic and tortuous expressions mixed with irrelevant information.It is a difficult task to understand sentiment in texts containing complex sentiment expressions such as multiple negations and contrastive turns.In addition,in the aspect-based sentiment analysis task,for the text containing multiple aspects,finding the sentiment expression related to the aspect and understanding its sentiment is also a problem worth studying.In this thesis,we propose effective models for the above two problems,on the sentence-level or document-level sentiment analysis and on the aspect-level sentiment analysis,as follows:1.In the sentence-level or document-level text sentiment analysis,the existing mainstream models only calculate and infer the context-sensitive partial word meaning of the text,and cannot accurately understand the complex sentiment affected by syntactic.For this kind of problem,this thesis proposes to construct an sentiment analysis model from the perspective of combining syntactic information and semantic information,so as to improve the understanding of complex sentiment in the model.Experiments show that the proposed model can better understand complex emotions,with accuracy rates of96.5%,96.5%,and 98.8% on the IMDB,SST-2,and Yelp-2 datasets,respectively.2.In the aspect level sentiment analysis task,the existing methods use syntax dependency to construct graph model,and locate the sentiment expression related to the aspect word through graph convolution,thereby complete the aspect sentiment analysis.Some dependencies are obviously beneficial to the recognition of sentiment expressions related to aspect words,but the existing models assign the same weight to all dependencies,which does not make full utilize of the information in syntactic dependencies.In addition,existing models have the problem of introducing interference information into aspect level sentiment discrimination if there are too many graph convolutional layers,while if there are too few,they cannot fully capture relevant sentiment descriptions,resulting in misunderstandings.In response to these problems,this thesis proposes a differential weight graph convolution method to help quickly establish a stable connection between aspect words and their related emotions.It also utilizes the knowledge provided by component trees to generalize the perspective of sentiment expression,capturing comprehensive and non redundant sentiment expressions related to aspects.Combined with the experience of complex emotion analysis,it achieves aspect level emotion analysis.The experiment shows that the method proposed in this article performs outstandingly on multiple datasets,with accuracy rates of 78.1%,80.2%,87.0%,84.4%,and 92.5% respectively in the TWITTER dataset and Semival’s finegrained sentiment analysis datasets LAP14,REST14,REST15,and REST16 from 2014 to 2016.The ablation experiment verifies that the proposed improved method is effective.更多还原