【作者】 王婷婷；

【导师】 杨静；

【作者基本信息】 华东师范大学 ， 计算机科学与技术， 2020， 硕士

【摘要】 随着社交网络与电商平台的兴起与发展,越来越多的人乐于在线上发表有关于购物、旅游、服务等领域的评论,这些带有个人主观情感态度的文本十分具有挖掘价值。在情感分析领域,文档级和句子级的情感分析只能挖掘整体的情感极性信息,但无法分析用户对于文本中各个实体或属性的情感及意见,因此方面级情感分析应运而生,其中方面可为实体或属性。近几年深度学习的发展与广泛运用,也为方面级情感分析任务提供了新的解决方法。本文主要关注基于深度学习的方面级情感分析问题,针对该研究内容现存的一些问题,尝试提出高效的解决方案。首先,目前的方面级情感分类模型大都基于整个句子进行预测,而这会引入与方面情感无关的上下文噪音信息。针对这一问题,本文提出了一个基于强化学习的神经网络模型来自适应地抽取与方面相关的描述片段,并基于该片段完成后续的情感分类任务。实验结果表明,本文所提方法可以有效地提取描述该方面的片段,从而实现情感分类性能的提高。此外,通过案例分析,可以直观地解释为什么本文所提模型会适用于方面级情感分类。其次,针对目前的方面级情感分析模型在建模过程中没有很好地利用方面抽取、情感词抽取和方面级情感分类三个子任务之间的关系这一问题,本文提出了一个端到端的方面级情感分析模型。该模型不但可以学习到子任务之间的关系,互相促进,还能解决一个方面对应多个情感词和一个情感词对应多个方面的问题。最终,本文在三个英文标准数据集上进行实验来验证模型的有效性。实验结果表明,本文提出的模型能取得比现有模型更好的效果。最后,为了探究所提模型的应用性与泛化能力,本文开发了一个方面级情感分析系统。该系统包括在线Demo、应用展示和数据分析模块。通过在线Demo模块,用户能在线完成方面级情感分析;应用展示模块提供了评论分类的功能,可以帮助用户了解评论中各实体属性的特性,还可用于算法模型的样例研究;而数据分析模块可以快速完成方面、情感词、情感极性等分布信息的统计和可视化,这些信息能帮助用户更好地了解数据的分布情况以及使用者对于产品各个方面的意见。更多还原

【Abstract】 With the development of social networks and e-commerce platforms,more and more people are willing to share online comments about shopping,tourism,services and other fields.Those texts with personal and emotional attitudes are valuable for mining.In the field of sentiment analysis,document -level and sentence -level sen timent analysis only concentrate on sentiment polarity upon the whole text,while the detailed sentiment or opinions on certain entities or attributes in the text are ignored.Thus,the task of aspect- level sentiment analysis is proposed,here the aspect referes to the entity or attribute.Deep learning has obtained great success in the artificial intelligence fields and it also provides new solutions for aspect -level sentiment analysis tasks.This paper proposes efficient solutions to some existing problems in aspect -level sentiment analysis with deep learning.Firstly,most of the existing aspect-level sentiment classification models predict the sentiment based on the entire sentence,which might introduce noise information irrele-vant to the sentiment of aspect.This paper proposes a novel neural network model based on reinforcement learning framework to adaptively extract aspect-related segments and apply them for aspect-level sentiment classification.The experimental results show that the proposed method can effectively extract the segment oriented to the aspect,and thus improve the classification performance.In addition,case studies are used to intuitively understand why the proposed model is suitable for aspect-level sentiment classification.Secondly,since the existing aspect-level sentiment analysis models are negligent in utilizing the relationship among aspect term extraction,opinion term extraction and aspect-level sentiment classification,this paper proposes an end-to-end aspect-level sentiment analysis model.This model can not only learn the relationship between sub-tasks,but also solve the problems that one aspect with multiple opinion words and one opinion term for multiple aspects.In the end,experiments are conducted on three stan-dard English datasets to verify the performance of the model.Experimental results show that the proposed model can achieve comparable or better results than strong baseline models.Finally,to explore the applicability and generalization of the proposed model,this paper develops an aspect -level sentiment analysis system.The system consists of on-line demo,application display and data analysis.Fisrtly,the online demo module can help users complete aspect-level sentiment analysis online.Furthermore,the applica-tion display module provides the function of reviews classification,it can help users to understand the characteristics of the product and conduct sample studies of proposed model.Eventually,the data analysis module can quickly complete the data analysis and visualization,including distribution information such as aspect terms,opinion terms and sentiment polarities,it can be used to better understand the data distribution and users’ attitudes on every aspect of the product.更多还原