【作者】 李强；

【导师】 夏鸿斌； 周丹平；

【作者基本信息】 江南大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 文本情感分析研究人们在文本中表达的情感、观点、态度。细粒度情感分析是其中的一个细分领域,相比于研究文本整体情感的粗粒度情感分析,细粒度情感分析直接对文本中的实体进行情感分析,具备更多的实用价值。细粒度情感分析的目标是提取出文本中实体,并对根据句中观点对每个实体的情感倾向进行分类。按提取的项目不同,其可以分为方面项抽取、观点项抽取及方面情感分类三个子任务。本文主要研究方面情感分类和结合三个子任务的方面情感三元组抽取任务。(1)对于方面情感分类,大部分的研究都是通过使用注意力机制及外部语义知识对全局上下文进行建模,以完成这项工作。方面的情感极性往往取决于与方面高度相关的局部上下文,但大多数模型将过多的注意力集中于全局上下文,这使得模型的参数量普遍都比较大,导致计算量也随之增大。为此,本文提出一种基于多头注意力机制的轻量化网络模型——局部与全局特征融合网络模型。根据与方面项的语义相关距离获取局部上下文,然后对局部上下文与全局上下文分别进行特征抽取,并融合两者的提取结果进行最后的分类。在三个标准数据集上进行的实验结果表明,该模型在参数量较小的情况下,取得了比其他基于方面的情感分类算法更好的结果。(2)对于方面三元组抽取,现有方法大多将这类型的任务分为多个子任务,将子任务组成流水线,完成这类任务。然而,基于流水线思想的方法在实际应用中会受到误差传播、不易使用等影响。为此,本文提出词对关系学习方法,将方面情感三元组抽取任务转化为端到端的词对关系学习任务。将句中的词对关系进行统一标注以表示所有三元组,并使用多头自注意力的注意力图作为词对关系输出,通过词对关系解码器即可获得句中所有三元组。在四个标准数据集上的实验表明,该方法性能较优。更多还原

【Abstract】 Text sentiment analysis studies the emotions,opinions,and attitudes that people express in texts.Fine-grained sentiment analysis is one of the sub-fields.Compared with the coarse-grained sentiment analysis that studies the overall sentiment of the text,the fine-grained sentiment analysis directly analyzes the sentiment of the entities in the text,which has more practical value.The goal of fine-grained sentiment analysis is to extract entities in the text and classify the sentiment tendency of each entity according to the sentiment in the sentence.According to the extracted items,it can be divided into three sub-tasks: aspect item extraction,opinion item extraction and aspect based sentiment classification.This paper mainly studies aspect sentiment classification and aspect sentiment triplet extraction task combining three sub-tasks.For aspect based sentiment classification,most of the researches accomplish this work by modeling the global context using attention mechanism and external semantic knowledge.The sentiment polarity of an aspect often depends on the local context that is highly related to the aspect,but most models focus too much attention on the global context,which makes the amount of parameters of the model generally larger,resulting in an increase in the amount of computation.To this end,this paper proposes a light weight network model based on the multi-head attention mechanism,local and global feature fusion network model.The local context is obtained according to the semantic correlation distance with the aspect term,and the feature extraction is carried out separately from the global context,and the extraction results of the two are combined to improve the classification accuracy.Experimental results on three standard datasets show that the model achieves better results than other aspect-based sentiment classification algorithms with a small amount of parameters.For aspect sentiment triplet extraction,in most of the existing methods,this type of task are regarded as some sub-tasks,and the researchers complete the task by a pipeline which composed of the sub-tasks.However,the methods based on pipeline will be affected by error propagation,and it is not easy to use the methods in practice.To this end,this paper proposes a word-pair relation learning method,which transforms the aspect sentiment triplet extraction task into an end-to-end word-pair relation learning task.The word-pair relation in the sentence is uniformly labeled to represent all triplets,and the attention map of multi-head self-attention is used as the word-pair relation output,and all triplets in the sentence can be obtained through the word-pair relation decoder.Experiments on four standard datasets show that the method has better performance.更多还原