【作者】 张颖；

【导师】 郑建国；

【作者基本信息】 东华大学 ， 管理科学与工程， 2020， 硕士

【摘要】 情感分类是自然语言处理的经典任务之一,具有广泛的应用场景和重要的应用价值。面向细粒度的要素级情感分类任务是预测句子在特定要素上所要传达的情感倾向,一般可以拆分成提取要素和基于要素进行情感分类两个子任务。大部分用户在发表评论及对某一要素做出情感表达时,并不会将要素词直接放入评论文本中。因此,本文以要素类提取和要素类情感分类作为研究任务,基于深度学习方面知识和文本挖掘相关技术,改进相关神经网络模型。第一,对于要素类提取任务,引入自注意力机制,提出结合自注意力机制的神经网络模型,使得评论文本中的每个位置的表示都具有全局的语义信息,模型能学习到带有上下文语义信息的词嵌入;引入相对位置表示,提出结合基于相对位置表示的自注意力机制的神经网络模型来模拟词语的顺序,以修正自注意力未考虑时序信息的缺点。第二,对于要素类情感分类任务,引入结合要素类嵌入的基于门控机制的卷积神经网络模型,利用更少的参数实现对输入的并行化训练,分别提取要素特征和情感特征;针对该模型的输入层所采用的要素类嵌入具有脱离上下文的缺点,联动要素类提取实验,引入基于上下文优化后的词向量和要素类嵌入;考虑到该模型的池化层所采用的最大值池化操作对特征信息的损失较大,引入K大值池化以减少特征的损失。在SemEval标准数据集上的消融实验结果表明,将自注意力机制融合进神经网络模型提取出的要素类具有较高的连贯性和可解释性,将优化后的词向量和要素类嵌入对门控卷积神经网络模型的输入进行初始化能有效提高模型的分类精度。更多还原

【Abstract】 Sentiment classification is a classic task of NLP having important application value and many application scenarios.Aspect-level sentiment classification is fine-grained with the goal being to predict the sentiment polarity of a sentence to be conveyed on a specific aspect.Due to the fact that most users comment on aspects without explicitly naming the aspect terms,aspect category sentiment classification digs deeper information being closer to actual needs compared to the other.Based on deep learning knowledge and text mining technology,this paper takes aspect category extraction and aspect category sentiment classification as research tasks and improves related neural network models.For the task of extracting aspect category,ABAE model utilized the characteristic information of the aspects to improve the coherence.Considering ABAE model was insufficient and still had room for improvement,this paper introduces self-attention mechanism and proposes SABAE model to make the representation of each position of the review text have global semantic information so that the model can learn word embedding with contextual semantic information,and introduces relative position representation to propose SABAEwith RPR model to simulate the order of words fixing the problem of self-attention mechanism not considering timing information.For the aspect category sentiment classification task,GCAE model trained in parallel was proposed,which can extract aspect category features and sentiment features separately.Considering the aspect category used in GCAE model is randomly initialized and context-free,this paper correlates aspect category extraction and aspect category sentiment classification,introduces trained aspect embedding matrix and word embedding matrix into GCAE model.Besides,this paper also introduces k-max pooling trying to improve the sentiment classification accuracy.Ablation experiment results on standard Sem Eval dataset showed that the aspect categories extracted by the SABAE model have higher coherence and better interpretability,adoption of corresponding word embeddings and aspect embeddings to initialize the inputs of GCAE model is effective to improve the accuracy of sentiment classification.更多还原