Research On Fine-grained And Cross-domain Sentiment Analysis For Text Data

With the development of Internet technology,extensive of text data has emerged in the online media,which is rich in academic research and commercial value.Sentiment analysis of text data has become an important task in the field of natural language processing.This thesis focuses on two main tasks in text sentiment analysis: fine-grained sentiment analysis and cross-domain sentiment analysis,and uses deep learning methods to analyze and study the two tasks.1.To address the interaction between aspect terms and context in fine-grained sentiment analysis and the inability to fully utilize the hidden feature information in the context,this thesis investigates fine-grained sentiment analysis from both semantic and syntactic perspectives.From the perspective of semantics,an interactive attention encoder network with local context features is proposed.We use a pre-trained BERT model to obtain embedding representations of context and aspect terms.The interaction between aspect term and context is fused using the Attention-over-Attention(AOA)module to learn the global features of context and aspect terms.The local information features around aspect terms in the text are obtained by using dynamic weighting.From the perspective of syntactics,a multi-interaction graph convolutional network model is presented.The two different interactions(syntactic interaction and semantic interaction)are fused by a gating mechanism to extract the syntactic and semantic feature information in the context.On the basis of Graph Convolutional Network(GCN),distance information(location distance and syntactic distance)features are considered.The positional distance information is fed into the GCN while the syntactic distance information and syntactic dependency information are considered to construct the adjacency matrix in the GCN.The experimental results show that the correlation model considering semantic and syntactic features can effectively identify the sentiment polarity corresponding to aspects,while the performance of the model can be further improved by using the features hidden in the context.2.For the domain adaptation in cross-domain sentiment analysis and the distribution differences between different domains,this thesis proposes a hierarchical attention transfer network model based on the union of multiple features.The document representation between contexts is extracted using a hierarchical attention network,and shared sentiment features between domains are learned using adversarial training.To reduce the distribution differences between domains,the Kullback-Leibler metric is used to constrain the distribution of features between domains.Then,the co-occurrence relationship between shared sentiment features and domain-specific features is used to extract domain-specific features by constructing auxiliary tasks.Finally,the cross-domain sentiment analysis task is performed by fusing the two different features through a gating mechanism.The experiments show that the proposed model improves the average accuracy by 0.33% over the optimal classification model and achieves optimal results on all five data sets relative to other baseline models.