Sentiment Analysis For Domain-specific Review Texts

With the booming e-commerce industry and various social media,a large number of users have posted product reviews on Internet platforms.The rise of social media platforms has provided a broad data base for sentiment analysis tasks.By using natural language processing techniques,reasonable handling of comment data on the Internet can create enormous commercial and social value.Take into account this,this paper takes the review texts in a specific field as the research object,and focuses on this task.Because the sentiment expressions in the network are some implicit,some unrestrained,some general and some specific,each has its own characteristics,therefore,the task of sentiment analysis has become a complex problem in nat ural language processing.Most of the current sentiment analysis systems train neural network models through labeled data,ignoring the characteristics of the sentiment analysis task itself,which leads to the following problems: First,the user’s sentiment expression usually has different characteristics when facing different fields,but the field characteristics are often ignored and not fully utilized;Second,human language expressions are usually concise and rich,and many of the information hidden in sentences is self-evident to human beings.However,the comment sentences in the network do not supplement this kind of knowledge,so it is difficult for machines to fully understand human language;Third,the traditional sentiment classification framework only encodes the textual semantic information of the comment sentence,ignoring the syntactic structure information,but these features can also provide reference for aspect-level sentiment classification models.In view of the above deficiencies,this paper conducts in-depth research from two perspectives,the utilization of domain features and the introduction of prior knowledge,in order to achieve knowledge enhancement of existing sentiment classification research.At the same time,the syntactic dependency information and the interaction between the aspects and the context in the review sentence are introduced in both research schemes,which effectively improves the accuracy of the model.Specifically,the main work of this paper is as follows:(1)In solving the aspect-level sentiment classification problem,this paper firstly addresses the problem that the existing word vector model is not optimised for a specific domain,and proposes a strategy to fuse the generic representation of the review text with its domain features through convolutional neural networks to achieve a double embedding of generic representation and domain features.Secondly,in order to solve the shortcoming that the generic text representation method cannot obtain the conceptual a nd background knowledge related to words,this paper improves the feature representation capability of word vectors and increases the diversity of features by introducing the entity concept knowledge in Microsoft Concept Graph.The experimental results sho w that a priori knowledge such as domain features and entity concepts can effectively improve the classification results of the model.(2)Considering part-of-speech information and the syntactic structure of commentary sentences during encoding is another channel to enhance knowledge.In this paper,the part-of-speech information in the text and the syntactic relationship between words are obtained by analyzing the dependency syntax,and the dependency syntax graph is constructed by taking words as nodes a nd the relationship between words as links.In the feature extraction framework after the word representation stage,this paper fully extracts the part-of-speech and syntactic information of the text through the GCN,the important influence of the interaction between the evaluation object and the context on the task is also considered.Related experimental results demonstrate that syntactic dependencies have a positive impact on fine-grained sentiment classification tasks.