Research And Application Of Knowledge Enhanced Aspect-based Sentiment Alanysis

Aspect-based sentiment analysis,which aims to predict the sentiment polarity for given aspects of a sentence,is one of the frontiers of research in natural language processing.The two most important sub-tasks are aspect term sentiment analysis and aspect category sentiment analysis,the main difference being whether the predicted object is explicitly present in the sentence or not.In recent years graph neural networks have achieved superior results in this area.However,most of the methods have shown limited performance improvement,mainly due to insufficient exploitation of external knowledge,poor construction of the relationship between contexts and aspect categories in sentences,and lack of modelling of the complementarity of syntactic structure and semantic relations.To address the above-mentioned problems,our works are as follows.1.For the aspect term sentiment analysis task,we propose a sentimentaware dual-channel graph convolutional neural networks model,where the two channels are knowledge enhanced syntactic channel and knowledge enhanced semantic channel.In the first channel,matrix constructed from dependency trees are first used to capture the syntactic structure of the sentence.Then,the naive matrix is enhanced by both the sentiment score from words in SenticNet and the lexical knowledge of part-of-speech;In the second channel,the word vectors trained based on the ConceptNet are fused with the hidden representations encoded by the Bi-LSTM to better utilize the knowledge semantics of words.Finally,the problem of insufficient use of external knowledge in both syntactic and semantic channels is solved.2.For the aspect category sentiment analysis task,we propose a knowledge-enhanced multi-channel graph convolutional neural networks model,in which the multi-channels are knowledge channel,syntactic channel and semantic channel,respectively.In the knowledge channel,we use the WordNet-based similarity function to calculate the concept-based similarity between the aspect category and the context,based on which,we construct the similarity matrix concerning the aspect category.This addresses the issue of unreasonable capture and exploitation of the relationship between the aspect category and the relevant words in the sentence;In the syntactic and semantic channels,dependency analysis and the self-attention mechanism are used to construct the corresponding adjacency matrices for GCN to generate syntactic and semantic features,which are then fused by a designed attention-based fusion module.This solves the lack of leveraging the complementarity of syntactic structure and semantic information.3.A review-oriented aspect-based sentiment analysis system is designed and implemented,which consists of a user information management module,a model management module and a sentiment analysis and visualization module.The system is platform-independent and user-friendly,allowing users to visually configure parameters,train models and perform aspect-based sentiment analysis on texts.