Research On Aspect-base Sentiment Classification Based On Deep Reinforcement Learning

With the popularization of the Internet,instant messaging,network news,online shopping and other activities have produced a large amount of short text,which contains a large amount of information closely related to people's production and life,which can greatly improve user experience and has high commercial value.However,the network short-text has the characteristics of short text length,sparse features,and lots of noise.The coarsegrained text classification task has mature models and methods,but the more fine-grained aspect-based sentiment classification task still has room for further improvement.In view of the characteristics of short text,such as sparse features and lots of noise,this thesis proposes to use deep reinforcement learning to extract the information related to aspects in the text for more fine-grained aspect emotion classification tasks.The main work contents and contributions are as follows:(1)In order to solve the problem that syntactic dependency trees may generate noise or contain information irrelevant to aspects,we propose Reinforced Dependency Graph Aspect-based Sentiment Classification(RDGSC),a reinforced dependency graph model for aspect-based sentiment classification.We use a strategy-based deep reinforcement learning algorithm to generate a reinforced syntactic dependency graph related target aspect by modeling the node information,the tag of contextual and relationship types of the syntactic dependency tree.Then,the graph attention network was used to make the aspect nodes fully integrate the information of the neighbor nodes,and then the retrieval-based attention mechanism is used on the context information to obtain a more refined final representation.At the same time,the classification result is used as delayed rewards to guide the updating of the policy network.It is worth noting that the proposed model is trained in an alternate way.Extensive experiments on five real-world datasets show that RDGSC approach performed significantly better in the task of aspect-based sentiment classification compared with strong baseline methods.(2)In order to solve the problem that some words in contextual information would be noisy or irrelevant to the target aspect,we develop a novel reinforcement learning framework for aspect-based sentiment classification,called Aspect-Aware Reinforcement Learningbased framework(A2RL).We introduce an aspect-guided reinforcement representation module(AG-RRM)to identify whether a word contributes to the final representation of the input text with respect to a target aspect.Then,an aspect-guided graph convolutional network(AG-GCN)is developed to generate the refined final representation over the reinforced representation outputted by AG-RRM by utilizing a retrieval-based attention mechanism and make the classification.A delayed reward is designed to guide the learning of the policy where we encourage the model to capture more irrelevant words as well as hold a small sentiment classification loss.It is worth noting that the proposed model is trained in an alternate way.Experimental results on five widely used datasets,i.e.,Twitter,Lap14,Rest14,Rest15 and Rest16,show that our proposed A2 RL can effectively generate aspectaware reinforced word representation and model the local dependency structure information with a multi-layer GCN upon the generated representation.