Deep Neural Networks Based Research On Aspect-based Sentiment Analysis

The key step of the Aspect-Based Sentiment Analysis(ABSA)task is how to obtain the aspect-specific sentence representation given the context sentence and the aspect terms.Recurrent Neural Networks(RNNs)armed with the attention mechanism can measure the relevance between the aspect terms and the words in the context sentence in an end-to-end manner,thus achieves the state-of-the-art performance recently.However,there are also some limitations in the current research works: on the one hand,current research works based on attention mechanism and recurrent neural networks can not model the potential dependency relationship among the aspects appearing in the same sentence,also,current attention based works may introduce some noise and attend irrelevant words when dealing with context selection;on the other hand,existing annotated datasets for the ABSA task are somewhat sufficient,which limits the further exploration of deep neural networks based research work.This thesis targets at the aforementioned issues.On the one hand,it proposes some improvements on current attention based research works and futher proposes a novel context selection mechanism different from attention mechanism;on the other hand,it introduces the transfer learning to alleviate the issue that current annotated datasets for the ABSA task are not sufficient,and it also evaluates these improvements by extensive experiments on the benchmark datasets.The main contributions are as follows:1.Targeting at the issues that current mainstream research works based on attention mechanism and recurrent neural networks can not model the potential dependency relationship among the aspects appearing in the same sentence,and do not take the overall sentence representation into the consideration,this thesis introduces the overall sentence representation into current attention mechanism and designs a tailor-made auxiliary task to guide the learning of the sentence representation;on the other hand,to model the potential dependency relationship among the aspects appearing in the same sentence,this thesis proposes to adopt the attention mechanism to fuse the different aspect representation.Extensive experiments on Semeval2014 datasets validate the efficiency of the method.2.Targeting at the issues that current attention mechanism based work may introduce some noise and the follow-up research still focuses on the further improvement under the framework of attention mechanism,this thesis proposes a novel Hierarchical Gate Mechanism for context selection.Meanwhile,this thesis leverages the POS tag and position information to enhance the selective capacity.This thesis proposes Hierarchical Gate Memory Network,it utilizes hierarchical gate mechanism for the context selection to obtain the aspect-specific sentence representation,and then adopts Convonlutional Neural Network for the polarity prediction.Extensive experiments on Sem Eval 2014 and Twitter datasets validate the efficiency of the method.3.Targeting at the issue that existing annotated datasets for the ABSA task are somewhat sufficient,this thesis introduces the transfer learning to alleviate the limitation.Specifically,it introduces the language model pre-trained on massive text corpus into the ABSA task,furthermore,this thesis integrates the domain-specific posting training tricks to improve the model performance.Extensive experiments on Sem Eval 2014 and Twitter datasets validate the efficiency of the method.