Aspect-level Sentiment Analysis Based On Deep Learning

Aspect-level sentiment analysis aims to analyze the sentiment orientation towards aspect terms from unstructured text.Heretofore,the task mainly includes three research problems: aspect term extraction,aspect-level sentiment classification and aspect term-polarity coextraction.In this thesis,we mainly studies the latter two research problems.The goal of aspect-level sentiment classification is to identify the sentiment polarity of a specific aspect term in context.The existing methods adopt multi-attention mechanism to alleviate the distraction problem,which can improve the performance of methods.However,these methods are unable to deeply integrate the semantics of aspect terms and its context,and ignore previous attention information.In order to solve these problems,this thesis proposes a novel aspect-level sentiment classification neural network based on dynamic attention gated recurrent unit(DAGRU).This model considers the previous attention information through the dynamic attention mechanism.The encoding layer adopts the union encoding scheme to encode sementics.And then these semantics are deeply integrated with dynamic attention.The experimental results on SemEval 2014 datasets(Laptop and Restaurant)show that our approach achieves the improvement of 1.73% in the accuracy rates in comparison to previous best model based on multi attention.Aspect term-polarity co-extraction aims to extract both aspect terms and their polarities in a sentence through an end-to-end model.Most of the existing works usually achieve this task through Collapsed based models.However,these methods are unable to obtain the interaction information between aspect term and sentence,and ignore associations between different aspect terms.To address these problems,this thesis designs a novel joint decoding scheme of aspect terms and their polarities,and proposes an end-to-end model based on encoder-decoder architecture.The joint decoding scheme adapts the encoder-decoder structure to the task,and achieves a complete prediction.On the other hand,the encoder utilizes message passing to consider the relationship between aspect terms.And the decoder adopts attention mechanism to achieve the interaction between aspect terms and sentence.The experimental results demonstrate that the proposed frameworks achieves F1 scores of 64.28%,74.76% and 54.52% on three benchmark datasets respectively,which outperforms state-of-the-art baselines.To sum up,for aspect-level sentiment classification,this thesis proposes an aspect-level sentiment classification neural network based on dynamic attention gated recurrent unit to fuse the semantic of aspect terms and context and consider the previous attention information.For the aspect term-polarity extraction,an end-to-end neural network based on encoder-decoder architecture is proposed to employ interaction between an aspect term and context and model the correlation between aspect terms.And a novel joint decoding scheme is designed to adapt the model to task and solve the problem of predictive integrity.The experimental results show that these proposed methods are effective.