Research On Aspect-level Sentiment Analysis And Sarcasm Detection For Web Text

With the rapid development of Internet technology,the amount of web text has exploded.These web text offer valuable reference for online e-commerce,government departments and potential user groups.On the one hand,only analyzing the sentiment polarity of the whole sentence can no longer meet the needs of these groups,so it is necessary to analyze the aspect-level sentiment of a comment.On the other hand,the existence of irony and sarcasm will reverse the sentiment polarity,leading to the misjudgment,which greatly affects the accuracy of sentiment analysis.Therefore,it is important to perform aspect-level sentiment analysis and sarcasm detection on web text.Existing methods typically treat aspect-level sentiment analysis as a standard text classification problem and use LSTM for it.But LSTM cannot capture local features and always ignores sentiment extraction of implicit comment objects.What's more,existing sarcasm detection systems mainly rely on contradictory word pair in sentences which have been considered as major indicators of sarcasm.When such cues are present in sentences,sarcasm detection can achieve high accuracy.However,when these indicators do not exist,the sarcasm detection task will lack the basis for discrimination and make a wrong judgment.This dissertation focuses on aspect-level sentiment analysis and sarcasm detection.Our main research work can be summarized as follows:(1)Our dissertation proposes an LSTM-CNN attention approach for aspect-level sentiment classification.Our approach combines LSTM with CNN for simultaneously leveraging LSTM to handle long-range dependencies and CNN's ability to identify local features,using CNN to compensate for the inadequacy of LSTM not capturing local features.Then we introduce the attention mechanism to focus on the important information about the aspect embedding by assigning higher weight values so that our model can get sentiment extraction of implicit comment objects.Our method performs well in Chinese restaurant comments and Sem Eval2014 in English.(2)For research work on the sarcasm detection,we propose a sarcasm detection model based on intra-sentence representations and contextual information such as user embeddings.Firstly,our model uses self-attention to calculate the similarity between words and words,modeling intra-sentence contrast.It not only solves the problem that standard LSTM can't capture long-distance dependence,but also find the inconsistency between word pairs more explicitly.Next,we use a pre-trained CNN mode to learn users' personality indicators and use Paragraph Vector to learn user's writing styles,which are then fused into a more comprehensive user embeddings,providing the model with a basis for judging implicit sarcasm.We conduct extensive experiments on four benchmark datasets from Reddit and the Internet Argument Corpus and our model has better prediction effect compared with the other four methods.