Research On Sentiment Classification In Dialogue Text

With the popularization of social media and the development of e-commerce platforms,sentiment analysis has been widely used in many fields such as social public opinion and commodity retail.Dialogue is the most basic way of human communication,such as chit-chat between people in daily life and the dialogue between customers and customer service agents in e-commerce platforms.The sentiment analysis in dialogue has a very important application value.Unlike opinion discussions on social media or product reviews on e-commerce platforms,the dialogue generated by the communication between different speakers contains the transmission and interaction of sentiments between speakers.In addition,the characteristics of the dialogue as a whole,such as the topic and motivation of the dialogue,also affect the judgment of the sentiment polarity of the utterances in the dialogue.In response to the above problems,this thesis studies sentiment classification in dialogue text.The main contributions are listed as follows:Firstly,for the characteristic that the sentiment polarity of utterances in dialogue de-pends on dialogue context,this thesis proposes a novel approach based on context information,to perform sentiment classification in dialogue.The model includes an encoder and a decoder and sequentially processes each utterance in the dialogue.In the encoder,for each utterance in the dialogue,a long short-term memory network with attention mechanism is used to mine relevant information from dialogue history which is incorporated into the semantic representation of the utterance.In the decoder,considering that sentiment labels of the previous utterances have an influence on the sentiment polarity of the current utterance,the probability distributions of the sentiment labels obtained from the decoded utterances are incorporated into the sentiment classification of the current utterance through a novel decoding method.The experimental results show that the proposed approach can effectively improve the performance of sentiment classification in dialogue.Secondly,for the characteristics of sentiment transmission and interaction between different speakers in dialogue,this thesis proposes a novel approach based on the role information,to perform sentiment classification in dialogue text.The proposed approach incorporates personal historical information and interaction information between different speakers when modeling the speaker's sentiment state.Specifically,the proposed approach first initializes an LSTM for each speaker in the dialogue to record the speaker's state,and then sequentially processes each utterance in the dialogue.At each moment,for the current speaker and the corresponding utterance to be classified,the attention mechanism is first used to mine context information in the current speaker's dialogue history and the interaction information in the other speaker's dialogue history.Based on this information,the state of the current speaker is updated while obtaining the final utterance representation for sentiment classification.The experimental results show that this approach is superior than other baselines for sentiment classification in dialogue text.Finally,for the influence of topic information on sentiment polarity of utterance in dialogue,this thesis proposes a novel approach based on topic information for sentiment classification in dialogue text and studies the effects of overall topic information and role-based topic information on sentiment classification in dialogue context.This approach includes a main task of sentiment classification and multiple auxiliary tasks of topic inference based on the multi-task learning framework.Specifically,in the sentiment classification task,the attention mechanism is used to model the connections between utterances in dialogue,thereby obtaining the utterance representation that incorporates dialogue context information.Auxiliary tasks of topic inference,including overall topic inference and dif-ferent role-based topic inference,are used to learn different topic information in dialogue.Finally,a gated fusion method is used to combine the topic representations from the auxiliary tasks with the utterance representation from the main task to obtain the final utterance representation for sentiment classification.The experimental results show that,besides considering the contextual connections between utterances,mining different topic information in dialogue can further improve the performance of sentiment classification in dialogue text.