Research On Emotion Analysis For Natural Spoken Language Text

With the advent of the Internet age,people's lives have undergone tremendous changes,and the situation of "one network connecting the world" has begun to take shape,and communication among people has no longer restricted by time and space.People are beginning to use colloquial text to communicate and comment on social media and other Internet platforms;as a result,they can express their views and opinions more freely and fully.A large amount of natural spoken language data,which usually contains the publishers' emotion tendency,has appeared on the Internet.By analyzing these natural spoken language text data,companies can obtain important commercial values such as user sentiment and user opinions,and the government can grasp the direction of public opinion and make a series of government decisions.This thesis focuses on the problems that extracting semantic emotion information insufficiently and being difficult to analyze sentiment directly in low resource corpus under the background of text sentiment analysis for natural spoken language.The main research results are as follows:(1)In view of the insufficient extraction of semantic information from existing text emotion recognition models,a multi-information neural network(AMINN)based on the attention mechanism,which consists of pre-trained word vectors,convolutional neural networks(CNN),two-way long and short-term memory network(BiLSTM)and attention mechanism,is proposed in this thesis.For the feature layer,pre-trained word vectors are used to obtain prior semantic emotional information;For the model layer,the local semantic emotional features of the text are extracted by CNN,the semantic emotional features of the text context are extracted by BiLSTM,and the two kinds of extracted features are merged dynamically by attention mechanism.Multi scale representation of semantic and emotional features is achieved through the fusion of multiple features.Experimental results show that the proposed method can significantly improve the performance of text emotion recognition.(2)In view of the lack of cross-language corpus which may causes that it is difficult to effectively perform multilingual sentiment analysis,a new type of cross-language sentiment recognition model is developed by integrating lan-guage translation module,cross-language word vector and AMINN model.Firstly,The bilingual parallel corpus is generated by the language translation module,and the word vector containing bilingual information is extracted.Then,transformation matrix of the bilingual word vector is further calculated to get the crosslanguage word vector.Finally,The cross-language word vector is combined with the AMINN model to analyze the emotion of the low resource corpus.The effectiveness of the proposed method is verified by experiments on the cross-language dataset(NLPCC2013).Experimental results show that the proposed method can significantly improve the performance of cross language text sentiment analysis.