Research On Text Emotional Analysis Method Based On ON-LSTM

With the rapid development of social network and the extensive participation of users,the network produces tens of millions of daily text data,which contains a large number of public views and attitudes towards hot events,as well as users' experience of products.It is of great social and commercial value to mine,analyse,identify and understand the emotional information contained in these text data.And text emotional analysis research attracts the attention of many researchers in academia and industry.Due to the oral and non-standard characteristics of the web texts,the research of emotional analysis faces many challenges.In terms of word embedding,most of the existing models used Word2 Vec,GloVe and other pre-training language models to obtain static word vectors,which could not be dynamically adjusted according to the specific context to solve the problem of polysemy.In terms of text feature extraction,CNN,Bi-LSTM and other neural network models can extract local semantic features and global semantic features,but they cannot learn the syntactic features of sentences.In addition,emotional category labels play an important role in emotion classification,but most classification models only symbolize category labels,and the semantic information of category labels is not fully utilized.Therefore,this thesis conducts an in-depth study on the text emotional analysis methods based on deep learning.The main work is divided into the following two aspects:(1)A text feature extraction method based on ON-LSTM and attention mechanism is proposed.Firstly,the dynamic word vector representation of emotional text is carried out by the pre-training language model BERT.Secondly,the emotional text is encoded by Ordered Neurous Long Short-Term Memory.The semantic information of the text is learned,and the syntactic structure information of the text is learned unsupervised,so as to get a more comprehensive and deep text feature representation.The attention mechanism is used to pay more attention to the emotion-related information in the text,which acheives the goal of extracting emotional text features.Experiments of comparison on the public evaluation data set show that the text feature extraction method could improve the effect of emotion analysis to a certain extent.(2)An emotion analysis model based on ON-LSTM and label semantic(ON-LSTM-LS)is proposed.Based on the characteristics that emotional labels have semantic information and can guide emotional analysis,this thesis combines the emotional features of text with the semantic features of labels to participate in emotional analysis.Three layer ON-LSTM network is used to extract emotional text features.There are two ways to obtain the semantic features of emotional category labels.The first is to obtain the word embedding of category labels by the pre-trained BERT.And the second approach is to extend category lables semantics which based on Bi-LSTM and attention mechanism,so as to solve the problem of insufficient semantics of category labels.In other words,introducing external knowledge to enrich labels,and then extracting the related features of emotional labels from external knowledge.Further,the extracted features as a supplement to the label semantics achieve the purpose of enriching the labels semantics.Ablation experiment and the comparison experiment of similar models on CLUE Emotion Analysis Dataset show that ON-LSTM-LS,an emotion analysis model incorporating semantic information of lables,could further effectively improve the effect of emotion analysis.In this thesis,ON-LSTM-LS,an emotion analysis model,is constructed based on Ordered Neurous Long Short-Term Memory and attention mechanism,and the semantic information of emotional category lables is integrated into the model to improve the performance of text emotion analysis.