Identifying Labels From Multi-label Texts Using Deep Learning

Mental health has become a focus of the society.With the development of Internet technology,Internet has become an important channel for people to consult mental health problems,so mental health websites came into being.With the accumulation of counseling problems,mental health websites have accumulated a large number of psychological texts,which contains emotion labels that express different mental illnesses.If these emotion labels can be automatically identified,it will help to quickly suggest similar psychological texts for the patient who is seeking help on the site.Each psychological text may belong to multiple emotion labels at the same time,so the psychological text of this thesis belongs to the multi-label text,and the task of automatically identifing these emotion labels belongs to the multi-label text classification problem.This thesis uses CNN(Convolutional Neural Network)and LSTM(Long-Short Term long short term memory Memory)to build combined deep neural network framework BLSTM_CNN(Bi-directional LSTM-CNN)classifier to automatically identify emotion labels from psychological text.This thesis firstly identifies the suitable word embedding method and the dimension of the word vector.Starting from the representation of the text,this thesis uses word2vec and GloVe as word embedding methods to train different dimensions of the word vector respectively with Chinese Wikipedia text data and psychological text data.We use the word vector to represent the psychological text,and compare the different word embedding methods and different dimensions of the word vector on the experimental results.and then select the most suitable word embedding method and word vector dimension for the task.And then this thesis builds classifiers to identify emotion labels from psychological text.From the view of network structure,we use the pre-trained word vector to build CNN classifier,LSTM classifier,LSTM_CNN classifier and BLSTM_CNN classifier for multi-label text classification.And according to the characteristics of psychological text,analyze the advantages and disadvantages of the CNN model and the LSTM model explored in this task.Use the combined deep neural network BLSTM-CNN to avoid the disadvantages and combine the advantages of the CNN model and the LSTM model.The experimental results show that the BLSTM_CNN classifier is better than the classifier built by CNN or LSTM model,or the LSTM_CNN classifier combined with CNN and LSTM model on the macro_fl evaluation index of multi-label classification.The effectiveness of BLSTM_CNN classifier in multi-label text classification is verified.