Classification Model And Text Enhancement For Suggestion Recognition

With the development of e-commerce,the value of customer message is gradually highlighted.Getting high-quality suggestion information from structured and semi-structured text data presents an immeasurable business prospect.Recommendation recognition is an application scenario of text classification in natural language processing.The purpose of suggestion recognition is to use natural language processing technology to mine the natural text composed of customer message,so as to accurately identify which text has the attribute of suggestion,so that service providers can focus on these suggestions and improve the service quality.To this end,based on deep learning technology,this thesis investigate how to establish an effective recognition model and how to use the features contained in the text to enhance the accuracy of the suggestion recognition model.The main work of this thesis is as follows:A classification model based on deep neural networks by treating suggestion recognition as a classification problem is proposed.The designing principle,loss functions and optimization methods are given in details.Two modules of text representation and feature learning constituting the proposed model is studied separately,different word vectors or language models,such as Word2 Vec,Glo Ve,Fasttext,ELMO,BERT and neural network components such as GRU,LSTM,Bi LSTM,Bi LSTM-Att,CNN-Att,CNN are combined to evaluate the accuracy of classification.Through intensive experiments,it was confirmed that the hybrid model based on BERT and CNN outperforms the remaining combinations.Two kinds of text enhancement methods,namely,word expansion and representation expansion are proposed to enhance classification.Word expansion recognizes important vocabulary through TF-IDF and then refers to the Microsoft Concept Graph database to supplement the important vocabulary-related features,complete the expansion of the original text.Representation expansion uses graph convolution technology to extract text features,merge with the text representation obtained by BERT-CNN,and make full use of the cooccurrence relationship between words,co-occurrence relationship between documents to realize the enhancement of text representation.Experiments show that these two methods can improve the accuracy of suggestion recognition.