Research And Application On Transfer Learning For Review Text

To transfer the knowledge from one scenario to another has always been a hot topic in the field of artificial intelligence research.Transfer learning aims to solve the problems with few or even not any labeled data in target domains.The main studies of the thesis for reviewers and review content are as follows:1.For the evaluation of the reviewer's reputation,in this thesis,a method based on emotional distance and domain self-adaption is proposed.To address cross-domain review sentiment tendency analysis from reviewers,a transfer learning framework based on two-layer convolutional neural network is proposed.This kind of automatic mechanism is very suitable for the following work which is for randomly and a variety of fields and need process in real-time.In addition,by calculating the emotional distance and consistency of emotional tendency between the reviews written by the target reviewer and those by other reviewers,the review objectivity and the emotional consistency between the target reviewer and the public are measured.Finally,by using product reviews on Amazon.com as the experimental data,results of the proposed method with the sequencing results of the Amazon Reviewer ranking are displayed to show the rationality of the method.Research on how to evaluate a reviewer's reputation effectively from his or her written reviews would be of great significance to research on text effectiveness and regularizing the behavior of consumers on e-commerce platforms.It also provides a new quantitative way for reviewer reputation research.2.For the problem of feature mismatch in cross-domain product review sentiment classification,in this thesis,a cross-domain semantic correlation auto-correspondence transfer learning method is proposed.Our method uses adjectives as features.Similar-pivot feature pairs that express similar sentiments but in different representations in either domain with the help of word embedding and pivot features are constructed.Finally,pivot features can then be applied to align similar sentiment feature representations.This process can avoid feature mismatches and reduce sentiment discrepancies between domains from the perspective of word semantic.The experimental results from testing on Amazon product reviews demonstrate that our method is suitable for the transfer learning task of product review sentiment classification and significantly outperforms previous approaches in sentiment classification.3.For the transfer learning tasks about identifying whether a post expresses confusion,determining the response urgency,and classifying the polarity of the sentiment,a transfer learning framework based on a convolutional neural network and a long short-term memory model is proposed.In this thesis,convolution operation can learn the feature representation by considering the local contextual feature.The feature representations can be input into LSTM model,which considers the long-term temporal semantic relationships of features.Finally,the possibility of transferring parameters from a model trained on one to another course and the subsequent fine-tuning are investigated.Experiments on Stanford MOOC forum post demonstrate that the transfer learning framework proposed in this thesis can effectively resolve the lack of annotated posts and biases across courses.It also can increase the effectiveness of monitoring MOOC forums in real time.