Emotion Analysis For Short Texts Based On Transfer Learning

Emotion Analysis is the task to automatically classify texts according to their emotions,such as angry,happy,surprise and so on.It is an extensive task of Sentiment Analysis which classifies texts into positive and negative by sentiment tendencies.Sentiment analysis is widely applied in products or movie comments,public opinion analysis and so forth.Sentiment Analysis is still a hot and difficult research field,especially for short texts.Supervised machine learning is the most common and popular approach to emotion analysis,but its performance is seriously dependent on labeled data.Unfortunately high quality labeled data is difficult to obtain and annotating is effort and time consuming.This paper compares the performances of word features and word embedding features for emotion analysis on Chinese short texts,then put the emphasis on transfer learning technologies for emotion analysis to alleviate the problem of short of labeled data.Main contributions and work of this paper are as follows:Analysis features of emotion analysis on short texts Feature engineering is the most important process of supervised machine learning.This paper analyses the characteristics of Sina Weibo texts,designs rules to discover the emotion icons and Internet slang,and proposes POS Tag based word2 vec word embedding generation pattern except character embedding and Segment word embedding.In addition,we compare the performances of word features and word embedding features,conduct experiments to analyse the effects of parameters on word embedding and the performances of CBOW and Skip-Gram model.Emotion analysis transfer learning based on instances Considering the problem of short of labeled data,this paper proposes an Emotion Analysis Transfer Learning model based on Ada Boost(EATAda Boost).The main idea of EATAda Boost is to make most use of source instances,we calculate the semantic similarity between source instances and non-domain specific sentiment words which both occur in source and target domain frequently.The similarity determines how to update instances' weights during the iteration of Ada Boost.Beyond that,we analyse the performances of instances represented by different word embedding compositions.Emotion Analysis transfer learning based on parameters In order to make full use of Chinese sentiment analysis datasets we have,this paper designs two neural network frameworks for emotion analysis transfer learning and realizes the transfer between different datasets and different classifications using two-step transfer learning framework,besides compares the transfer ability of CNN model and LSTM model for emotion analysis.