An Attention-based Model For Stance Detection In Chinese Microblog

With the rapid development of social networks,users have generated massive amounts of data every day on the Internet.These data is always valuable,social network data in the stance detection has gradually become an important research direction.The fifth session of the Natural Language Processing and Chinese Computing Conference(NLPCC2016)in 2016 proposed position detection task based on the Chinese microblogging.In the existing work,the researchers mainly through the construction of feature engineering,add emotional dictionary and expert knowledge in areas such as resources to obtain valuable semantic features.But this way it takes a lot of manpower about designing features and training machine learning models and final detection effect is heavily dependent on the quality of the feature design and the tuning of the model parameters.So some other researchers use the deep learning model to learn the features and achieved good results.Chinese microblogging length is limited and contains more Internet terms,emoticons and other non-normative text,resulting in microblogging text to accommodate the information is very limited.A few words can roughly reflect the sentence's overall position information in the microblogging.How to extract the richer semantic features from the limited microblogging information is the focus of the research on the microblogging stance detection task.We propose an attention-based BiLSTM-CNN model for Chinese microblog stance detection.First of all,we use a variety of word segmentation model for microblogging text word segmentation in order to reduce the noise generated by the interference.Moreover,emotional tendencies affect the position of the sentence,so we add emotional features to the sentence representation.Based on the convolution neural network(CNN),the Chinese microblogging position detection model is constructed.Due to the loss of information in the traditional pooling strategy,we propose a bi-directional short-term memory neural network(LSTM)and a convolution neural network(CNN)hybrid network detection model based on attention mechanism.Attention mechanism makes microblogging key words and features can be highlighted at the same time to improve the CNN traditional pooling strategy.This experiment uses the task notes and evaluation criteria provided by NLPCC2016.Experiments show that the proposed BiLSTM-CNN-ATT network model can extract the position information contained in Chinese microblogging and obtain good results.