Stance Detection Via Condition-CNN Model

With the increasing development of Internet technology and the widespread use of new media,the social media platforms is becoming more diverse.As the most commonly used social media platform,Weibo can let users as a presenter to post topic which may related to a social phenomenon.Other users can use the # to express opinions to following the topic.When researchers need to analyze the public opinion of a given target topic,they can use this function to collect the relevant texts to construct a datasets conviniently.Therefore,Automatically detecting stance has widespread applications in information retrieval,text summarization,and textual entailment.Over the last decade,the focus of modeling stance is often on the analysis of Weibo text,while ignoring the influence of target.In some improved researches,most work focuses on combining target with Weibo text.The combined method is mainly to simply concat two sequences or using Attention calculation so that the target features can be added to the features of the Weibo text with different weights.This paper proposes a model framework based on Condition-CNN,which not only combines the features of topic and text,but also expands the topic by extracting the topic phrase,so that the role of topic information in the stance detection task is fully considered.This paper considers following two aspects:First,this paper uses LDA to extract the topic words from the Weibo text and calculates the pointwise mutual information of adjacent words to determine whether this adjacent words can be use as a phrase.When extracting the topic phrase from the Weibo text,firstly calculate the topic words scale and point mutual information for the phrase which appearing in the corpus,and use the phrases with high scores as the candidates of the topic phrases;then use the Trie structure to storage them to filter the incomplete phrases.Second,the extracted topic phrases are used as a supplement to the targets to form target set.Before the classification,we calculate a relation matrix(we call it Condition layer in this paper)which can make the target set and the Weibo text to better reflect;The product neural network performs feature extraction on the Condition layer and predicts the stance.This paper calls this model based on the relation matrix as Condition-CNN.We evaluate our approach on the NLPCC 2016 Task 4: Stance Detection in Chinese Microblogs corpus achieving the best performance.