Semantic And Sentiment Analysis In Twitter Event Detection

With the rapid development of social networks,people share what they see,hear and think on the social platforms anytime,anywhere.Many researchers believe that social networks are a sensor network that reflects the real world.Consequently,social media data analysis has a wide range of applications,such as detecting criminal activities,predicting public behaviors,et al.As the proportion of text data in social media data is very high and text data is rich of information,text semantic analysis is critical to social media data analysis.Text semantic analysis research in the past is mainly designed for text in formal language,such as news texts,Wikipedia texts,and so on.However,social media text is limited in length and contains a lot of misspellings,slangs,grammatical errors,and other non-canonical language usage.Therefore,it is not ideal to directly apply traditional semantic analysis techniques to social media texts.Based on the existing semantic analysis techniques,this thesis proposed a method to learn tweet features which contain both semantic and sentiment information,and applied these tweet features to Twitter event detection.The main work of this thesis can be summarized as the following two parts:1.Generate word representation which contains both semantic and sentimental information.The vectorization and semantic features of words are the basis of text semantic analysis.This thesis analyzed the state-of-the-art neural network language model,word2 vec.As the word2 vec word embedding performs poorly on distinguishing synonyms and antonyms,this thesis proposed a method to simultaneously use the word context's semantic and sentimental information to construct word embedding to enhance word embedding's ability to distinguish antonyms from synonyms.Specifically,this thesis used a distant supervision method,extended the word2 vec neural network model by using the emoticons in tweets as weak sentiment labels,and encoded the semantic and sentiment information of the context into word embeddings.This thesis calls this word embedding containing semantic and sentiment information as senti-word2 vec word embedding.2.Combine semantic and sentiment information to detect event in Twitter.The traditional Twitter event detection methods organize semantically similar tweets to represent events.However,many of automatic semantic feature extraction methods have limited ability to distinguish synonyms and antonyms.Therefore,the tweets in the same event cluster may express different sentiment attitudes towards the same event.Under the constraint of sentiment information,this thesis proposed to split the Twitter event cluster into event favor cluster,event against cluster and event neutral cluster.Specifically,this thesis used the senti-word2 vec word embedding to generate tweet features containing semantic and sentiment information,analyzed the impact of this kind of tweet features on tweet semantic similarity judgment and sentiment analysis,and finally applied the tweet features to Twitter event detection.In order to obtain a large number of accurate and objective test data,this thesis used four public test data sets to evaluate experiment results.Compared with the original word2 vec word embedding,on the Sim Lex-999 semantic polarity test dataset,senti-word2 vec word embedding's capability of distinguishing antonyms and synonyms increased by 112.97%;on the three tweet test datasets from SemEval,the semantic similarity detection results produced by senti-word2 vec word embedding increased by an average of 59.37%,the sentiment analysis results of senti-word2 vec word embedding increased by an average of 1.67%,the Twitter event detection results increased by an average of 3.99%.