Deep Representation Learning For Sarcasm Detection In Twitter Using Attention Mechanism

Sentiment analysis is becoming an extremely active research field in Natural Language Processing.Recently,many textual resources containing customers' opinions are available on the internet: Internet user's thoughts,forums,social networks,consumer's surveys.This study is fascinating for companies who want to have an idea about customer feedback on their products as well as for people wishing to inquire about reviews on products or trips.At present,Twitter is one of the most popular micro-blog which connects people around the world,and it also has a high level of user participation,it is a part of their daily lives.In this thesis,the author presents a study of sarcasm detection in Twitter messages.Although,most of the researches on sarcasm detection highlight the use of syntactic,pragmatic,or lexical features that are often expressed over figurative devices such as words,exclamation,and emoticons.In this thesis,the author exploits the deep representation learning in sarcastic tweets detection tasks by integrating the Attention Mechanism with the deep neural networks models and the models results with the state-of-the-art feature engineering approaches.The author also built a dataset of tweets manually annotated concerning the presence of sarcasm.Since Recurrent Neural Network(RNN)model usually cannot cover all the important information from its final hidden state,the author focused on the impact of the Attention Mechanism,first integrating it with Long Short-Term Memory(LSTM)and then with Bidirectional LSTM to find out the relative contribution of every word in the sentence.The results show that the proposed models achieve competitive compared to the state-of-the-art results by getting an F1 score of 0.91 with Attention-based LSTM model and 0.90 with the Attention-based BLSTM model.