As Twitter gains popularity, millions of messages are generated every day via this platform allowing people to communicate with each other through daily chatter and public conversations. This social content is usually considered more subjective than professional articles involving a lot of opinions and thoughts. This has stimulated many companies to use tweets to keep track and have general overview on their customer opinions about the brand. The conversational element of Twitter is of particular interest to the marketing community. However, most studies on summarizing customer opinions through Twitter, have so far employed single tweets rather than considering the whole conversations. As tweets are limited to 140 characters and usually written in an informal way, it is frequently hard to detect the exact meaning of the tweet when taken separately. In this paper, we propose a new approach for customer opinion summarization based on Twitter conversations named COSTwiC. We employ a new conversation-based method that employs conversation interactions to effectively extract the different product features as well as the polarity of the conversation messages. Experimentations show promising results. In particular, we have proved that incorporating conversation structure in the opinion summarization task contributes to improving system performance.
Deep learning models such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models have recently emerged as effective solutions to various NLP tasks with comparatively remarkable results. The CNN model efficiently extracts higher level features using convolutional layers and max-pooling layers while the LSTM model allows capturing long-term dependencies between word sequences. In this paper, we propose a hybrid CNN-LSTM model taking advantage of both models to overcome the sentiment analysis problem on Twitter data. We employ a multi-channel CNN that extracts local n-gram features in different sizes using several filters of different lengths. We also employ a weighted average word embeddings method which incorporates sentiment information in the continuous representation of words based on an adapted version of the delta TFIDF measure. These word representations will be the input for the CNN-LSTM model. Experiment results substantiate our intuition by reaching good macro average recall and accuracy scores beating several existing models as well as our same model using individual CNN and LSTM.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.