![]() | Only 14 pages are availabe for public view |
Abstract Most of the tasks that infer features from text are addressed in a way that either ignores the environment’s impact, or narrows the context down, in the best scenario, to capturing long-term semantic dependencies of a sentence, or a review. This work explores the impact of taking the environment within which a tweet is made, on the task of analyzing sentiment orientations of tweets produced by people in the same community. The thesis proposes- as we call it- C-GRU (Context-aware Gated Recurrent Units), which extracts the contextual information (topics) from tweets and incorporates them into finding sentiments conveyed by the tweet. The proposed architecture learns direct co-relations between such information and the (sentiment) predictions. With a multi-modal model, the architecture combines both outputs learnt (from topics and sentences) by learning the contribution weights of the two sub-modals to the predictions. The evaluation of the proposed model which is carried out by applying it to the SemEval-2018 dataset for Arabic multi-label emotion classification, demonstrates that the model outperforms the highest reported performer on the same dataset, with an accuracy of 54.4%. Also, it shows comparable results on the Stanford Sentiment Tree English dataset, and a further Arabic tweets dataset for emotion detection, with accuracies of 46% and 73.4% respectively |