Ανάλυση συναισθήματος στο twitter με βαθιά νευρωνικά δίκτυα
KeywordsΑνάλυση συναισθήματος ; Εξόρυξη γνώμης ; Επεξεργασία φυσικής γλώσσας ; Κατηγοριοποίηση κειμένων ; Μηχανική μάθηση ; Τεχνητά νευρωνικά δίκτυα ; Βαθιά νευρωνικά δίκτυα ; Sentiment analysis ; Opinion mining ; Natural language processing ; Text classification ; Machine learning ; Artificial neural networks ; Deep neural networks
Sentiment analysis is an area in Natural Language Processing (NLP), studying the identification and quantification of the sentiment expressed in text. The thesis addresses the problem of predicting the sentiment of messages from Twitter microblogging service. The task is approached using Artificial Neural Networks (ANN), utilizing distributed text representations (word embeddings). Firstly, a survey of the field of sentiment analysis is performed. Next, the most important approaches for addressing the problem are reviewed. Special focus is given to the resurgence of research in ANNs. Τhe most common ANN architectures, which have been applied to sentiment analysis, are presented. Furthermore, a comparison is being made between ANNs with the more traditional machine learning approaches, providing theoretical justifications for choosing ANNs for modeling natural language. Moreover, a text pre-processing tool was developed, for preparing the Twitter messages before passing them as inputs to the machine learning models. The tool is geared towards texts from social networks, which are very challenging to deal with, because of their informal and “creative” writing style, with improper use of grammar, figurative language, misspellings and slang. The text processing tool, is able to utilize most of the information in text, performing sentiment-aware tokenization, spell correction, word normalization, word segmentation (for splitting hashtags) and word annotation. Finally, in the context of my research, we participated in Semeval-2017, which is an international competition for semantic evaluation of computational semantic analysis systems. The models that were developed for the participation in Semeval, were essentially the result of my research. A thorough analysis of these models is given, along with the rationale behind each design decision. The models were very competitive, achieving the first place in Task 4:“Sentiment Analysis in Twitter” and the second place in Task 6: “#HashtagWars: Learning a Sense of Humor”.