Συνάντηση εργαστηρίου SDE: Opinion Mining in the Deep

Abstract: Opinion Mining (OM) or Sentiment Analysis (SA) is one of the promising fields of text mining that its major goal is to assign opinions, expressed on social media using natural language, a sentiment valence. Well established methods addressing OM use natural language processing tools and sentiment lexicons for creating learning features so as to train the sentiment classifier. The availability of such resources is not always the case for most of  the human natural languages. Deep neural networks (NNs) promise end-to-end solutions where the learning features extraction is part of the learning process. However NNs require a large amount of labeled data for the training process. Opinionated text is available at scale on social media, unlabeled. This study adopts deep recurrent neural networks for addressing multi-lingual OM by proposing a novel method of transfer learning. A classifier is trained on an available large dataset in one natural language. Learning weights are transferred so as to train an OM classifier in another natural language with a constrained in size labeled dataset. The method is validated by state-of-the-art classification results.