Text-based Sentiment Analysis and Music Emotion Recognition

Nowadays, with the expansion of social media, large amounts of user-generated texts like tweets, blog posts or product reviews are shared online. Sentiment polarity analysis of such texts has become highly attractive and is utilized in recommender systems, market predictions, business intelligence and more. We also witness deep learning techniques becoming top performers on those types of tasks. There are however several problems that need to be solved for efficient use of deep neural networks on text mining and text polarity analysis. First of all, deep neural networks are data hungry. They need to be fed with datasets that are big in size, cleaned and preprocessed as well as properly labeled. Second, the modern natural language processing concept of word embeddings as a dense and distributed text feature representation solves sparsity and dimensionality problems of the traditional bag-of-words model. Still, there are various uncertainties regarding the use of word vectors: should they be generated from the same dataset that is used to train the model or it is better to source them from big and popular collections that work as generic text feature representations? Third, it is not easy for practitioners to find a simple and highly effective deep learning setup for various document lengths and types. Recurrent neural networks are weak with longer texts and optimal convolution-pooling combinations are not easily conceived. It is thus convenient to have generic neural network architectures that are effective and can adapt to various texts, encapsulating much of design complexity. This thesis addresses the above problems to provide methodological and practical insights for utilizing neural networks on sentiment analysis of texts and achieving state of the art results. Regarding the first problem, the effectiveness of various crowdsourcing alternatives is explored and two medium-sized and emotion-labeled song datasets are created utilizing social tags. One of the research interests of Telecom Italia was the exploration of relations between music emotional stimulation and driving style. Consequently, a context-aware music recommender system that aims to enhance driving comfort and safety was also designed. To address the second problem, a series of experiments with large text collections of various contents and domains were conducted. Word embeddings of different parameters were exercised and results revealed that their quality is influenced (mostly but not only) by the size of texts they were created from. When working with small text datasets, it is thus important to source word features from popular and generic word embedding collections. Regarding the third problem, a series of experiments involving convolutional and max-pooling neural layers were conducted. Various patterns relating text properties and network parameters with optimal classification accuracy were observed. Combining convolutions of words, bigrams, and trigrams with regional max-pooling layers in a couple of stacks produced the best results. The derived architecture achieves competitive performance on sentiment polarity analysis of movie, business and product reviews. Given that labeled data are becoming the bottleneck of the current deep learning systems, a future research direction could be the exploration of various data programming possibilities for constructing even bigger labeled datasets. Investigation of feature-level or decision-level ensemble techniques in the context of deep neural networks could also be fruitful. Different feature types do usually represent complementary characteristics of data. Combining word embedding and traditional text features or utilizing recurrent networks on document splits and then aggregating the predictions could further increase prediction accuracy of such models.