Text Representation through Multimodal Variational Autoencoder for One-Class Learning

• Autor(a) principal: Gôlo, Marcos Paulo Silva
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-23052022-150550/
• topic: Aprendizado de máquina baseado em uma única classe; Classificação de textos; Multimodal variational autoencoder; One class learning; Text classification; Variational autoencoders multimodais

Automatic text classification has become increasingly relevant for several applications, both for academic and business purposes. Traditionally, multi-class learning methods perform text classification, which requires prior labeling of textual datasets for all classes. These methods fail when there is no well-defined information about the texts classes and require a great effort to label the training set. One-Class Learning (OCL) can mitigate these limitations since the model training is performed only with labeled examples of an interest class, reducing the users labeling effort and turning the classification more appropriate for open-domain applications. However, OCL is more challenging due to the lack of counterexamples for model training. Thus, OCL requires more robust text representations. On the other hand, most studies use unimodal representations, even though different domains contain other types of information that can be interpreted as distinct modalities for textual data. In this sense, the Multimodal Variational Autoencoder (MVAE) was proposed. MVAE is a multimodal method that learns a new representation from the fusion of different modalities, capturing the characteristics of the interest class in a more adequate way. MVAE explores semantic and syntactic representations, density, linguistic and spatial information as modalities. Furthermore, MVAE is based on a Variational Autoencoder, considered one of the state-of-the-art for learning representations. Finally, the main contributions of this dissertation are: (i) a multimodal method to represent texts in the OCL scenario; (ii) detection of fake news through representations generated by MVAE; (iii) applying MVAE to represent app reviews in the filtering of relevant app reviews; and (iv) sensing events represented by the MVAE.