Text structure-aware classification

Abstract

Bag-of-words representations are used in many NLP applications, such as text classification and sentiment analysis. These representations ignore relations across different sentences in a text and disregard the underlying structure of documents. In this work, we present a method for text classification that takes into account document structure and only considers segments that contain information relevant for a classification task. In contrast to the previous work, which assumes that relevance annotation is given, we perform the relevance prediction in an unsupervised fashion. We develop a Conditional Bayesian Network model that incorporates relevance as a hidden variable of a target classifier. Relevance and label predictions are performed jointly, optimizing the relevance component for the best result of the target classifier. Our work demonstrates that incorporating structural information in document analysis yields significant performance gains over bag-of-words approaches on some NLP tasks.