Inference during reading: multi-label classification for text with continuous semantic units

摘要

With the developing of electronic platform such as education, commerce and etc., text with continuous semantic units (CSU-text) emerges in large numbers. Each CSU-text is usually short but contains series of independent semantic units. Mining CSU-text is helpful to determine users’ preferences or intentions, and further improve service quality in real life. Even though there are lots of text mining techniques, they are hard to well handle CSU-text because they usually learn one representation for a whole document. In this case, the information hidden in various semantic units can not be sufficiently captured. Inspired by how a human being understands a text and acquires knowledge in cognitive science, in this paper, we treat multi-label classification for CSU-text as a sequence tagging task and propose a novel inference during reading (InfDR) model. The model is able to simultaneously partition continuous semantic units and map them to semantic labels. Extensive experiments are conducted on three real-world datasets, demonstrating that the proposed model is effective and significantly outperforms the existing baselines with one single text representation.