Granular transfer learning using type-2 fuzzy HMM for text sequence recognition

Context information plays an important role in text sequence recognition, but it is difficult to harness the uncertainty caused by conflicting implications. In this paper, we propose a novel Granular Transfer (GT) learning with type-2 fuzzy Hidden Markov Model (HMM) called GT2HMM, in which interpretable granules' representation is introduced to describe the contextual uncertainty for its transfer learning. The correspondences among words are transformed into information granules using fuzzy c-means. To fulfill the utilization of granular information in sequence recognition, we construct a type-2 fuzzy HMM which fuses labeled data and unlabeled observations. With a tunable granularity, correspondence information is refined in a coarse-to-fine manner in GT2HMM. Experiments on transductive and inductive transfer learning in part-of-speech (POS) tagging tasks verify the effectiveness of our proposed GT2HMM.