Rule meta-learning for trigram-based sequence processing

Predicting overlapping trigrams of class labels is a recently-proposed method to improve performance on sequence labelling tasks. In this method, sequence elements are effectively classified three times, therefore some procedure is needed to post-process those overlapping classifications into one output sequence. In this paper, we present a rule-based procedure learned automatically from training data. In combination with a memory-based leaner predicting class trigrams, the performance of this meta-learned overlapping trigram post-processor matches that of a handcrafted post-processing rule used in the original study on class trigrams. Moreover, on two domain-specific entity chunking tasks, the class trigram method with automatically learned post-processing rules compares favourably with recent probabilistic sequence labelling techniques, such as maximum-entropy markov models and conditional random fields.