Minimal Combination for Incremental Grammar Fragment Learning

The ability to identify text fragments can benefit many information extraction tasks. The main challange in this matter is to recognise the multiple patterns that exist across possibly heterogeneous sources. Traditional machine learning methods can generate grammars for a set of training examples but the disadvantage is their potential heavy computational overhead. The proposed incremental grammar learning method offers immediate learning of new examples by slight alteration of the existing grammar collections by measuring the changes needed to equalize two grammars. This minimal combination approach avoids overgeneralization and maintains the coverage of newly learnt and past examples without a need to repeat the training process. Our experimental results suggest that the proposed algorithm outperforms a fuzzy grammar learning algorithm implemented in normal supervised learning setting.