Efficient distributional approaches for unsupervised grammar learning David Brooks, Dr. Mark Lee
This event took place on 8th March 2006 at 9:00am (09:00 GMT)
Knowledge Media Institute, Berrill Building, The Open University, Milton Keynes, United Kingdom, MK7 6AA
Distributional approaches to grammar learning are typically inefficient, enumerating large numbers of candidate constituents. In this paper, we introduce an efficient incremental procedure for learning syntactic structure by distributional analysis, Directed Alignment, which uses heuristics to reduce the space of candidate constituents. We apply Directed Alignment to a large corpus containing over 400000 words, and evaluate the results using EVALB. We show that the performance of this approach is limited, and provide a detailed analysis of learnt structure and the learning process. We complement this with an analysis of actual constituentcontexts, showing that some of the problems result from diminishing probability over constituent-context distributions. Our findings suggest that distributional methods do not generalise enough to learn effectively from raw text, and that we should investigate ways to increase this generalisation.
Distributional approaches to grammar learning are typically inefficient, enumerating large numbers of candidate constituents. In this paper, we introduce an efficient incremental procedure for learning syntactic structure by distributional analysis, Directed Alignment, which uses heuristics to reduce the space of candidate constituents. We apply Directed Alignment to a large corpus containing over 400000 words, and evaluate the results using EVALB. We show that the performance of this approach is limited, and provide a detailed analysis of learnt structure and the learning process. We complement this with an analysis of actual constituentcontexts, showing that some of the problems result from diminishing probability over constituent-context distributions. Our findings suggest that distributional methods do not generalise enough to learn effectively from raw text, and that we should investigate ways to increase this generalisation.
Click below to play the event (27 minutes)