Minimally-Supervised Natural Language Learning

Bootstrapping is a minimally supervised machine learning algorithm used in natural language processing (NLP) to reduce the cost of human annotation. It starts from a small set of seed instances (e.g., (cat, animal) for learning is-a relation) to extract context patterns (e.g., “X such as Y”) from a corpus. The extracted patterns are used to extract other target instances which co-occur with the patterns, and the extracted instances are then used for inducing other context patterns. By applying these steps iteratively, one can easily multiply the number of seed instances with minimal human annotation cost. The idea of bootstrapping has been adopted to many NLP tasks such as relation extrac-tion and named entity recognition. However, bootstrapping has a tendency, called semantic drift, to select instances unrelated to the seed instances as the iteration proceeds. The main contribution of this thesis is to demonstrate the semantic drift of bootstrapping has the same root as the topic drift of Kleinberg’s HITS, usin