Research Project Name
Learning Strategies for Story Comprehension: A Reinforcement
Learning Approach
Objective
The objective is to construct a computational model that
can answer Who/What/Why/When/How questions about a natural language story. The
scientific goal is to achieve this by having all knowledge in the computational
model be learned, using statistical and machine learning techniques, with no use of manually entered knowledge.
Results
A computational
model, called QABLEe, has been implemented.
The initial lexical knowledge is constructed using off the
shelf statistical language preprocessors over a subset of the target story
corpus.
Then the performance is improved using inductive generalization and
reinforcement learning over classified examples to select the most useful
lexical pieces of information to be used by the inference procedure when
answering questions about the story. Finally, the model is tested on
a subset of stories that have not been previously seen.
QABLe is compared to three prior non-learning systems on the Remedia corpus of
fourth grade stories. The conditions
under which learning is effective are evaluated. The experimental results show
that a learning-based approach significantly improves upon “matching and
extraction only” techniques.
Publications
- Grois, E. and Wilkins, D. C., "Learning Strategies for Open-Domain
Natural Language Question Answering," Nineteenth International Joint
Conference on Artificial Intelligence, Edinburgh, Scotland, July 30 –
August 5, 2005.
pdf
- Grois, E. and Wilkins, D. C., "Learning Strategies for Story
Comprehension: A Reinforcement Learning Approach," Twenty-Second
International Conference on Machine Learning (ICML),
Bonn,
Germany,
August 7-11, 2005.
pdf