Learning Natural Language Interfaces over Expresive Meaning Representation Languages

This thesis focuses on learning natural language interfaces using synchronous grammars, λ-calculus and statistical modeling of parse probabilities. A major focus of the thesis has been to replicate Mooney and Wong’s λ-WASP [17] algorithm and implement it inside the C-PHRASE [12] Natural Language Interface (NLI) system. By doing this we can use C-PHRASE’s more expressive and transportable meaning representation language (MRL), rather than the PROLOG-based MRL Mooney and Wong used. Our system, the C-PHRASE LEARNER, relaxes some constraints in λ-WASP to allow use of more flexible MRL grammars. We also reformulate the algorithm in terms of operations on trees to clarify and simplify the approach. We test the C-PHRASE LEARNER over the US geography corpus GEOQUERY and produce precision and recall results slightly below those achieved by λ-WASP. This was expected as we have fewer domain restrictions due to our more expressive and portable MRL grammar. Our work on the C-PHRASE LEARNER system has also revealed some promising avenues of future research including, among others, alternative statistical alignment strategies, integrating linguistic theories into our learning algorithm and ways to improve named entity recognition. C-PHRASE LEARNER is presented as open source to the community to allow anyone to expand upon this work.