A Hybrid Tree Framework for Semantic Parsing and Natural Language Generation

In this thesis, we present a novel framework based on hybrid trees that aims to bridge natural language sentences and their underlying meaning representations. The framework is guided by theoretical principles related to language and semantics. The purpose of the framework is to facilitate the development of systems that transform natural language sentences into their underlying meaning representations (called semantic parsing), as well as systems that transform meaning representations into their corresponding natural language sentences (called natural language generation). Within the framework, we build a novel generative model that jointly generates both a natural language sentence and its meaning representation. We also develop efficient training and decoding algorithms for the proposed generative model. The generative model has a natural symmetry, allowing for easy transformation from natural language to meaning representation, and vice versa. In practice, though the generative model gives reasonable performance, it still exhibits some limitations. To address these limitations, additional discriminative techniques are added to the generative model when performing both semantic parsing and natural language generation. We demonstrate through experiments that such a pipelined approach gives significant improvements in performance. In particular, the generative model, when augmented with discriminative techniques, outperforms previous state-of-the-art systems when evaluated on standard benchmark corpora. Acknowledgements I am fortunate to have the privilege to work closely and have interaction with 3 excellent professors in the special and unique SMA programme: Prof. is an expert in the field of natural language processing. He helped me a lot on finding research topics and directions, contacting various other researchers for enquiries, pointing to relevant papers and researchers in the community. Prof. Ng is very good at finding good research directions. He has a very good view about the big picture in the field, which is very important. He is also extremely careful in academic writings and presentations. From him I have learned a lot. Prof. Lee is very good at machine learning and statistical methods. We had lots of very useful discussions on how to build enhance statistical models effectively, how to improve algorithms and so on. He is very approachable and I can always hear many good technical suggestions from him during discussions and emails. Although Prof. Kaelbling's main research focus was on machine learning and robotics, she is still a very responsible supervisor from whom I have learned a lot. She is very good at posing high level questions for my research. These …