Meaning from structure in natural language processing

The goal of natural language processing is to construct a computer-digestible representation of the meaning of the typed sentence, i.e. a semantic representation. The development of larger scale natural language systems has been hampered by the need to manually create mappings from syntactic structures into meaning representations. A new approach to semantic interpretation is described, which uses partial syntactic structures as the main unit of analysis for interpretation rules. The approach can work for a variety of syntactic representations corresponding to directed acyclic graphs and is designed to map into meaning representations based on frame hierarchies with inheritance. Semantic interpretation rules are defined in a compact format which is suitable for automatic rule extension or generalization, when existing hand-coded rules do not cover the current input. Furthermore, automatic discovery of semantic interpretation rules from input/output examples is made possible by this new rule format. The principles of the approach are validated in a comparison to other methods on an independently developed domain. In experiments performed on an English language corpus of sentences, the approach allowed semantic interpretation rules to be created manually in about 50 percent less time, with 78 percent coverage of the test corpus, as opposed to the 66.1 percent coverage which had been achieved before with the original rules written for this application by independent sources. In addition, automatic rule discovery on the English test corpus produced semantic interpretation rules that accurately mapped 70.5 percent of the corpus. Similar experiments performed on a Japanese corpus of sentences yielded comparable results, with a slight disadvantage for both manual rule creation as well as automatic rule discovery using the new approach, due to external factors such as incomplete lexical coverage.;Instead of relying purely on painstaking human effort, this thesis shows that a combination of human expertise with learning strategies by the computer on representative examples is successful to overcome the bottleneck of semantic interpretation.