Semantic Operations for Transfer-based Machine Translation

Thesis (Ph.D.)--University of Washington, 2018This dissertation describes a new approach to the automatic extraction of semantic mappings (transfer rules) for rule-based machine translation. This approach continues previous work in combining HPSG rule-based grammars, whose precise bidirectional implementation facilitates deep semantic analysis of sentences and the enumeration of grammatical realizations of semantic representations, and data-driven techniques of machine translation, whose automatic extraction of knowledge and statistical inference allow models to be quickly built from bitexts and to rank extracted patterns by their frequency. I define two new methods for bilingually aligning semantic fragments (or semantic subgraphs) and a heuristic strategy for aligning nodes between source and target subgraphs, which together allow me to design transfer systems that meet, and at times exceed, the translation coverage and quality of the prior state of the art with a significantly reduced dependence on idiosyncratic language-pair definitions (i.e., improved language independence). These improvements are made possible by a number of semantic operations, either designed or implemented by me and defined within this dissertation, that fully model the semantic representations and allow for inspection and transformation as graph operations. I apply my methods to the task of translating Japanese sentences into English—a typologically distant language pair