To facilitate the application of semantics in statistical machine translation, we propose a broad-coverage predicate-argument structure mapping technique using automated resources. Our approach utilizes automatic syntactic and semantic parsers to generate Chinese-English predicate-argument structures. The system produced a many-to-many argument mapping for all PropBank argument types by computing argument similarity based on automatic word alignment, achieving 80.5% F-score on numbered argument mapping and 64.6% F-score on all arguments. By measuring predicate-argument structure similarity based on the argument mapping, and formulating the predicate-argument structure mapping problem as a linear-assignment problem, the system achieved 84.9% F-score using automatic SRL, only 3.7% F-score lower than using gold standard SRL. The mapping output covered 49.6% of the annotated Chinese predicates (which contains predicate-adjectives that often have no parallel annotations in English) and 80.7% of annotated English predicates, suggesting its potential as a valuable resource for improving word alignment and reranking MT output.
展开▼
