| Popis: |
Translation divergences are a challenge for MT and alignment. In this paper, we investigate whether an alignment method based on semantic knowledge improves over approaches for linguistically uninformed word alignment and purely syntax-based tree alignment. We annotate sentences with rolesets from PropBank and NomBank (verbal and nominal predicates and their semantic roles), and link predicates to their auxiliary words (auxiliary, modal and support verbs) using parse trees. We study two language pairs, English-French and English-Dutch. As no extensive semantic resource is available for French and Dutch, the annotation strategy we choose is crosslingual semantic annotation projection, combined with automatic SRL. A manual evaluation of our system on an English-Dutch sample shows our system is successful at adding links for predicates to the output of a word alignment system (GIZA++) and two tree alignment systems (Lingua-Align and Sub-Tree Aligner). The performance for role linking is significantly lower, due to errors in the English or target parses. ispartof: pages:73-82 ispartof: Proceedings of Workshop on Annotation and Exploitation of Parallel Corpora (AEPC) pages:73-82 ispartof: Workshop on Annotation and Exploitation of Parallel Corpora (AEPC) location:Tartu (Estonia) date:2 Dec - 2 Dec 2010 status: published |
