Hierarchical Rule Generalization with Automatically Derived Multi-level Word Classes for MT

Hierarchical rule generalization is a central aspect of Chiang’s hierarchical decoding model. “Hierarchical substitution” is the process by which generalized rules are used in decoding. We believe that hierarchical substitutions of semantically similar words are more likely to be grammatically well-formed than hierarchical substitutions of arbitrary words. In this paper, we extend the rule generalization process to use a hierarchy of semantic bilingual word classes, which are derived automatically using a bilingual semantic similarity score. The result is that during decoding, hierarchical substitution of semantically specific classes (e.g., the class of colors) will have a higher probability than hierarchical substitution of more general classes (e.g., the class containing all words). Although we achieved a substantial increase in the number of generalized rules used during decoding, more investigation is required to determine how this will improve MT scores (BLEU) over those generated by the standard rule generalization process.