Generalizing Local Translation Models

We investigate translation modeling based on exponential estimates which generalize essential components of standard translation models. In application to a hierarchical phrasebased system the simplest generalization allows its models of lexical selection and reordering to be conditioned on arbitrary attributes of the source sentence and its annotation. Viewing these estimates as approximations of sentence-level probabilities motivates further elaborations that seek to exploit general syntactic and morphological patterns. Dimensionality control with `1 regularizers makes it possible to negotiate the tradeoff between translation quality and decoding speed. Putting together and extending several recent advances in phrase-based translation we arrive at a flexible modeling framework that allows efficient leveraging of monolingual resources and tools. Experiments with features derived from the output of Chinese and Arabic parsers and an Arabic lemmatizer show significant improvements over a strong baseline.