Unification-Based Glossing

We present an approach to syntax-based machine translation that combines uniication-style interpretation with statistical processing. This approach enables us to translate any Japanese newspaper article into English, with quality far better than a word-for-word translation. Novel ideas include the use of feature structures to encode word lattices and the use of uniication to compose and manipulate lattices. Uniication also allows us to specify abstract features that delay target-language synthesis until enough source-language information is assembled. Our statistical component enables us to search eeciently among competing translations and locate those with high English uency. 1 Background JAPANGLOSS Knight et al., 1994; 1995] is a project whose goals are to scale up knowledge-based machine translation (KBMT) techniques to handle Japanese-English newspaper MT, to achieve higher quality output than is currently available, and to develop techniques for rapidly constructing MT systems. We built the rst version of JAPANGLOSS in nine months and recently participated in an ARPA evaluation of MT quality White and O'Connell, 1994]. JAPANGLOSS is an eeort within the larger PANGLOSS NMSU/CRL et al., 1995] MT project. Our approach is to use a KBMT framework, but to fall back on statistical methods when knowledge gaps arise (as they inevitably will). We syntactically analyze Japanese text, map it to a semantic representation, then generate English. Figure 1 shows a sample translation. Parsing is bottom-up, driven by an augmented context-free grammar whose format is roughly like that of Shieber, 1986]. Our grammar rules look like this: OUTPUT: The new company plans to establish in February. The semantic representation contains conceptual tokens drawn from the 70,000-term SENSUS ontology Knight and Luk, 1994]. Semantic analysis proceeds as a bottom-up walk of the parse tree, in the style of Mon-tague and Moore Dowty et al., 1981; Moore, 1989]. Semantics is compositional, with each parse tree node assigned a meaning based on the meanings of its children. Leaf node meanings are retrieved from a semantic lexicon , while meaning composition rules handle internal nodes. Semantic rules and lexical entries are sensitive to syntactic structure,