📄 Machine Translation: A View from the Lexicon

Summary

This item is not an original research article. It is Daniel Radzinski's critical review of Bonnie Jean Dorr's 1993 MIT Press book, which describes UNITRAN, an interlingual machine translation system. UNITRAN uses Chomskyan government-binding (GB) theory for syntax and Jackendoff's lexical conceptual structure (LCS) for lexical semantics; LCS also serves as the interlingua. Dorr claims that lexical-semantic mismatches among English, Spanish, and German fall into roughly seven divergence classes. In her design, these are resolved by a parameterized mapping instead of language-specific rules.

Radzinski finds the book interesting but doubts that the approach is promising. He uses Wall Street Journal sentences to show that the system's GB constraints wrongly reject verb–PP–object orders that are common in real English. He also criticizes several other features:

• the system can output an ungrammatical subjunctive form when translating German to Spanish;
• the lexicon is very small;
• the test sentences are short;
• some core constructions are not handled;
• some divergences exist only because of the chosen representation.

He considers the tense/aspect extension of LCS the most valuable contribution. For NLP practitioners, the review shows why grammar claims should be checked against corpus evidence.

Key points

• UNITRAN combines GB syntax with an extended LCS interlingua for English, Spanish, and German. Divergence types it handles include categorial, conflational, demotional, and promotional.
• The coverage gaps are substantial. The morphological lexicon holds only 305 English verbal roots, and the reviewer suggests roughly ten times that for a basic system. Most test sentences are under ten words, and existential there constructions are not supported.
• Translating the five-word German sentence Johann brach ins Zimmer ein into Spanish reportedly took close to 3 hours.
• The counterexamples to the verb–PP–object ban come from Wall Street Journal material on the ACL/DCI CD-ROM I.
• Chapter 9 adds a telicity parameter that distinguishes ransack, destroy, and obliterate. It also gives corpus evidence linking LCS to an established aspectual classification, which the reviewer considers promising.