📄 Verb Semantics and Lexical Selection

Summary

The paper looks at how machine translation chooses a target verb, using English verbs translated into Mandarin. Transfer-based systems list verb pairs with argument restrictions in a bilingual dictionary, and this breaks down for very polysemous verbs such as break. Mandarin has no general "break" verb. It needs serial verb compounds that name both the action and the resulting state of the object. The authors tested the commercial system TranStar on Brown corpus sentences with break and found low accuracy. They argue that selection restrictions only give a default reading, which context often overrides. Their proposal treats each verb as a set of concepts spread across several conceptual domains, drawing on Levin's domains such as motion, force, contact and change-of-state. Placing these concepts in a hierarchy makes a similarity measure possible, so the system can pick a good translation even when no exact lexical match exists. Their prototype, UNICON, is much more accurate than TranStar. Relaxed similarity-based matching also handles some metaphorical uses.

Key points

• TranStar baseline on break:
  • The Brown corpus has 246 sentences with break as the main verb. After removing idioms and particle constructions, 157 remain.
  • TranStar has only 13 Chinese entries for break. It translated just 30 of the 157 sentences correctly (19.1%), mostly falling back on dasui as a default.
  • A human translator used 68 distinct Chinese expressions for the same sentences. The distribution looks Zipf-like: a few verbs are very frequent, with a long tail of rare ones.
• Evidence against selection restrictions: Mandarin VA compounds (action + result, e.g. da-sui, "hit-in-pieces") require the action to be stated explicitly. Selection restrictions predict the result state fairly well but say little about the action.
• UNICON setup: 21 English verbs from Levin's hit, touch, break and cut classes. The data were 400 Brown corpus sentences, 100 of which were used for training.
• Test set one (154 sentences with concrete patients): 57.8% correct at first, rising to 99.45% once unknown argument words were given conceptual meanings.
• Test set two (116 sentences, including non-concrete objects and metaphors):
  • 31% correct at first.
  • 75% after adding meanings.
  • 88.8% with the extended, constraint-relaxing selection, a gain of 13.8 points.

Interesting ideas

• I1. Multi-domain verb definitions: A verb is defined by several concepts in different domains rather than one node in a single tree. This fits the fact that verbs often combine motion, force, contact and result. It could help any lexical resource where one strict hierarchy cannot hold word meanings.
• I2. Similarity fallback instead of a fixed candidate list: When no exact match exists, the system picks the nearest lexicalized concept in the hierarchy. This addresses the long tail of rare usages that dictionaries never fully cover. The same strategy fits low-resource language pairs.
• I3. Restrictions as soft defaults: Argument restrictions are treated as defaults that can be relaxed, with conceptual similarity taking priority. This lets the system handle metaphorical uses, such as bonds "hitting" a price peak. The approach is useful wherever strict type constraints are too brittle.
• I4. Checking the frequency shape of translations: Comparing how translations are distributed in a system versus a human reference reveals whether the system covers the tail. It is a cheap diagnostic for any translation lexicon.
• I5. Separating result-state choice from action choice: Result states can be predicted fairly well from properties of the object, while actions depend on context. Splitting the decision this way could apply to other languages with resultative or serial verb constructions.