📄 Two Languages Are More Informative Than One

Summary

The paper addresses lexical ambiguity in unrestricted text, focusing on target word selection in machine translation: picking the target-language equivalent that best fits a source word in its context. The authors avoid hand-built semantic knowledge and sense-tagged source corpora. Instead, they make the decision on the target side.

Each ambiguous source sentence is reduced to its syntactic lexical relations, such as verb–object, subject–verb, noun compounds and noun–PP. A bilingual lexicon expands each relation into every possible target-language version. These candidates are then counted in a large, untagged, parsed English corpus of about 150 million words. A statistical model uses a bound on the odds ratio between alternatives to decide which relation to trust first and which translation to choose.

On a random sample of Hebrew foreign-news sentences, the model applied to 70% of ambiguous words and was right in 92% of those cases. A German set was also tested. The study shows that cheap monolingual statistics in a second language can replace costly sense annotation.

Key points

• Motivating example. The Hebrew verb 'laḥtom' can mean sign, seal, finish or close, and 'hoze' can mean contract or treaty.
  • Choosing the most frequent English word alone gives the wrong output "closing a peace contract".
  • Relation counts fix this: 'peace treaty' appeared 49 times and 'peace contract' never. 'Sign a treaty' appeared 79 times, while each competing verb appeared at most twice.
• Counting method. Relation frequencies come only from sentences in the English corpus (~150M words) that the parser handled successfully. When a sentence has several ambiguous words, they are resolved one at a time, in an order set by the statistical confidence of each relation.
• Evaluation. There were two example sets, one Hebrew→English and one German→English. On the randomly drawn Hebrew news set, applicability was 70% and precision 92%.
• Error categories.
  • Trusting the wrong relation. For 'matzav' (state/position), put-position (320) outweighed put-state (18). The modifier relation, state-war (13) vs position-war (2), pointed to the correct choice.
  • Mixing senses. The target corpus ignores the source word, so counts for one Hebrew word can leak into another (e.g., 'young prospect' affecting 'sikkuy' vs 'tiqva').
  • Needing deep understanding. Hebrew 'hagira' covers both emigration and immigration, and only the wider context could tell them apart.
• Authors' explanation of the high precision. Correct readings are a small fraction of the possible ones (sparseness), and independent cues tend to agree with each other (redundancy).

Interesting ideas

• I1. Resolve ambiguity where it is cheaper. The source sentence is treated as a noisy signal, and a target-language model picks the most plausible output. This framing is useful whenever one side of a task has abundant raw text and the other does not, such as a lower-resource source paired with English.
• I2. Use another language as free sense labels. Distinct translations often line up with distinct senses of a source word. Counting relations among those translations therefore yields sense-level statistics for monolingual WSD, and a third language can help when two senses share a translation. This could bootstrap sense inventories without manual tagging.
• I3. Resolve the most confident decision first. The system ranks ambiguous relations by a confidence bound rather than raw counts, settles the surest one, and lets it constrain the rest. This greedy, uncertainty-aware ordering transfers to many joint-decision problems.
• I4. Weight relation types differently. The 'matzav' failure suggests that a noun's own modifiers or objects should count more for its sense than the verb governing it. Typed weighting of evidence by syntactic role is a cheap refinement for any count-based model.
• I5. Report coverage and precision separately and classify failures. Giving applicability (70%) and accuracy on covered cases (92%) as separate figures, plus grouping errors by cause, shows clearly where the method breaks down. It also frames future gains as an explicit recall/precision tradeoff.