📄 Word-Sense Disambiguation for Machine Translation

Summary

Classic word sense disambiguation (WSD) struggles for two reasons. Fine-grained inventories such as WordNet separate senses that are hard to tell apart, and hand-labelled training data is scarce. The authors reframe the problem as word translation: given the context, predict which target-language word or phrase a source word should become. Bilingual corpora then supply large amounts of partially labelled data almost for free. Using the European Parliament proceedings (English→French), they align words with GIZA++ (IBM Model 4 output) in both directions and keep only candidate translations that occur frequently enough. The resulting examples are treated as a labelled corpus for a relatively simple discriminative classifier that uses source-side context. The model clearly beats a baseline that is notoriously hard to surpass in WSD. It also helps on a simplified translation task, a newly proposed blank-filling evaluation. Its scores can also shrink candidate translation sets much more effectively than alignment models can. This matters because standard source-channel MT systems discard most source-side context.

Key points

• Training labels are derived automatically from Europarl. Word alignments come from GIZA++ (IBM Model 4) run S→T and T→S, and rare aligned candidates are dropped to reduce alignment noise.
• No stemming was applied, so suffix information remains available. As a result, the label sets mix sense distinctions with grammatical-form distinctions. For example, the candidates for English "rise" include montée, hausse, augmentation, lever and others.
• The paper introduces a blank-filling task to evaluate word translation's effect on MT separately from syntax and other factors. On this task, higher word-translation accuracy yields better translation.
• In candidate pruning at 95% recall, the cutoff strategy needs an average set size of 4.2 with the word-translation model, versus 6.7 with a discriminative alignment model. A generative alignment model performed worst.
• The authors note that "hard" integration does not improve MT, such as forcing the decoder to use the top choice (Carpuat and Wu, 2005). They argue instead for soft scoring or reranking.

Interesting ideas

• I1. Define labels by a downstream task instead of an ontology. Treating "senses" as observed translations removes distinctions that no application needs and makes annotation essentially free. This could transfer to Ukrainian WSD by using Ukrainian–English parallel corpora.
• I2. Bidirectional alignment plus a frequency threshold as a noise filter. Requiring agreement across both alignment directions and a minimum count is a cheap way to turn noisy automatic alignments into usable training labels. The same trick applies to any projection-based data creation.
• I3. Isolating a component with a controlled proxy task. The blank-filling setup measures one module's contribution to a larger system without confounds from the rest of the pipeline. This is a useful evaluation pattern whenever end-to-end metrics are too noisy.
• I4. Evaluating a model as a pruner via recall vs. candidate-set size. Plotting how often the correct answer survives against average set size shows how much a classifier can shrink a search space. This is relevant for speeding up decoders or retrieval pipelines.
• I5. Acknowledging single-reference evaluation limits. Accuracy against one observed translation underestimates quality when synonyms are interchangeable. Keeping this in mind helps when interpreting scores built from automatically projected gold labels.