Word Sense Disambiguation (WSD)—the task of determining which "sense" of a word is meant in a given context—has long been stymied by the "knowledge acquisition bottleneck." Supervised machine learning models perform well on WSD, but they require vast amounts of manually tagged data. Labeling every instance of "bank" as either a financial institution or a river slope is tedious and expensive.
A recurring insight across research by Hwee Tou Ng, Zhi Zhong, and others is that we do not need to label this data by hand. Instead, we can treat professional translators as implicit annotators. By leveraging parallel corpora—large datasets of the same text in two languages—researchers can automatically project semantic labels from a target language back to a source language.
Translation as Implicit Disambiguation
The core intuition is that polysemy is rarely identical across languages. A word that is ambiguous in English is often translated into distinct, unambiguous words in another language like Chinese or French.
Consider the English word interest. In a financial context, it might translate to the Chinese word 利息 (lìxī). In a hobbyist context, it might translate to 兴趣 (xìngqù). As Hwee Tou Ng and Yee Seng Chan demonstrate in Word Sense Disambiguation with Distribution Estimation, these distinct translations serve as "proxy labels." If we have a million sentences of English-Chinese parallel text, we can automatically label every English instance of interest based on its Chinese counterpart.
This approach transforms WSD from a manual labeling problem into an alignment problem. By using bilingual dictionaries and word-alignment algorithms, researchers can generate "silver-standard" training data at a scale that manual effort could never reach.
Scaling to All Words
Early WSD systems were often "lexical samples," focusing on a small handful of difficult words. However, the goal of modern NLP is "all-words" WSD. In Word Sense Disambiguation for All Words without Hard Labor, Hwee Tou Ng and Zhi Zhong showed that this cross-lingual projection method could scale to the entire vocabulary of a language.
The process generally follows three steps:
- Alignment: Align words in a parallel corpus (e.g., English-Chinese).
- Mapping: Map the target language translations to a specific sense inventory, like WordNet. For example, a human expert might define a rule that says when bank translates to 岸 (àn), it corresponds to WordNet Sense #2 (river bank).
- Training: Use these automatically labeled examples to train a standard supervised classifier, such as a Support Vector Machine (SVM).
This methodology was the engine behind the It Makes Sense (IMS) system (It Makes Sense: A Wide-Coverage Word Sense Disambiguation System for Free Text), which became a standard tool in the NLP community by providing pre-trained models boosted by parallel corpora.
Beyond Nouns: Modals and Cognates
The power of cross-lingual projection extends beyond simple nouns. Oana Frunza and Diana Inkpen applied a similar "bilingual bootstrapping" technique to disambiguate partial cognates—words that look the same in two languages but only share some meanings (Semi-Supervised Learning of Partial Cognates Using Bilingual Bootstrapping). For instance, the French word facteur can mean "factor" or "mailman." By looking at how these words are translated into English, the system can learn to distinguish the "true friend" from the "false friend."
Similarly, Mengfei Zhou, Alexis Palmer, and Annemarie Friedrich used this logic for Modal Sense Classification (Semantically Enriched Models for Modal Sense Classification). They used "bridge languages" to determine if a word like must indicates a logical necessity (epistemic) or an obligation (deontic), proving that translation captures even subtle grammatical nuances.
Why It Works
The success of these systems, such as the top-performing NUS-PT at SemEval-2007 (NUS-PT: Exploiting Parallel Texts for Word Sense Disambiguation in the English All-Words Tasks), suggests that translation is a high-quality form of semantic feature engineering.
When a translator chooses a specific word in a target language, they have already performed the mental work of disambiguation. By aligning these texts, we are essentially "back-projecting" human intelligence into our training sets. This bypasses the need for manual tagging and allows models to adapt to new domains by simply finding parallel text relevant to that field, effectively solving the domain-shift problem that plagues traditional WSD.