Resolving Lexical Polysemy in Machine Translation

28 concepts28 papers2 proposals

When we translate language, we rarely translate words in isolation. Consider the word bank. If you see it next to river, it means a slope of land; next to money, it means a financial institution. This phenomenon—where a single word possesses multiple distinct meanings—is known as lexical polysemy.

For machine translation (MT) systems, polysemy represents a fundamental hurdle. A system that defaults to a word's most common translation will frequently produce outputs that are fluent but semantically incorrect. Resolving this requires models to move beyond surface-level text matching and develop mechanisms for contextual word sense selection.

Why Local Context Fails

Early statistical machine translation (SMT) systems attempted to resolve ambiguity by looking at immediate neighbors, relying on local n-gram context windows and fixed phrase translation tables. However, these local models are fundamentally limited. As Marine Carpuat demonstrated, traditional phrase-based MT systems are consistently outperformed by dedicated Word Sense Disambiguation (WSD) models, proving that local context windows fail to capture broader semantic nuances.

Because these traditional mechanisms rely on static probability tables, they cannot adapt when a source word requires a completely different target translation based on its specific sentence-level usage (Marine Carpuat et al.; Deyi Xiong et al.). For instance, if the crucial clue that disambiguates a word sits at the beginning of a sentence, a local phrase window at the end of the sentence will miss it entirely.

To overcome this, researchers began explicitly injecting WSD classifiers into the translation pipeline. By using a broader semantic window to score translation candidates across the entire input sentence, systems could dynamically override static phrase tables (Marine Carpuat et al.). Yee Seng Chan et al. successfully integrated a state-of-the-art WSD system into a hierarchical phrase-based model, providing the mathematical proof that integrating wide-context semantic cues directly improves translation quality.

Injecting Semantic Structures and Embeddings

How do we give a mathematical model the "intuition" to choose the right word sense? One approach is to anchor translations to structured semantic frameworks.

Pascale Fung et al. addressed this by using an Expectation-Maximization (EM) algorithm to align bilingual semantic ontologies. By treating semantic frames as a universal "map" shared across languages, their system could determine whether an ambiguous word fit the conceptual frame of the surrounding sentence, drastically improving lexical selection. Rather than relying on fine-grained dictionary definitions, Deyi Xiong et al. introduced the use of "supersenses"—coarse-grained semantic categories from WordNet (such as noun.food or verb.motion). Integrating these high-level categories via neural embeddings allowed translation models to rule out conceptually impossible translation candidates.

[Source Sentence] ---> [Context-Sensitive Embedding] ---> [Dynamic Translation Selection]
                             ^
                             | (Influenced by global topic & surrounding tokens)

As representation learning matured, the field shifted from static tables to context-sensitive embeddings. Haiyang Wu et al. designed a bilingual semantic embedding model that integrates the entire surrounding source text into the vector representation of a phrase. Instead of a word having a single fixed vector, its representation shifts dynamically based on its neighbors. To capture both immediate and overarching context, Jinsong Su et al. introduced a Context-Aware Topic Model (CATM). The CATM models local sentence-level context alongside global document-level topics, recognizing that a document about "finance" heavily biases how a polysemous word like stock should be translated.

The Neural and Large Language Model Era

With the transition to Neural Machine Translation (NMT), explicit feature engineering gave way to implicit representation learning within deep neural networks. A natural question emerged: where exactly does an NMT model resolve polysemy?

Gongbo Tang et al. investigated this "black box" and discovered that the encoder layers bear the primary responsibility for word sense disambiguation. The encoder contextualizes ambiguous surface forms into distinct semantic representations before passing them to the decoder for generation. When this process fails, it results in severe translation errors, which Denise Díaz et al. categorized as foundational lexical-semantic failures in their large-scale qualitative audit of NMT systems.

Surprisingly, even modern Large Language Models (LLMs) with hundreds of billions of parameters still stumble over polysemy. Yichong Huang et al. identified a striking paradox: an LLM may possess the general knowledge to explain a complex, ambiguous word in a question-answering format, yet it will frequently default to a literal, word-for-word surface mapping when executing a translation. To fix this, the authors developed the DUAT framework, which forces the LLM to first generate an explicit cross-lingual interpretation of a difficult word, aligning its general understanding with its translation mechanism to achieve a true sense-for-sense output.

Cascading Failures in Downstream Tasks

When translation models fail to resolve polysemy, the errors cascade into downstream applications that rely on cross-lingual alignment.

  • Cross-Language Information Retrieval (CLIR): If a user searches for documents in another language, translating the query incorrectly due to polysemy will retrieve entirely irrelevant data. Jianqiang Wang et al. demonstrated that introducing human-in-the-loop back-translations and usage examples can boost retrieval effectiveness by up to 48% by correcting these automatic translation errors. Furthermore, Longyue Wang et al. proved that translating entire documents rather than short queries resolves this "context gap," giving the underlying MT engine enough linguistic context to accurately disambiguate search terms.
  • Automated Lexicon and Dataset Generation: Building massive multilingual semantic networks like BabelNet (Roberto Navigli et al.) or automatically constructing Wordnet synsets (Khang Nhut Lam et al.) is frequently bottlenecked by polysemy noise. In pivot-based dictionary generation (e.g., using English to bridge Basque and Spanish), high-frequency terms present a major mathematical challenge. As one study noted, common words appear in so many diverse contexts that their distributional signatures become indistinct, allowing incorrect translation candidates to corrupt the lexicon. Similarly, Gregor Geigle et al. showed that translating short, isolated dataset labels (like those in ImageNet) without surrounding context introduces severe semantic mismatches, requiring careful filtering or manual intervention.
  • Specialized Domains and Code-Switching: In professional medical indexing, Yuka Tateisi et al. found that translating isolated technical terms fails because it strips away the contextual cues required to select the correct word sense; translating full sentences is mandatory for accuracy. Meanwhile, in cross-lingual spoken language understanding, Zhihong Zhu et al. emphasized that substituting keywords across languages without a context-aware polysemy mechanism creates semantic mismatches that fundamentally confuse a model's internal representation of user intent.

Ultimately, resolving lexical polysemy requires translation systems to treat words not as fixed tokens with static equivalents, but as dynamic variables whose true meanings can only be decoded by looking at the entire linguistic environment.

Key Foundational Work

Pooling and Threading in Translation Term Banks (1982) — contributed 1 concept elements to this direction.
Une troisieme version du Dicautom (1967) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_lexical_polysemy_in_translation_753da456,
  title = {lexical polysemy in translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-polysemy-in-translation/rdr_753da456bd304923a10608cd7654acb8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_lexical_polysemy_in_translation_753da456_v1,
  title = {lexical polysemy in translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-polysemy-in-translation/rdr_753da456bd304923a10608cd7654acb8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake