Semantic Matching for Contextual Word Sense Disambiguation

This direction is split into other directions: dual-encoder similarity scoring. Its last version stays readable and citable here.

Version 1 of 126 papers28 conceptsfrozen 2026-10-04

Traditional Word Sense Disambiguation (WSD) was long treated as a multi-class classification problem: given a word like "bank," a model had to choose the correct ID from a fixed inventory like WordNet (e.g., bank%1:14:00 for a financial institution vs. bank%1:17:00 for a river side). This approach, however, collapses when faced with rare senses, new languages, or the "long tail" of human expression.

A new direction in NLP moves away from these rigid, categorical buckets. Instead, researchers are reframing word meaning as a semantic matching task. By comparing a word’s context directly against dictionary definitions (glosses) or other usage examples, models can understand polysemy through relationship rather than memorization.

The Relational Shift: Word in Context

The foundation of this shift is the "Word-in-Context" (WiC) framework. As Mohammad Taher Pilehvar and José Camacho-Collados argue in WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations, the core of the problem isn't labeling a word, but determining if its meaning is the same across two different sentences.

This relational approach removes the need for a model to "know" every possible sense of a word beforehand. Instead, it evaluates the model’s ability to generate contextualized embeddings that are sensitive enough to shift when the meaning shifts. This framing has been extended to multilingual settings by Alessandro Raganato and colleagues in XL-WiC, allowing models to evaluate semantic boundaries in languages where fixed sense inventories don't even exist.

Dual-Encoder Architectures for Gloss Matching

If we no longer use fixed IDs as labels, how does a model "pick" a meaning? The dominant mechanism is the dual-encoder (or bi-encoder) architecture.

In this setup, the model uses two separate pathways:

  1. Context Encoder: Processes the sentence containing the target word to produce a vector representing that specific usage.
  2. Gloss Encoder: Processes the text of a dictionary definition (the "gloss") to produce a vector representing the abstract concept.

The model then calculates the similarity between these two vectors. If the context vector for "the key to the door" is closer to the gloss vector for "a metal instrument" than to "a crucial element," the model has successfully disambiguated the word.

Researchers like Hongming Zhang and Yangqiu Song have used this to create mBERT-UNI, a framework that maps different languages into a unified sense space. By matching a French sentence to an English definition, the model can transfer knowledge across languages. Similarly, Wenlin Yao and the Tencent AI Lab team developed Connect-the-Dots, which treats words and their definitions as semantic "surrogates," allowing the model to learn the general relationship between usage and definition.

Solving the Long Tail and Zero-Shot Bottlenecks

The primary advantage of matching over classification is its ability to handle the "long tail" of language. Traditional models struggle with Zipf’s Law—the fact that most word senses appear very rarely in training data.

By treating WSD as a matching task, models can perform zero-shot disambiguation. If a model understands how to match a context to a definition for common words, it can apply that same logic to a word it has never seen before, provided it has access to a dictionary. Hongming Zhang and Tong Zhang 0001 demonstrated this in Rare and Zero-shot Word Sense Disambiguation using Z-Reweighting, showing that reweighting training data to focus on these rare instances significantly boosts performance without losing accuracy on common words.

This flexibility extends beyond standard vocabulary:

  • Acronyms: Daphné Chopard and Irena Spasić applied this to medical acronyms, matching ambiguous abbreviations to their full expansions.
  • Metaphors: Hai Wan and colleagues used gloss-based interpretations in MDGI-Joint to detect when a word's contextual meaning has drifted far from its literal dictionary definition.
  • Entities: The BED model by Nantao Zheng uses the same dual-encoder logic to link entities in text to descriptions in a knowledge base, outperforming static ID-based systems.

Why It Works: Interaction and Relevance

While dual-encoders are fast, some researchers argue that the model needs to see the context and the definition at the same time to capture fine-grained nuances. Boon Peng Yap and Eng Siong Chng explored this in Adapting BERT for Word Sense Disambiguation, framing WSD as a relevance ranking task. By feeding the context and gloss into a single encoder (a cross-encoder), the model can attend to specific words in the definition that explain the specific usage in the sentence.

Whether using the speed of dual-encoders or the precision of cross-encoders, the research direction is clear: understanding a word is no longer about picking a number from a list. It is about the dynamic act of matching a specific moment of language to the broader map of human knowledge.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_matching_for_contextual_word_se_0ad7ebff,
  title = {Semantic Matching for Contextual Word Sense Disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-matching-for-contextual-word-sense-disambiguation/rdr_0ad7ebff316c482188c2dbc3a64f020d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_semantic_matching_for_contextual_word_se_0ad7ebff_v1,
  title = {Semantic Matching for Contextual Word Sense Disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-matching-for-contextual-word-sense-disambiguation/rdr_0ad7ebff316c482188c2dbc3a64f020d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}