Automated Sense Tagging via Monosemous Anchors

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 219 papers24 conceptsfrozen 2026-10-04

The "knowledge acquisition bottleneck" has long been a primary obstacle in Natural Language Processing. To teach a machine to distinguish between the "bank" of a river and a "bank" for money, researchers traditionally needed thousands of sentences manually tagged by human linguists. However, manual annotation is slow, expensive, and difficult to scale.

A recurring solution across decades of research is the use of monosemous anchors: leveraging words that have only one possible meaning to automatically generate labeled training data from raw, unannotated text. By identifying these "unambiguous" words, researchers can create high-quality "silver-standard" datasets that simulate the complexity of human language without the human price tag.

The Intuition of the Monosemous Proxy

The core insight, pioneered by researchers like Rada Mihalcea and Dan Moldovan in An Automatic Method for Generating Sense Tagged Corpora, is that while many words are ambiguous (polysemous), many others are not.

If you find the word "sofa" in a sentence, you know exactly what it means because it has only one sense. If "sofa" is a synonym for a specific sense of the ambiguous word "couch," then the context surrounding "sofa" provides a "pure" signal for that specific meaning. By retrieving thousands of sentences containing "sofa" and programmatically replacing it with "couch," a model can be trained to recognize the "furniture" sense of "couch" using naturally occurring data.

This approach transforms Word Sense Disambiguation (WSD) from a manual labeling task into a data-mining task. As Juanzi Li and Changning Huang demonstrated in A Model for Word Sense Disambiguation, the contextual patterns of these single-meaning words can define the boundaries of semantic classes, allowing models to learn the "shape" of a meaning without a human ever drawing the lines.

Creating Synthetic Ambiguity

While using single words is effective, researchers also developed "Equivalent Pseudowords" (EPs) to create more robust training signals. In An Equivalent Pseudoword Solution to Chinese Word Sense Disambiguation, Haifeng Wang and colleagues showed how to construct artificial ambiguous words by grouping several monosemous synonyms together.

Imagine creating a fake word, "BANK_PROXY," which represents both "financial institution" and "sloping land." By training a classifier to decide which monosemous synonym actually belongs in a given context, the model learns the features of those distinct senses. This "synthetic ambiguity" allows for the training of high-performance classifiers on massive raw corpora, often outperforming supervised models limited by small, hand-tagged datasets.

Scaling to Entity Linking and Social Media

The same logic applies to Entity Linking—the task of mapping a name like "Abbott" to a specific database entry (e.g., the pharmaceutical company vs. the politician). Wei Zhang and Chew Lim Tan extended this in Entity Linking Leveraging Automatically Generated Annotation by identifying unambiguous entity mentions in news text and replacing them with their ambiguous synonyms or acronyms.

This technique is particularly powerful in noisy environments like Twitter. Romaric Besançon and Omar Adjali have used unique social media handles as anchors. In Building a Multimodal Entity Linking Dataset From Tweets, they showed that a mention of a unique handle (like @Apple) can be programmatically anonymized to a surface form ("Apple") to create a ground-truth training instance. Because the original handle was unique, the "correct" link is known with 100% certainty, providing a scalable way to train multimodal models that use both text and images for disambiguation.

Modern Relevance: Seeding the Latent Space

Even in the era of large language models and Transformers, monosemous anchors remain vital. Modern WSD systems often struggle with "long-tail" senses that rarely appear in labeled data. Daniel Loureiro and José Camacho-Collados addressed this in Don't Neglect the Obvious by creating the UWA (Unambiguous Word Annotations) dataset.

By automatically labeling millions of monosemous word occurrences in web text, they created "seed" embeddings. These high-precision representations of specific meanings propagate through semantic networks, helping Transformer-based models better understand the nuances of polysemous words. This demonstrates that the "obvious" occurrences of unambiguous words are not just noise—they are the foundation upon which complex semantic understanding is built.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 15 papers
  2. Version 1 · 2026-10-04 · 19 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_monosemous_proxy_labeling_76dafc82,
  title = {monosemous proxy labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/monosemous-proxy-labeling/rdr_76dafc82eaf04d14b793c599fd85452b},
  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_automated_sense_tagging_via_monosemous_a_76dafc82_v1,
  title = {Automated Sense Tagging via Monosemous Anchors},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-sense-tagging-via-monosemous-anchors/rdr_76dafc82eaf04d14b793c599fd85452b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}