Abstract
Lexical ambiguity is a major hurdle in Natural Language Processing (NLP), particularly the distinction between polysemy (related senses) and homonymy (unrelated senses). This paper formalizes and empirically validates four "repulsion" hypotheses which posit that homonyms almost never overlap in translation, discourse, or collocation. Using etymological data as a ground truth, we demonstrate that these hypotheses hold with over 99% accuracy across multiple corpora, providing a theoretically grounded method for creating coarse-grained sense inventories.
Introduction
In the field of lexical semantics, researchers distinguish between polysemy—where a word has multiple related meanings (e.g., "bank" as a financial institution vs. the building itself)—and homonymy—where a word has unrelated meanings that happen to share a spelling (e.g., "bank" as a river edge vs. a financial institution).
Current NLP tools, including Word Sense Disambiguation (WSD) and Neural Machine Translation (NMT) systems, often struggle with these distinctions. Standard resources like WordNet are frequently criticized for being too fine-grained, which hampers system accuracy. This research addresses the gap by providing a computational framework to distinguish homonymy from polysemy, facilitating the creation of more effective "coarse-grained" sense inventories.
The Repulsion Hypotheses
The core of this work is the proposal of four hypotheses based on the observation that homonyms "repel" each other in linguistic usage. These are formalized as follows:
- One Homonym per Translation (OHPT): Distinct homonyms of a word will have disjoint sets of translations in another language. If two senses share a translation, they are likely polysemous (related), not homonymous.
- One Homonym per Discourse (OHPD): Within a single document or conversation, a word will represent only one of its homonyms.
- One Homonym per Collocation (OHPC): In a specific local context (neighboring words), only one homonym of a word is likely to appear.
- One Homonym per Sense Cluster (OHPSC): Any valid clustering of word senses should never group senses from different homonyms together.
Methodology and Data
To test these hypotheses, the authors established a "gold standard" using Type-A homonyms—those with distinct etymological origins. For example, "sole" (fish) originates from Latin solea, while "sole" (only) comes from Latin solus. By mapping these etymologically distinct homonyms to WordNet senses, the researchers could objectively evaluate their hypotheses against annotated corpora without relying on subjective speaker intuition.
The study utilized several key datasets:
- SemCor: A large sense-annotated English corpus for discourse and collocation testing.
- MultiSemCor (MSC) & JSemCor (JSC): Parallel corpora (English-Italian and English-Japanese) for testing translation-based repulsion.
- OntoNotes: Used to evaluate the sense clustering hypothesis.
Results and Discussion
The empirical evidence for the repulsion hypotheses was remarkably strong. Across nearly all tests, the hypotheses were supported at levels exceeding 99%.
The slightly lower support for the collocation hypothesis (OHPC) at 97.9% was largely attributed to limitations in the WSD system (IMS) used for testing rather than a flaw in the underlying theory. Notably, when data appeared to violate the hypotheses, manual inspection revealed that the errors almost always resided in the underlying datasets (e.g., WordNet or SemCor) rather than the theory itself.
Conclusion
This research demonstrates that homonyms act like particles with identical charges, constantly pushing away from each other in usage. The OHPT hypothesis, in particular, suggests that shared translations are a sufficient condition for polysemy, offering a concrete path toward automating the grouping of related senses. Future work will focus on identifying "Type-B" homonyms (senses that split over time) and developing operational methods to generate high-quality, coarse-grained sense inventories to improve NMT and WSD performance.
Summary updated