Choosing Sense Distinctions for WSD: Psycholinguistic Evidence

Susan Windisch BrownACL 2008

Introduction

Word Sense Disambiguation (WSD) is a fundamental task in Natural Language Processing (NLP) that involves identifying the correct meaning of a word in context. While resources like WordNet 2.1 provide highly detailed sense distinctions, these fine-grained definitions often pose a challenge for both human annotators and automated systems. Human agreement on WordNet senses typically caps at around 70%, suggesting that many distinctions may be too subtle for practical use. This study investigates whether these fine-grained distinctions reflect actual cognitive representations or if humans naturally group related senses together.

Experimental Design

To explore how the human brain processes different word senses, a semantic decision task was conducted. The study categorized verb-phrase pairs into four distinct groups based on their level of relatedness, using WordNet 2.1 for fine-grained distinctions and the Oxford English Dictionary (OED) to identify homonyms.

Examples of verb-phrase pairs used in the semantic decision task across four levels of relatedness.

Participants were presented with a "prime" phrase and asked to determine if it made sense. Immediately following, they were shown a "target" phrase using the same verb. The researchers measured response latency (speed) and accuracy. The underlying hypothesis was that if two senses share a mental representation, the prime phrase would facilitate the processing of the target phrase, resulting in faster and more accurate responses.

Results

The experimental data revealed a clear linear progression in processing difficulty as the relationship between senses became more distant.

Mean response latencies (ms) for target phrases across the four experimental conditions.
  • Same Sense: Participants were fastest when the sense remained identical (1056ms).
  • Closely Related: Senses involving different concrete objects (e.g., "broke the glass" vs. "broke the radio") were processed nearly as efficiently as identical senses, with a significant speed advantage (173ms) over more distant categories.
  • Distantly Related and Unrelated: These categories showed significantly higher latencies and lower accuracy. While "Same Sense" pairs achieved 95.6% accuracy, "Distantly Related" and "Unrelated" pairs dropped to an average of 72%.

The findings suggest that the "cost" of switching between senses is negligible for closely related meanings, indicating that these senses likely overlap within the same mental lexicon entry.

Discussion and Implications

The results provide psycholinguistic justification for clustering fine-grained WordNet senses in NLP applications. The human mind appears to represent word meanings not as a discrete list of definitions, but as a semantic space where closely related concrete usages are treated as essentially the same.

A key observation from the study is that "closely related" senses often involve different concrete objects, whereas "distantly related" senses typically involve a shift from concrete to metaphorical usage. For WSD systems to better mimic human understanding, researchers should prioritize coarse-grained distinctions that reflect these cognitive boundaries. Future work should focus on identifying the specific semantic features—such as animacy or agency—that trigger the brain to differentiate between word senses.

Summary updated

BibTeX
@inproceedings{brown2008choosing,
  author = {Susan Windisch Brown},
  title = {{Choosing Sense Distinctions for WSD: Psycholinguistic Evidence}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2008},
}
Figure