Abstract
Verbs like make, have, and get present challenges for applications requiring automatic word sense discrimination. These verbs are both highly frequent and polysemous, with semantically “full” readings, as in make dinner, and “light” readings, as in make a request. Lexical resources like WordNet encode dozens of senses, making discrimination difficult and inviting proposals for reducing the number of entries or grouping them into coarser-grained supersenses. We propose a data-driven, linguistically-based approach to establishing a motivated sense inventory, focusing on make to establish a proof of concept. From several large, syntactically annotated corpora, we extract nouns that are complements of the verb make, and group them into clusters based on their Word2Vec semantic vectors. The results show that the clusters reflect an intuitively plausible sense discrimination of make. As an evaluation, we test whether words within a given cluster co-occur in coordination phrases, such as apples and oranges, as prior work has shown that such conjoined nouns are semantically related. Our results pave the way for a WordNet sense inventory that, while not inconsistent with the present one, would reduce it significantly and hold promise for improved automatic word sense discrimination.
Introduction: The Challenge of Light Verbs
In natural language processing, Word Sense Disambiguation (WSD) is the task of determining which "sense" of a word is being used in a specific context. This is notoriously difficult for "light verbs"—highly frequent verbs like make, have, and get that carry little semantic weight on their own and instead rely on their complements to form a predicate (compare "make dinner" vs. "make a request").
The primary challenge is that lexical resources like WordNet are often too detailed for practical computational use. WordNet lists 49 distinct senses for the verb make, a granularity that often confuses automated systems and human users alike. There is a significant need for a "coarser-grained" inventory of senses that is linguistically motivated but computationally practical.
Methodology: A Data-Driven Pipeline
The researchers developed a pipeline to discover these coarse senses automatically by focusing on the relationship between the verb and its direct object. The core hypothesis is that for light verbs, the "sense" is primarily determined by the noun it selects.
- Extraction: Using several Universal Dependencies (UD) corpora (including EWT, GUM, and PUD), the team extracted nouns acting as direct objects (the
objrelation) for the verb make. This allows the system to isolate the specific semantic contexts in which the verb operates.
- Embedding and Reduction: Each noun was converted into a 300-dimensional vector using pre-trained Word2Vec embeddings. To filter out noise and identify robust structures, Principal Component Analysis (PCA) was applied, identifying that the first 150 components captured the most vital information.
- Clustering: The team applied $k$-means clustering to these vectors. They tested a "fine" version with $k=30$ and a "coarse" version with $k=15$, informed by an "elbow" analysis of cluster inertia.
Results: Discovering Intuitive Supersenses
The $k=15$ model produced highly recognizable sense groupings that correspond to intuitive "supersenses." For example, one cluster focused on food-related items (vodka, soup, potato), while another centered on errors and failures (blunder, mistake, error).
Visualizing these clusters using PCA (as seen in Figure 5) reveals clear semantic neighborhoods. The model successfully distinguished between building physical objects (e.g., statues, sculptures) and more abstract creations or gains (e.g., money, progress).
Validation via Coordination
To validate the clusters without relying on the same embeddings used to create them, the authors employed "coordination structures"—phrases where two nouns are joined by and or or. The logic is that humans rarely conjoin semantically unrelated things (a phenomenon known as zeugma when used for comedic effect, such as "she made a salad and a mess").
Using the Corpus of Contemporary American English (COCA), the researchers found that nouns within the same cluster were significantly more likely to be conjoined than nouns from different clusters. For instance, in the "food" cluster, they found 93 instances of intra-cluster coordination (e.g., "lunch or dinner") compared to only 21 instances of coordination with outside clusters. Furthermore, using the Wu-Palmer similarity metric, nouns within the same cluster showed an average similarity of 0.60, compared to just 0.27 for nouns in different clusters.
Conclusion
This research demonstrates that the "sense explosion" of light verbs can be managed through data-driven clustering of their complements. By reducing 49 senses into a handful of linguistically motivated clusters, the researchers provide a path toward more efficient Word Sense Disambiguation. Future work will involve scaling this method to other light verbs like take or give and integrating these findings into a revised, more practical WordNet inventory.
Summary updated