Introduction
In the field of Natural Language Processing (NLP), Word Sense Disambiguation (WSD) remains a critical challenge. Identifying whether the verb "express" refers to conveying a feeling or mailing a package is foundational for higher-level tasks like event coreference and information extraction. Traditionally, high-performing WSD systems have relied on supervised learning, which necessitates massive, human-annotated datasets that are both expensive and time-consuming to produce.
Researchers at Carnegie Mellon University have proposed a "knowledge-lean" alternative. Their approach leverages distributional semantics—the principle that words used in similar contexts share similar meanings—to disambiguate verb senses without requiring a labeled training corpus. By combining a massive syntactic database called PropStore with dictionary definitions from OntoNotes, the system identifies event types based on the unique syntactic "signature" of a verb's usage.
The PropStore Framework
The proposed pipeline operates through three primary stages: building a relational database, extracting local syntactic context, and scoring potential senses using vector-based similarity.
1. Building the Triple Store
The foundation of the system is the PropStore, a "triple store" generated by parsing the entirety of Wikipedia using the Fanse Dependency Parser. It stores relations in the format <word1, relation, word2>. For example, it records every instance where a specific verb takes a specific noun as a direct object or is modified by a particular prepositional phrase. This provides a massive repository of real-world verb usage patterns.
2. Extracting Syntactic Signatures
When the system encounters a target verb in a sentence, it extracts a "signature"—a specific slice of the dependency parse tree. This signature typically includes the verb, its direct object, and its prepositional modifiers. This structural "slot" is then used to query PropStore for other verbs that appear in identical syntactic environments across Wikipedia.
3. Sense Scoring via Distributional Semantics
Once PropStore returns a list of candidate verbs that fit the signature, the system must map these candidates to the correct sense of the target verb. It utilizes OntoNotes as a sense dictionary, extracting the short, human-readable definitions for each possible meaning. Using word2vec (CBOW model), the system calculates the semantic similarity between the candidate verbs from PropStore and the words within the OntoNotes definitions. The sense that achieves the highest aggregate similarity score is selected as the correct disambiguation.
Qualitative Analysis: The Case of "Express"
The effectiveness of this approach is best illustrated through the disambiguation of the verb "express." In the sentence "They express these packages to corporate headquarters," the system extracts a signature where "express" takes "packages" as a direct object and is followed by the preposition "to."
Querying PropStore with this signature returns candidate verbs like "deliver," "send," and "provide." When compared against OntoNotes definitions, these candidates show high similarity to Sense 3 ("to mail or post") and low similarity to Sense 1 ("to convey a feeling").
Conversely, when the system encounters a signature involving abstract nouns (e.g., "expressing horrors"), the PropStore candidates shift toward verbs like "voice" or "convey," correctly steering the system toward the psychological sense of the word. The researchers also noted that treating different Parts-of-Speech (POS) as distinct vectors (e.g., mail.verb vs. mail.noun) significantly improved accuracy, as POS tags are strong indicators of semantic boundaries.
Conclusion and Future Directions
This research demonstrates that syntactic context, when backed by massive distributional data, can serve as a powerful proxy for human annotation in WSD. By treating dictionary definitions as semantic anchors rather than just text, the system achieves robust disambiguation with minimal manual overhead.
Future work aims to expand the complexity of the syntactic signatures beyond simple objects and modifiers. Additionally, the authors plan to integrate this WSD module into full-scale event coreference pipelines and conduct large-scale benchmarking against standard supervised models to further validate the "knowledge-lean" hypothesis.
Summary updated