Disambiguating and Interpreting Verb Definitions

Yael Ravin

Abstract

This paper presents a Disambiguation Module (DM) designed to interpret dictionary definitions of the form "to VERB with NP." By categorizing the preposition "with" into five distinct semantic relations, the system achieves over 80% coverage in identifying how verbs and objects relate. This process transforms static dictionary text into a structured lexical knowledge base, addressing long-standing challenges in prepositional phrase attachment within Natural Language Processing (NLP).

Introduction

Natural Language Processing has historically struggled with the ambiguity of prepositional phrase attachment. A classic example is "I saw the man with the telescope," where it is unclear if the telescope is an instrument for seeing or an object held by the man. This research focuses on a specific subset of this problem: the interpretation of verb definitions in dictionaries (e.g., angle: "to fish with a hook").

While these definitions appear simple to human readers, they are linguistically dense. The word "with" acts as a "Swiss Army knife" preposition, signaling instruments, manners of acting, physical changes, or purely syntactic links. To build a comprehensive lexical knowledge base, machines must determine the precise semantic role of "with" in context.

Methodology

The authors developed a Disambiguation Module (DM) that operates in three primary stages:

  1. Syntactic Parsing: Utilizing the PEG (PLNLP English Grammar) parser, the system identifies the core components of a definition: the head verb and the noun phrase (NP) head following "with."
  2. Semantic Categorization: The system consults machine-readable versions of Webster’s 7th (W7) and the Longman Dictionary of Contemporary English (LDOCE). It assigns the noun to one of 16 semantic categories, such as +INSTRUMENT or +BODYPART.
  3. Heuristic Mapping: The DM applies rules to map the definition to one of five semantic relations:
  4. USE: The NP is an instrument or substance (e.g., "fish with a hook").
  5. MANNER: The NP describes the mode of action (e.g., "examine with intent").
  6. ALTERATION: The action changes the object's state (e.g., "fill with air").
  7. CO-AGENCY: Two entities participate equally (e.g., "combine with parts").
  8. PROVISION: The agent provides the NP to a recipient (e.g., "fit with clothes").

A "final preference heuristic" resolves conflicts by prioritizing verb-based interpretations (complements) over noun-based ones (adjuncts).

Experimental Setup

The study utilized two primary lexical resources: Webster’s 7th New Collegiate Dictionary for definitions and LDOCE for subcategorization data. The evaluation was conducted using two benchmarks:

  • Training Corpus: 170 verb definitions randomly extracted from the 'a' and 'b' sections of W7.
  • Testing Corpus: 132 previously unseen "to VERB with NP" definitions.
  • Human Baseline: A subset of 58 strings was evaluated by human subjects to validate the module's logic.

Results and Analysis

The DM demonstrated significant robustness. As shown in Figure 1, the system successfully assigned interpretations to 78.2% of the training set and 83.3% of the testing set.

Coverage ratios for training and testing datasets.

Beyond coverage, the system's accuracy was high. In the training set, the DM provided a single, correct interpretation for 85% of attempted cases. On the unseen test set, the system achieved a "Correct-to-Incorrect" ratio of 68%. When including cases where the system provided two interpretations (one of which was correct), the reliability reached 80% (see Figure 2).

Detailed performance metrics including correct (COR), partial (1/2), and incorrect (INC) counts.

Human validation further supported these results, with subjects agreeing with the module's "correct" judgments in 91% of test cases. This suggests that the system's heuristic logic closely mirrors human linguistic intuition.

Conclusion

While the DM shows promise, limitations remain regarding coordinate structures (e.g., "strength and vigor") and rare word senses. For instance, the system may misidentify a sense if a rare dictionary definition (like "dog" as a mechanical device) is prioritized. Future work aims to scale these heuristics across the entire dictionary to populate a fully linked lexical knowledge base (COMPLEX) for advanced NLP applications.

Summary updated

BibTeX
@inproceedings{ravin1990disambiguating,
  author = {Yael Ravin},
  title = {{Disambiguating and Interpreting Verb Definitions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1990},
  doi = {10.3115/981823.981856},
}
Figure