Introduction
Achieving true Natural Language Understanding (NLU) requires mapping raw syntax to deep semantic representations. Traditionally, this has been divided into two isolated tasks: Word Sense Disambiguation (WSD), which identifies the specific meaning of a word, and Semantic Role Labeling (SRL), which identifies the relationships between a verb and its arguments. This paper argues that such separation is counterproductive, as word senses and semantic roles are deeply interdependent. We present a unified, knowledge-based approach that performs both tasks concurrently by matching candidate interpretations against a formal ontology.
Methodology
Our approach leverages a single process to generate and evaluate candidate sentence interpretations using the Component Library (CLib), a domain-independent upper ontology containing approximately 500 generic concepts and 80 semantic relations.
Candidate Generation and Semantic Matching
For any given sentence, the system identifies all possible CLib concepts (senses) and relations (roles) for every word and syntactic link. It then generates all valid combinations of these possibilities. The core engine, the Semantic Matcher, evaluates these candidates by comparing them against the background knowledge in CLib. Scoring is based on taxonomic distance, measuring the hierarchical "steps" between concepts in the ontology.
Transformation Rules
A critical challenge in semantic mapping is the "mismatch" between natural language and formal logic. To bridge this gap, we employ a library of approximately 200 transformation rules. For instance, if a text states "The man hit the table," but the ontology specifies that a "hand" is the instrument of hitting, a part descension rule allows the system to recognize that an action involving a part (the hand) relates to the whole (the man).
Experimental Setup
The system was evaluated on a corpus of 196 sentences collected from human subjects across four domains: Chemistry, Pollution Prevention, Employee Safety, and Nuclear Deterrence. To isolate the semantic matching performance from parsing complexities, sentences were written in "controlled English" (simplified syntax without pronouns). Two human graders provided the gold-standard annotations for comparison.
Results and Discussion
The unified approach was compared against a most-frequent-sense baseline and the Lesk algorithm. As shown in the table below, our approach significantly outperformed both benchmarks in both precision and recall.
Key Findings
- Synergistic Disambiguation: By performing WSD and SRL simultaneously, the system uses the constraints of one to resolve the other. Interpretations that are logically inconsistent with known semantic roles are naturally penalized.
- The Role of Transformations: Ablation studies (the "NoXForm" variant) revealed that disabling transformation rules led to a nearly 20% drop in SRL recall. This confirms that bridging the gap between "what is said" and "what is known" is essential for robust NLU.
- Knowledge-Driven Robustness: Unlike data-hungry deep learning models, this approach relies on the quality of the underlying ontology, making it highly explainable and effective in niche domains where training data is scarce.
Conclusion and Future Work
While the current system demonstrates high accuracy in controlled environments, future work will focus on scaling to unrestricted natural language. This will involve integrating sophisticated parsers to handle complex syntax and pronouns. Additionally, we aim to improve computational efficiency to mitigate the combinatorial explosion of candidate interpretations and enhance robustness against incomplete or biased ontological data.
Summary updated