Towards Effective Correction Methods Using WordNet Meronymy Relations

Javier Álvez, Itziar Gonzalez-Dios, German RigauGWC 2023 · OpenReview

Abstract

For AI systems to perform effective commonsense reasoning, they must bridge the gap between natural language resources and formal logic ontologies. This paper describes a semi-automated pipeline designed to correct misalignments between WordNet and the Suggested Upper Merged Ontology (SUMO). By addressing both semantic mapping errors and overly restrictive ontological axioms, the researchers improved the validation rate of "member-of" relations from a baseline of 0.15% to nearly 82%, significantly enhancing the utility of these knowledge bases for automated theorem proving.

Introduction

Commonsense reasoning—the ability for an AI to understand that a "parent" is a member of a "family"—often relies on linking human-readable dictionaries like WordNet to formal, machine-readable ontologies like SUMO. However, these resources are frequently misaligned. A WordNet synset might be mapped to an incorrect SUMO concept, or the two systems may have fundamental logical disagreements. For instance, WordNet might treat a biological genus as a group, while SUMO treats it as an individual. Such discrepancies create a "knowledge gap" where automated theorem provers (ATPs) cannot logically verify facts, even when the data is present.

Methodology

The researchers focused specifically on the meronymy (part-whole) relation "member." They proposed a three-pronged strategy to resolve misalignments:

  1. Mapping Correction: The team manually inspected "Basic Level Concepts" (BLCs)—the most frequent and salient terms—and corrected their links to SUMO. They then applied four heuristics to automatically propagate these corrections to thousands of related sub-terms (hyponyms).
  2. Ontology Correction: The study identified that SUMO’s internal logic was often too rigid. For example, SUMO’s member predicate required the "part" to be a physical object, which conflicted with biological group definitions. The researchers proposed relaxing these domain restrictions and adding new axioms to better represent plant and animal taxonomies.
  3. Joint Intervention: This approach combined both mapping and ontology fixes to maximize reasoning performance.
Figure 1: Examples of valid and invalid mappings between WordNet and SUMO concepts, highlighting the complexity of biological taxonomies.

As shown in Figure 1, the "Invalid" example illustrates a common friction point: WordNet and SUMO often clash on whether a species or genus should be treated as an abstract class or a concrete group, leading to logical redundancies where both the part and the whole are subsumed by the same concept (e.g., Canine).

Experimental Evaluation

The methods were tested using Adimen-SUMO (a first-order logic version of SUMO) and state-of-the-art theorem provers like Vampire and E. The primary dataset was WordNet v3.0, specifically its 12,293 "member" meronymy pairs. To test generalizability, the researchers also used WebChild, a large-scale collection of commonsense knowledge extracted from the web.

Key Results

The improvement across the different phases of intervention was dramatic. While mapping or ontology fixes alone provided modest gains, the joint intervention yielded a massive leap in system performance.

Figure 2: Performance metrics across different intervention phases for WordNet and WebChild datasets.

As detailed in Figure 2, the initial validation rate for WordNet was a dismal 19 pairs (0.15%). The joint intervention increased this to 10,071 validated pairs (~82%). When applied to the unseen WebChild dataset, the method achieved a Precision of 0.84, demonstrating that the corrections were not merely overfitting to WordNet but were improving the system's general reasoning capabilities.

Discussion and Future Work

A major finding of this research is the "Rigidity Trap": the ontology itself is often the bottleneck. Even with perfect mappings, theorem provers fail if the ontology's internal rules are too strict for natural language nuances. Furthermore, the success of the BLC-based heuristics proves that strategic human intervention at the top of a hierarchy is far more efficient than manual correction of every individual node.

Despite these gains, limitations remain. The current method does not yet handle metonymy (e.g., using "the White House" to refer to the US Government) or certain misalignments in human vs. animal classifications. Future work will explore mapping based on deep logical formulae rather than text labels and extending this logic to other meronymy relations like "part-of" (physical components) and "substance-of."

Summary updated

BibTeX
@inproceedings{alvez2023towards,
  author = {Javier Álvez and Itziar Gonzalez-Dios and German Rigau},
  title = {{Towards Effective Correction Methods Using WordNet Meronymy Relations}},
  booktitle = {GWC},
  year = {2023},
  url = {https://openreview.net/forum?id=AU0R36mPDY2},
}
Figure