Do Dogs have Whiskers? A New Knowledge Base of hasPart Relations

Sumithra Bhakthavatsalam, Kyle Richardson, Niket Tandon, Peter ClarkarXiv 2020 · arXiv

Abstract

Understanding part-whole (meronymic) relationships is a cornerstone of common-sense reasoning. However, existing knowledge bases (KBs) often struggle to balance scale, precision, and salience. This paper introduces hasPartKB, a high-quality resource containing over 50,000 part-whole relationships. By leveraging generic sentences and neural extraction, hasPartKB achieves 90% precision and significantly outperforms existing KBs in covering common, salient terms.

Introduction

In artificial intelligence, reasoning about physical objects requires understanding their components—for instance, knowing that if a car moves, its wheels must also move. This domain of meronymy (the hasPart relation) is fundamental but poorly represented in current AI resources. Existing KBs typically fall into three categories: small but accurate ontologies (e.g., WordNet), massive but noisy automated KBs (e.g., PWKB), or highly specialized domain-specific resources.

The challenge lies in creating a KB that is simultaneously accurate, broad in vocabulary, and salient (containing facts humans actually find relevant). To address this, we present hasPartKB, which bridges the gap between small-scale manual curation and large-scale automated extraction.

Methodology

The construction of hasPartKB follows a five-step pipeline designed to maximize both precision and salience:

  1. Generic Sentence Selection: The process begins with GenericsKB, a collection of 3.4 million generic statements (e.g., "Birds have wings"). Generics are naturally biased toward essential, category-level truths, providing a pre-filtered source of salient information.
  2. Neural Extraction: A RoBERTa-large model was fine-tuned on a small hand-annotated dataset to identify hasPart relationships within these sentences.
  3. Entity Normalization: Modifiers are stripped to maintain a clean KB (e.g., "large elephant" becomes "elephant"), with the original modifiers preserved as metadata.
  4. Thresholding: To ensure high quality, a strict confidence score cutoff of 0.9985 was applied, targeting a precision of 90%.
  5. Entity Linking: Entries are disambiguated and linked to Wikipedia and WordNet senses using a customized GlossBERT model.

Evaluation and Results

We evaluated hasPartKB against several established benchmarks, including WordNet, ConceptNet, PWKB, and WebChild.

Coverage of Common Vocabulary

A key metric for a general-purpose KB is its coverage of "common" terms. We used a "Fifth-Grade Vocabulary" benchmark to assess how well each KB captures relationships between everyday objects. As shown in Figure 3, hasPartKB contains 15,200 relations within this vocabulary—significantly more than WordNet or ConceptNet.

Figure 3: Yield comparison across different knowledge bases, highlighting hasPartKB's superior coverage of common (5th-grade) vocabulary.

Salience and Precision

While massive KBs like PWKB contain millions of entries, they often include low-salience facts that humans rarely mention (e.g., "a nucleolar organiser is part of a poet laureate"). We measured salience using a web-search co-mention test. hasPartKB achieved a salience score of 79.2%, significantly higher than the massive PWKB (12.1%) and WebChild (18.4%), as detailed in Figure 6.

Figure 6: Salience comparison showing that hasPartKB contains information much more relevant to human communication than larger, automated KBs.

Furthermore, hasPartKB maintains a precision of 90%, rivaling manually curated resources while offering much greater scale.

Discussion and Limitations

The success of hasPartKB stems from the use of generic sentences, which act as a natural filter for salience. However, the model still faces challenges:

  • Ambiguity: Prepositions like "of" do not always denote part-whole relationships.
  • Metonymy: The model occasionally misses intermediate concepts (e.g., jumping from "monkey" to "fingers" and missing "hand").
  • Over-generalization: Specific facts (e.g., about aquatic species) are sometimes incorrectly generalized to broader categories.

Conclusion

hasPartKB provides a robust, large-scale resource for meronymic reasoning. By focusing on salient, common-sense relationships, it offers a significant improvement over existing KBs for downstream AI tasks. The dataset is publicly available for the research community at https://allenai.org/data/haspartkb.

Summary updated

Abstract

We present a new knowledge-base of hasPart relationships, extracted from a large corpus of generic statements. Complementary to other resources available, it is the first which is all three of: accurate (90% precision), salient (covers relationships a person may mention), and has high coverage of common terms (approximated as within a 10 year old's vocabulary), as well as having several times more hasPart entries than in the popular ontologies ConceptNet and WordNet. In addition, it contains information about quantifiers, argument modifiers, and links the entities to appropriate concepts in Wikipedia and WordNet. The knowledge base is available at https://allenai.org/data/haspartkb
BibTeX
@misc{bhakthavatsalam2020do,
  author = {Sumithra Bhakthavatsalam and Kyle Richardson and Niket Tandon and Peter Clark},
  title = {{Do Dogs have Whiskers? A New Knowledge Base of hasPart Relations}},
  year = {2020},
  eprint = {2006.07510},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.07510},
}
Figure