TL;DR
This paper critically examines the WordNet verb hierarchy, revealing significant topological anomalies and questioning the validity of its semantic categories. By proposing methods to correct these issues and suggesting a new top ontology, the research aims to improve the performance of NLP applications that rely on WordNet's verb relations.
What Problem Does This Paper Address?
WordNet is a cornerstone lexical database for Natural Language Processing (NLP), particularly for tasks involving semantic distance. However, applications that leverage WordNet's verb hierarchy, specifically its hypernymy/troponymy relations, often suffer from suboptimal performance. This paper investigates the correctness of these verb relations, aiming to identify and rectify anomalies that hinder the accuracy and efficiency of NLP systems. The core challenge lies in the potential for these hierarchical relationships, especially when dealing with complex phenomena like multiple inheritance, to be inconsistently encoded, leading to "topological anomalies" and questionable semantic categorizations.
Methods
The authors employ a computational approach to analyze WordNet 3.0. They implement a model where synsets, word senses, and relations are represented as Java objects. Their methodology focuses on three main aspects: topological anomalies, anomalies related to semantic categories, and the top ontology.
Key techniques include:
- Algorithmic Anomaly Detection: An algorithm is developed to recursively construct "upside-down trees" from each synset to identify cycles and rings (instances where a synset appears more than once in its own ancestry). This approach is an extension of previous work, specifically designed to detect rings that span across different semantic categories.
- Isolator Identification: Isolators, synsets whose only hypernym resides in a different semantic category, are identified and analyzed to assess the validity of WordNet's semantic categories.
- Candidate Unique Beginner Identification: Synsets lacking a hypernym are identified as candidate unique beginners, prompting a review of the top ontology.
- Manual Annotation and Proposed Ontology: For candidate unique beginners, a new top ontology is proposed through manual annotation, defining hypernyms or verbal phrases. This involves identifying necessary auxiliary verbs and modal verbs, leading to a more compact and consistent top ontology.
Datasets
The primary dataset used is WordNet 3.0. The study also references findings from previous analyses of WordNet 2.0 (Liu et al., 2004; Smrž, 2004). No new datasets were introduced; rather, the focus was on analyzing and improving the existing WordNet structure.
Key Insights and Ideas
- Topological Anomalies Abound: The WordNet verb hierarchy is not a perfect tree. The study found a significant number of "rings" (where a synset has multiple hypernyms with a common ancestor) and "isolators" (synsets with hypernyms in different categories).
- Semantic Categories are Problematic: The current semantic categories in WordNet are not mutually exclusive and lack a consistent theoretical basis, leading to misclassifications and anomalies. The paper suggests that verbs might need to belong to multiple semantic categories.
- Dual Inheritance Needs Justification: While multiple inheritance (dual inheritance) is present, many instances are not semantically justified. Only cases where distinct but compatible properties are inherited are considered legitimate.
- A New Top Ontology is Possible: By analyzing candidate unique beginners and proposing a new top ontology based on auxiliary verbs and modal verbs, the authors demonstrate the possibility of creating a more compact and consistent hierarchy, reducing the number of unique beginners from 254 to 18.
- Encoding Errors are Prevalent: Many hypernymy relations appear to be encoded arbitrarily, with little respect for the distinctions between different types of relations (e.g., causation vs. entailment).
Key Results
- Cycle: A single cycle was identified in WordNet 3.0 involving the synsets "restrain, keep, keep back, hold back" and "inhibit, bottle up, suppress." Merging these synsets is proposed.
- Rings: Out of 8 rings in the verb hierarchy, 4 were asymmetric and 4 symmetric. In 7 out of 8 cases, rings could be corrected, with one case ("turn") deemed justified.
- Isolators: 1593 verb isolators were found, with a sample analysis showing 63% of semantic relations among isolators were considered correct.
- Dual Inheritance: A review of dual inheritance without rings identified 46 verb hypernyms, with only 34.78% being considered "correct."
- Candidate Unique Beginners: 254 verbs were identified as candidate unique beginners (no hypernym), prompting the development of a new top ontology.
- Proposed Top Ontology: A new top ontology with 18 unique beginners was proposed, incorporating modal verbs and auxiliaries.
Limitations and Future Work
- Manual Correction Overhead: The authors acknowledge that applying manual corrections to the entire verb hierarchy or constructing a new lexical database would be a substantial undertaking.
- Borrowing Relational Structure: An alternative suggested for improvement is to borrow the relational structure from another lexical database.
- Formalizing Definitions: The proposed definitions for the new top ontology are tentative and could benefit from formalization using frameworks like Jackendoff's lexical conceptual structures.
- Further Investigation of Dual Inheritance: While some cases of justified dual inheritance were identified, the authors suggest that many more could be discovered, potentially leading to further justifiable rings.
- Category Overlap: The paper strongly advocates for allowing verbs to belong to multiple semantic categories to better reflect real-world semantics.
Summary updated