Learning Word-Class Lattices for Definition and Hypernym Extraction

Roberto Navigli, Paola VelardiACL 2010

Abstract

Automatically identifying definitions and hypernymic relationships is a critical task for knowledge base construction and question answering. Traditional lexico-syntactic patterns often suffer from low recall due to linguistic diversity or low precision due to noise. This paper introduces Word-Class Lattices (WCLs), a graph-based representation that generalizes over sentence structures by combining specific words with Part-of-Speech (POS) tags. Evaluated on both clean Wikipedia data and noisy web crawls, WCLs demonstrate near-perfect precision and superior generalization capabilities compared to standard pattern-matching techniques.

Introduction

Extracting definitions—sentences that describe the meaning of a term—is a foundational challenge in Natural Language Processing. A typical definition follows a structure like "A pixel is a dot...", where "pixel" is the term being defined (definiendum) and "dot" is its broader category (hypernym or genus). While humans easily recognize these structures, machines struggle with the inherent variability of natural language.

Traditional approaches rely on fixed "search templates" or lexico-syntactic patterns. However, these methods face a trade-off: specific patterns like "X is a Y" are often too noisy, capturing non-definitional statements like "London is a city", while broader patterns fail to capture the diverse ways definitions are phrased in the wild.

Methodology: Word-Class Lattices

The researchers propose Word-Class Lattices (WCLs) as a solution. A WCL is a Directed Acyclic Graph (DAG) where nodes represent "Word Classes"—either specific frequent words (e.g., "is", "a") or POS tags (e.g., NN for nouns, JJ for adjectives).

Example of a Word-Class Lattice (WCL) representing a set of definitional sentences.

The construction process involves three steps:

  1. Star Pattern Generation: Frequent words are retained, while infrequent words are replaced by wildcards (*).
  2. Clustering: Sentences are grouped based on shared star patterns.
  3. Lattice Induction: For each cluster, a lattice is built to represent the common structural "skeleton" of the sentences.

The authors developed two variants: WCL-1, which models the entire sentence as a single lattice, and WCL-3, which decomposes the sentence into three functional fields: the Definiendum (the term), the Definitor (the verb phrase), and the Definiens (the hypernym). This modularity allows WCL-3 to recognize novel definitions by combining learned components in new ways.

Experimental Setup

The models were tested on two datasets:

  • Wikipedia Corpus: A relatively clean dataset of ~4,600 sentences used to test precision in a controlled environment.
  • ukWaC Subset: A noisy, large-scale web crawl containing over 300,000 sentences from diverse domains like Astronomy and Cardiology. This dataset serves as a rigorous test of the system's robustness to real-world linguistic "messiness."

Results and Discussion

The WCL approach demonstrated exceptional performance, particularly in maintaining high precision across different data qualities. On the Wikipedia dataset, WCL-1 achieved a precision of 99.88%, effectively eliminating false positives.

The most significant advantage of WCLs appeared in the noisy ukWaC dataset. While traditional methods like "Star Patterns" and Bigram classifiers saw their precision drop below 50%, WCL-3 maintained a precision of 94.87%.

Performance comparison on the noisy ukWaC dataset, highlighting the precision of WCL models.

Beyond identifying definitions, the system was highly effective at Hypernym Extraction. By identifying the "genus" within the Definiens field, WCL-3 achieved over 96% precision in extracting correct hypernyms (or valid substrings) from the web data, significantly outperforming baseline Hearst patterns.

Conclusion

Word-Class Lattices provide a powerful middle ground between rigid templates and overly flexible statistical models. By capturing the underlying structure of definitions through a combination of lexical and syntactic information, WCLs offer a robust method for large-scale knowledge extraction. Future work may extend this lattice-based approach to other information extraction tasks, such as automated taxonomy generation and relation extraction.

Summary updated

BibTeX
@inproceedings{navigli2010learning,
  author = {Roberto Navigli and Paola Velardi},
  title = {{Learning Word-Class Lattices for Definition and Hypernym Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
}
Figure