Taxonomy Construction Using Syntactic Contextual Evidence

Anh Tuan Luu, Jung-jae Kim, See-Kiong NgEMNLP 2014 · OpenReview

Abstract

Taxonomies are the backbone of many structured, semantic knowledge resources. Recent works for extracting taxonomic relations from text focused on collecting lexical-syntactic patterns to extract the taxonomic relations by matching the patterns to text. These approaches, however, often show low coverage due to the lack of contextual analysis across sentences. To address this issue, we propose a novel approach that collectively utilizes contextual information of terms in syntactic structures such that if the set of contexts of a term includes most of contexts of another term, a subsumption relation between the two terms is inferred. We apply this method to the task of taxonomy construction from scratch, where we introduce another novel graph-based algorithm for taxonomic structure induction. Our experiment results show that the proposed method is well complementary with previous methods of linguistic pattern matching and significantly improves recall and thus F-measure.

Introduction

Taxonomies—hierarchical structures where concepts are linked by "is-a" relationships—are essential for semantic search, question answering, and knowledge graph construction. Manually building these structures is labor-intensive and difficult to scale. Existing automated methods typically rely on linguistic patterns (e.g., Hearst patterns like "A such as B") or statistical co-occurrence. While linguistic patterns offer high precision, they suffer from low recall because many obvious relationships are never explicitly stated in a single sentence. Statistical methods offer better coverage but often struggle to distinguish specific taxonomic relations from general associations.

This paper addresses the "missing link" problem: identifying taxonomic relations between terms that are clearly related but do not appear together in standard linguistic patterns.

Methodology

The proposed approach follows a three-stage pipeline to construct taxonomies from scratch.

1. Term Extraction and Filtering

Candidate terms are extracted from domain-specific corpora. To ensure relevance, terms are filtered using a combination of TF-IDF, domain relevance, and domain consensus metrics.

2. Taxonomic Relation Identification

The core innovation lies in combining three distinct evidence scores:

  • String Inclusion with WordNet (SIWN): A baseline checking for substring relationships or existing WordNet links.
  • Lexical-Syntactic Patterns (LSP): Validation of traditional patterns using Google search counts.
  • Syntactic Contextual Subsumption (SCS): This novel method posits that a hypernym (general term) appears in a wider variety of syntactic contexts than its hyponyms (specific terms). By comparing Subject-Verb-Object triples (e.g., "Animals eat" vs. "Tigers eat"), the system infers a hierarchy if the context set of one term subsumes another. This allows for evidence collection across different sentences.

3. Taxonomy Induction

Identified relations are treated as a directed graph. To resolve redundancies and loops, the authors apply Edmonds’ algorithm to find the Maximum Optimum Branching, pruning the graph into the most probable tree structure based on confidence scores.

Graph-based taxonomy induction: pruning a noisy graph (a) into a tree structure (b).

Evaluation and Results

The system was evaluated across five domains, including Artificial Intelligence (IJCAI/ACL papers) and Terrorism (State Department reports).

The results demonstrate that the SCS method significantly boosts the number of extracted relations. In the AI domain, the system identified 9.8% more relations than previous state-of-the-art methods. Crucially, SCS identified nearly 30% of relations that were entirely missed by traditional pattern-matching.

Number of extracted relations across different methods and domains.

Despite the increase in recall, the system maintained high precision, ranging between 90% and 92% across domains. On benchmark datasets like "Animals," the system achieved an F-measure of 71%, outperforming existing benchmarks.

Conclusion

By leveraging syntactic contextual evidence, this approach bridges the gap left by traditional pattern-matching methods. While SCS introduces a slight precision trade-off and higher computational requirements due to web-scale evidence gathering, its ability to find "hidden" relations significantly improves the robustness of automated taxonomy construction. Future work includes applying this method to complex fields like biomedicine and integrating it into collaborative human-AI ontology authoring tools.

Summary updated

Abstract

Taxonomies are the backbone of many structured, semantic knowledge resources. Recent works for extracting taxonomic relations from text focused on collecting lexical-syntactic patterns to extract the taxonomic relations by matching the patterns to text. These approaches, however, often show low coverage due to the lack of contextual analysis across sentences. To address this issue, we propose a novel approach that collectively utilizes contextual information of terms in syntactic structures such that if the set of contexts of a term includes most of contexts of another term, a subsumption relation between the two terms is inferred. We apply this method to the task of taxonomy construction from scratch, where we introduce another novel graph-based algorithm for taxonomic structure induction. Our experiment results show that the proposed method is well complementary with previous methods of linguistic pattern matching and significantly improves recall and thus F-measure.
BibTeX
@inproceedings{luu2014taxonomy,
  author = {Anh Tuan Luu and Jung-jae Kim and See-Kiong Ng},
  title = {{Taxonomy Construction Using Syntactic Contextual Evidence}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
  url = {https://openreview.net/forum?id=HJEZoGzOZr},
}
Figure