Cornetto: A Combinatorial Lexical Semantic Database for Dutch

Piek Vossen, Isa Maks, Roxane Segers, Hennie van der Vliet, Marie‐Francine Moens, Katja Hofmann, Erik Tjong Kim Sang, Maarten de RijkeEssential Speech and Language Technology for Dutch 2013 · DOI · DBLP

Abstract

Cornetto is a large-scale semantic database for the Dutch language that unifies two distinct linguistic resources: a combinatorial lexicon (Referentiebestand Nederlands) and a relational wordnet (Dutch Wordnet). By aligning word-to-meaning and meaning-to-word perspectives and anchoring them to the formal SUMO ontology, Cornetto provides a robust infrastructure for advanced Natural Language Processing (NLP) tasks. This paper describes the architecture, alignment methodology, and evaluation of Cornetto in applications such as news classification and sentiment analysis.

Introduction

In Natural Language Processing (NLP), researchers typically rely on two types of resources: lexicons and wordnets. Lexicons, such as the Referentiebestand Nederlands (RBN), focus on lexical behavior, including syntax, collocations, and idioms. Wordnets, like the Dutch Wordnet (DWN), organize concepts into "synsets" (synonym sets) and map logical relationships like hypernymy. The Cornetto project addresses the gap between these resources by unifying them into a single, cohesive database, enabling systems to understand both how a word is used and what it conceptually represents.

Methodology

The creation of Cornetto involved a sophisticated alignment process across a four-layer architecture:

  1. Lexical Units (LU): Rich morphological and syntactic data derived from RBN.
  2. Synsets: Shared concepts derived from DWN.
  3. Ontology Mappings: Linking synsets to the Standard Upper Merged Ontology (SUMO) for logical anchoring.
  4. Domain Mappings: Categorizing concepts into topical fields (e.g., transport, music).
Cornetto Architecture: Linking Lexical Units and Synsets

Alignment was achieved through a combination of automatic heuristics—such as overlapping definitions and synonym matches—and manual revision by lexicographers for the 10,000 most frequent and polysemous words.

Data Resources

Cornetto integrates several key datasets:

  • RBN: The combinatorial and syntactic backbone.
  • DWN: The source for semantic relations and synset structure.
  • Princeton Wordnet (PWN): A bridge for importing ontological and domain labels.
  • SUMO: The external formal ontology for logical grounding.
  • Wordnet Domains: Used for topical clustering.

Key Insights and Ideas

A central innovation of Cornetto is its dual-perspective architecture. Rather than merging data into a flat list, it maintains linked collections for Lexical Units and Synsets. This allows for querying from either a word-centric or concept-centric starting point.

Polysemy in Cornetto: Different senses of the Dutch word "band"

As shown in the example of the Dutch word "band" (Figure 2), Cornetto captures distinct combinatorial patterns (e.g., "inflate a tire" vs. "start a band") alongside their respective semantic hierarchies and domain labels. This highlights that combinatorics are as vital for machine understanding as isolated definitions.

Key Results

The final database comprises over 92,000 lemmas and 118,000 word meanings. Task-based evaluations demonstrated its utility:

  • News Classification: Using Cornetto-extracted word combinations as features improved the F-measure of a news topic classifier by 2.8% over a standard baseline.
  • Subjectivity Lexicon: Bootstrapping a sentiment analysis tool with Cornetto resulted in a 7% improvement in accuracy compared to using DWN alone, thanks to the richer internal relations.

Conclusion and Future Work

While Cornetto represents a significant advancement, challenges remain. Automated extraction of hypernyms from the web yielded insufficient precision (31%) for inclusion. Furthermore, high-quality alignment remains labor-intensive. Future work will focus on mapping Cornetto to FrameNet to handle semantic roles and developing Word Sense Disambiguation (WSD) systems that leverage its unique structure.

Summary updated

BibTeX
@inproceedings{vossen2013cornetto,
  author = {Piek Vossen and Isa Maks and Roxane Segers and Hennie van der Vliet and Marie‐Francine Moens and Katja Hofmann and Erik Tjong Kim Sang and Maarten de Rijke and Erik F. Tjong Kim Sang},
  title = {{Cornetto: A Combinatorial Lexical Semantic Database for Dutch}},
  booktitle = {Essential Speech and Language Technology for Dutch},
  year = {2013},
  doi = {10.1007/978-3-642-30910-6\_10},
}
Figure