TL;DR
This paper tackles the challenge of representing and interlinking German multiword expressions (MWEs) by porting the OdeNet lexical resource and the MMORPH morphological analyzer into the OntoLex-Lemon framework. This approach allows for a richer, more detailed morphological description of MWEs, bridging the gap between lexical semantics and full linguistic descriptions.
What Problem Does This Paper Address?
Lexical resources like WordNets are invaluable for natural language processing, but they often lack detailed morphological information, especially for complex linguistic phenomena like multiword expressions (MWEs) and compounds. German, with its rich morphology and compounding capabilities, presents a particular challenge. Existing resources like OdeNet (an open German WordNet) and MMORPH (a German morphological analyzer) exist, but they are not easily interoperable or rich enough to capture the full complexity of MWEs. The authors aim to bridge this gap by using OntoLex-Lemon, a semantic web framework for lexicons, to represent and interlink these resources, enabling a more comprehensive understanding of German MWEs.
Methods
The core of the paper's methodology involves mapping two existing German linguistic resources, OdeNet and MMORPH, into the OntoLex-Lemon model.
- OntoLex-Lemon Adoption: The authors leverage OntoLex-Lemon, an extension of the lemon model, which is designed to provide rich linguistic grounding for ontologies. It allows for the representation of lexical entries, their morphological and syntactic properties, and their connection to ontological entities. OntoLex-Lemon's integration with SKOS (Simple Knowledge Organization System) is particularly useful for representing conceptual hierarchies like synsets.
- OdeNet Porting: OdeNet, an open German WordNet, was chosen because it's available under an open-source license, making it suitable for inclusion in the Open Multilingual WordNet. The paper describes the process of converting OdeNet's data into OntoLex-Lemon, including handling its "phrasal entries" and ensuring its inclusion in a SKOS-based conceptual hierarchy.
- MMORPH Integration: An updated version of the MMORPH morphological analyzer was also ported to OntoLex-Lemon. MMORPH provides extensive morphological information, including full-forms and specific details about compounds. A script was developed to transform MMORPH's data into OntoLex-Lemon's RDF (Resource Description Framework) format.
- Interlinking Strategy: The key innovation lies in linking these two resources. The authors demonstrate how OdeNet's MWEs can be morphologically specified using MMORPH's detailed information. This includes representing compounds as
ontolex:MultiWordExpressionand linking them to their constituent parts, as well as capturing specific usage restrictions (e.g., singular-only use) through properties likeontolex:usage.
Datasets
- OdeNet: A German lexical semantics resource, described as an open German WordNet. It combines data from OpenThesaurus German and the Open Multilingual WordNet (OMW), which itself is linked to Princeton WordNet.
- MMORPH: An updated German version of the MMORPH morphological analyzer, which also covers English, Spanish, French, and Italian. The German version contains over 2.6 million full-forms and has improved coverage of compounds.
- External Resources: The authors also consulted online editions of Duden and CanooNet, as well as German Wiktionary for manual checking of lexical features. Wikidata and DBpedia are mentioned as potential sources for adding ontological references.
Key Insights and Ideas
- OntoLex-Lemon as a Bridge: OntoLex-Lemon proves to be a powerful framework for bridging the gap between lexical semantic resources (like WordNets) and detailed morphological analyzers. It allows for a richer representation that goes beyond simple lemmas and part-of-speech tags.
- MWEs as First-Class Citizens: The framework enables the explicit representation of Multiword Expressions (MWEs) and compounds as distinct entities (
ontolex:MultiWordExpression), allowing for the detailed description of their components and their relationships. - Morphological Richness for German: The combination of OdeNet and MMORPH within OntoLex-Lemon allows for the capture of German's complex morphology, including the numerous inflectional variants of adjectives and the structure of compound nouns.
- Capturing Usage Restrictions: OntoLex-Lemon's
ontolex:usageproperty and other mechanisms can be used to encode specific constraints on how MWEs are used (e.g., singular-only), which is often not possible in simpler WordNet structures. - Enabling Linked Data Publication: By using OntoLex-Lemon, the authors pave the way for publishing these enriched lexical resources within the Linguistic Linked Open Data cloud.
Key Results
The paper focuses on the methodology and the potential of the OntoLex-Lemon framework rather than presenting quantitative performance metrics. However, key qualitative results include:
- Successful Porting: Both OdeNet and MMORPH data were successfully ported into OntoLex-Lemon.
- Rich Representation: The resulting OntoLex-Lemon representation of OdeNet includes over 120,000 lexical entries, 36,000 synsets, and a significant portion (approximately half) of entries can be considered MWEs or compounds.
- Detailed Morphological Descriptions: The paper showcases how complex German MWEs, like compounds ("Kernspaltung") and idiomatic phrases ("geistiges Eigentum"), can be represented with their constituent parts and detailed morphological variants derived from MMORPH.
- Demonstration of Usage Restrictions: The authors successfully demonstrate how to represent specific usage restrictions for MWEs, such as the singular-only use of "geistiges Eigentum."
Limitations and Future Work
- Manual Curation: While OdeNet is partly crowd-sourced, it requires curation. The paper notes that automatically generated entries have a confidence score of 0.7, while manually curated ones have a score of 1.0.
- Compound Identification: The current figure for compounds in OdeNet (14,080) is based on a comparison with existing lists. The authors expect a larger number of compounds by applying more sophisticated decomposition algorithms.
- Future Work: The authors mention considering the
lemonUbyresource in a future step. They also suggest that further work could involve adding reference information by querying external knowledge bases like DBpedia or Wikidata. The potential for using OntoLex-Lemon to bridge WordNet-like resources with full lexical descriptions is highlighted as a promising avenue.
Summary updated