Accessing GermaNet Data and Computing Semantic Relatedness

Iryna Gurevych, Hendrik NiederlichACL 2005 · OpenReview

Abstract

We present an API developed to access GermaNet, a lexical semantic database for German represented in XML. The API provides a set of software functions for parsing and retrieving information from GermaNet. Then, we present a case study which builds upon the GermaNet API and implements an application for computing semantic relatedness according to five different metrics. The package can, again, serve as a software library to be deployed in natural language processing applications. A graphical user interface allows to interactively experiment with the system.

Introduction

In English Natural Language Processing (NLP), integrating semantic knowledge is relatively straightforward thanks to WordNet and established tools like the WordNet::Similarity package. These tools enable researchers to compute conceptual relatedness, which is essential for tasks such as word sense disambiguation, text summarization, and spelling correction.

However, researchers working with German face significant challenges. While GermaNet serves as the German equivalent to WordNet, its underlying data model differs substantially. GermaNet includes non-lexicalized "artificial" concepts and, most critically, lacks the extensive textual definitions (glosses) found in its English counterpart. Because of these structural divergences, existing WordNet-based implementations cannot be directly applied. This paper addresses this gap by introducing a dedicated Java-based API and a tailored semantic relatedness suite specifically designed for GermaNet.

System Architecture and Methods

The authors developed a two-part system: a foundational API and a semantic relatedness calculator.

The GermaNet API

Built in Java, the API parses the XML version of GermaNet using the Apache Xerces parser. It incorporates a German Porter stemmer to facilitate querying by both exact words and word stems. The API exposes methods at both the Word Sense level (e.g., fetching antonyms) and the Synset level (e.g., fetching hypernyms, holonyms, and entailments).

Semantic Relatedness Software

The system implements five semantic relatedness metrics, inheriting from a custom RelatednessComparator class:

  • Information Content (IC) Metrics: Adaptations of the Resnik, Jiang & Conrath, and Lin metrics. These rely on the GermaNet hierarchy combined with statistical corpus evidence.
  • Lesk-based Metrics: The traditional Lesk algorithm relies on dictionary definitions. To overcome GermaNet's lack of glosses, the authors introduced a novel PseudoGlossGenerator. This tool automatically synthesizes definitions by extracting text from neighboring nodes in the conceptual hierarchy, such as hypernyms or related synsets.

System Workflow

The interaction between the user and the system is designed to handle the specific requirements of the German lexical model through a multi-step process.

Workflow of the GermaNet semantic relatedness system

As illustrated in Figure 1, the workflow begins when a user specifies two German words, their part of speech, and a chosen metric. The system then (1) retrieves word senses corresponding to stems, (2) displays relevant info such as pseudo-glosses or lowest common subsumers, (3) generates possible combinations of word senses, (4) computes the relatedness scores, and (5) returns the best word sense combination and final semantic score.

Datasets

The system was developed and tested using the following resources:

  • GermaNet V4.0 (XML): The core lexical semantic database used to construct the graph of German word senses.
  • German taz corpus: A collection of texts from the newspaper die tageszeitung. This was processed with TreeTagger to compute the Information Content (IC) maps required for the Resnik, Jiang & Conrath, and Lin metrics.

Key Insights and Results

The project demonstrates that direct translation of NLP tools between languages is often insufficient. GermaNet’s lack of textual glosses required a structural workaround—the PseudoGlossGenerator—to enable standard algorithms like Lesk. Interestingly, the authors found that applying morphological stemming to the German corpora made negligible difference when computing information content maps for synsets.

The primary result of this work is a functional Java library and a Graphical User Interface (GUI) that allows users to interactively compute semantic relatedness. While quantitative evaluation against human gold standards is detailed in companion research, this system provides the necessary infrastructure for integrating German semantic knowledge into broader NLP applications, such as automated dialogue summarization and context-aware spelling correction.

Summary updated

Abstract

We present an API developed to access GermaNet, a lexical semantic database for German represented in XML. The API provides a set of software functions for parsing and retrieving information from GermaNet. Then, we present a case study which builds upon the GermaNet API and implements an application for computing semantic relatedness according to five different metrics. The package can, again, serve as a software library to be deployed in natural language processing applications. A graphical user interface allows to interactively experiment with the system.
BibTeX
@inproceedings{gurevych2005accessing,
  author = {Iryna Gurevych and Hendrik Niederlich},
  title = {{Accessing GermaNet Data and Computing Semantic Relatedness}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2005},
  url = {https://openreview.net/forum?id=rkWs-TeO-H},
}

Loading related content...

Figure