Abstract
This paper presents an extensible, end-to-end supervised system for Target-Word Sense Disambiguation (WSD). By integrating and modifying three powerful open-source tools—GATE, NSP, and WEKA—the authors created a streamlined pipeline that automates feature extraction and machine learning, overcoming significant architectural limitations in previous NLP frameworks.
Introduction
Word Sense Disambiguation (WSD) is a fundamental task in Natural Language Processing that involves determining the correct sense of a word based on its context. Target-word WSD narrows this focus to a specific set of pre-defined words. While frameworks like GATE (General Architecture for Text Engineering) were available in the mid-2000s, they were not optimized for WSD tasks. The authors identified several critical gaps: restrictive feature extraction limited to immediate context, high workflow friction between text processing and machine learning, and a lack of automation for batch experiments.
System Architecture
The authors developed a unified system that coordinates three major components: GATE for text processing, the Ngram Statistics Package (NSP) for feature identification, and WEKA for classification. The architecture is designed to facilitate both interactive experimentation and automated batch processing.
Methods and Implementation
The core innovation of the system lies in the Modified ML PR (Machine Learning Processing Resource). By rewriting parts of the standard GATE learning component, the authors enabled the extraction of "floating features"—important context clues located at variable distances from the target word.
The system provides two primary interfaces:
- Graphical User Interface (GUI): Integrated within GATE for individual, interactive experiments.
- Command Line Interface (CLI): Utilizes a Perl script (
mkconfig.pl) and Java wrappers to facilitate batch processing and automated "train-test" cycles, allowing researchers to programmatically generate and test hundreds of configurations.
Datasets and Compatibility
While the paper focuses on architectural delivery, the system is designed for broad compatibility with industry standards:
- WordNet: For lexical knowledge integration.
- SenseTools/SyntaLex: Ensuring compatibility with major competition formats like SENSEVAL.
- WEKA ARFF: The standard format for machine learning datasets.
Key Insights
- Solving the Floating Feature Problem: By allowing the system to handle variable-distance features, the authors made the pipeline more linguistically aware than standard tools of the time.
- Extensibility: Built on the GATE infrastructure, the system can easily incorporate new components, such as named entity recognizers or co-reference resolvers, without requiring a redesign.
- Automated Experimentation: The inclusion of
mkconfig.plallows for the systematic exploration of the feature and classifier space, significantly reducing the manual labor involved in model optimization.
Results and Future Work
The primary achievement of this work is a functional, open-source, end-to-end system that bridges the gap between raw text and trained classifiers. The tools, WSDGate and NSPGate, were released on SourceForge to provide the research community with a ready-to-use solution for supervised WSD and related tasks like acronym expansion.
Future improvements aim to address current limitations, including support for non-contiguous n-grams and reducing the system's dependency on specific POS taggers. Additionally, the authors proposed a web-based interface to further lower the barrier to entry for researchers.
Summary updated