Abstract
This paper presents a novel approach to overcoming the data bottleneck in Word Sense Disambiguation (WSD) by automatically generating large-scale, sense-tagged corpora. By leveraging WordNet definitions and the vast resources of the internet via search engines, the authors developed a method to acquire training data that is 30 times larger than existing manually tagged sets, achieving a precision of over 90%.
Introduction
In the late 1990s, Natural Language Processing (NLP) began a significant shift toward statistical models. While these models offered superior performance for tasks like Word Sense Disambiguation, they were constrained by a "data hunger" that manual linguistic labor could not satisfy. Mihalcea and Moldovan (1999) addressed this bottleneck by proposing a method to transform the internet into a live, searchable corpus for linguistic training.
The Challenge of Semantic Tagging
Word Sense Disambiguation—identifying which specific meaning of a word is intended in a given context—is a foundational NLP task. However, creating the "gold standard" datasets required to train these models is prohibitively expensive. For instance, the LEXAS system required approximately one man-year of human effort to tag fewer than 200,000 word occurrences. This scarcity of data limited research to a narrow set of words and hindered the development of robust, general-purpose WSD systems.
Methodology: From WordNet to the Web
The authors' pipeline bypasses manual tagging by using WordNet as a knowledge base and AltaVista as a retrieval tool. The process consists of three primary phases:
- Preprocessing: For a target word sense, the system identifies "monosemous" synonyms (words with only one meaning) and extracts unique identifying phrases from the WordNet "gloss" (definition).
- Search Phase: The system constructs boolean queries to find these synonyms or descriptive phrases in context. By finding a monosemous synonym, the system effectively finds a "clean" context for the ambiguous target word.
- Postprocessing: Sentences are retrieved and verified using a Part-of-Speech (POS) tagger. The search phrase is then replaced with the target word and automatically labeled with the corresponding sense tag.
As shown in the table above, the search queries utilize boolean operators like AND and OR to ensure the retrieved context is highly specific to the intended sense.
Experimental Results
The method was tested on a set of 20 polysemous words encompassing 120 distinct senses. The results demonstrated both high quality and massive scalability.
Key findings include:
- Precision: Manual verification of 1,080 generated examples revealed a precision of 91%. Some words, such as the noun "interest," achieved accuracy as high as 95.7%.
- Volume: The system acquired 80,741 examples for the test set, representing a 30-fold increase over the data available in SemCor, the primary tagged corpus of the era.
- Efficiency: By automating the acquisition process, the researchers were able to generate in hours what would have previously taken years of manual effort.
Discussion and Limitations
While the results were transformative, the authors identified several constraints. The reliance on search engine APIs (like AltaVista) limited the number of hits per query. Furthermore, "rare senses" remained difficult to populate, as they appeared infrequently even on the web. The authors also noted a "distributional bias," where the frequency of senses in the generated corpus reflects online writing habits rather than real-world linguistic frequency.
Despite these challenges, the paper marked a pivotal transition in NLP, proving that the web could serve as a dynamic, self-labeling resource for training complex semantic models.
Summary updated