Introduction
Building comprehensive lexical resources like WordNet is a resource-intensive endeavor. To mitigate costs, researchers often attempt to align existing resources—a process known as Word Sense Alignment (WSA). This paper, presented at EMNLP 2014, addresses the challenge of aligning the Senso Comune De Mauro (SCDM) Italian lexicon with MultiWordNet (MWN). While MWN is highly structured and labeled with subject domains (e.g., "Medicine"), SCDM lacks such metadata, leading to alignment errors where words with the same spelling but different domains (e.g., "bank" in finance vs. "bank" in geography) are incorrectly paired.
Methodology
The authors proposed a two-stage hybrid approach to automatically assign domain labels to SCDM entries, leveraging the "One Domain per Discourse" (ODD) hypothesis. This hypothesis suggests that within a specific sense description, words will generally belong to the same domain.
1. Binary Classification
The first stage involves a machine learning classifier to distinguish between "FACTOTUM" senses (generic words like thing) and "OTHER" senses (domain-specific words). Evaluation of various classifiers showed that Naive Bayes utilizing Word Sense Disambiguation (WSD) features performed most effectively.
2. Rule-Based Domain Assignment
For senses identified as "OTHER," a rule-based system analyzes the definition (gloss). By applying WSD to the words within the gloss and mapping them to their corresponding domains in MultiWordNet, the system assigns the most frequent domain label to the SCDM entry.
Evaluation and Results
The effectiveness of the domain assignment was evaluated against a manually aligned dataset of 166 sense pairs. The hybrid system (NaiveBayes_wsd+Rules) achieved a Precision of 0.70 and an F1 score of 0.58, outperforming baselines that relied solely on lemmas or WSD without the hybrid rules.
The ultimate goal was to use these assigned labels as a "sanity check" filter for alignment. If an alignment algorithm suggested a pair with mismatched domains, the pair was discarded. This filtering mechanism significantly enhanced alignment quality.
As shown in Figure 4, the domain filter consistently improved precision across different configurations. For the LexicalMatch+Cosine > 0.2 setup, precision increased from 0.67 to 0.77. In stricter settings, precision reached as high as 0.87. While this approach involves a trade-off with recall (as some valid alignments with broader domain labels may be filtered out), the substantial gain in precision makes the resulting lexical resource far more reliable for downstream NLP applications.
Conclusion
This research demonstrates that automatic domain assignment, even when derived from short dictionary definitions, is a powerful tool for refining word sense alignments. The study highlights that performing WSD on definitions acts as a "force multiplier" for accuracy. Future work aims to address the classifier's tendency to over-assign the "FACTOTUM" label and to explore the use of domain clusters to allow for more flexible, semantically-aware matching.
Summary updated