Abstract
This paper presents an approach to Multilingual Lexical Substitution using Word Sense Disambiguation (WSD) and automatic word alignment. By mapping English WordNet senses to foreign translations via parallel corpora, the authors developed the COLEUR and COLSLM systems. These systems achieved top rankings in SemEval 2010 Tasks 2 and 3, demonstrating that WSD is a robust bridge for cross-lingual tasks and that the choice of training corpus domain significantly impacts performance.
Introduction
In natural language processing, polysemy—where a single word has multiple meanings—poses a significant challenge for translation. For instance, the English word "bank" requires different translations depending on whether it refers to a financial institution or a riverbank. Choosing the correct target word depends entirely on the specific "sense" used in context.
This paper addresses Multilingual Lexical Substitution, specifically for SemEval 2010 Tasks 2 and 3. The goal is to automatically provide a ranked list of valid translations for a target English word within a given sentence. This requires a system to disambiguate the English word's context and map that meaning to the appropriate vocabulary in a second language.
Methodology
The authors proposed a pipeline that bridges monolingual sense disambiguation and multilingual translation through three primary pillars:
- Word Sense Disambiguation (WSD):
- Task 2 (Supervised): A supervised system trained on SemCor was used to create "sense profiles" (vectors of surrounding words) for WordNet senses. Test contexts were compared to these profiles to select the best sense.
- Task 3 (Unsupervised): A graph-based "In-Degree" algorithm was employed. Senses were represented as nodes in a graph with edges indicating semantic similarity (JCN measure). The sense with the highest centrality was selected.
- Automatic Word Alignment: The authors used GIZA++ on parallel corpora to map English words to foreign equivalents. To ensure high precision, they retained only bidirectional alignments (English-to-L2 and L2-to-English).
- Word Sense Translation Table (WSTT): This core innovation combines WSD tags with word alignments to build a lookup table mapping specific English senses to probability distributions of foreign words. This was further expanded using WordNet synonyms to increase translation variety.
Experimental Setup
The study utilized several key datasets to evaluate the systems:
- EuroParl: A standard parallel corpus of European Parliament proceedings.
- SALAAM: An "artificial" parallel corpus created by translating English corpora (Brown, WSJ, SensEval) into other languages using machine translation. This provided a more balanced dataset compared to the parliamentary focus of EuroParl.
- SemCor and WordNet: Used for training the supervised WSD component and as the underlying sense inventory, respectively.
Results and Discussion
The systems, COLEUR and COLSLM, performed competitively in the SemEval 2010 workshop. A key finding was that the artificial SALAAM corpus (T2-COLSLM) outperformed the natural EuroParl corpus (T2-COLEUR) in Task 2. This suggests that for lexical substitution, a broad and balanced domain is more critical than the high-quality human translations found in narrow domains like politics.
In Task 3, the unsupervised approach (T3-COLEUR) achieved the highest ranking among all participants for French, German, and Italian in both the "best" and "out-of-ten" (OOT) categories. For French, the system reached a precision of 21.96% for the best substitute and 49.44% for the OOT metric.
Conclusion
The research reinforces the idea that translation can be effectively modeled as a two-step process: disambiguating the source word's meaning followed by mapping that sense to the target language. While the systems ranked highly, absolute F-measures remained relatively low, highlighting the ongoing difficulty of fine-grained cross-lingual disambiguation. Future work may involve incorporating target-language WordNets or moving toward Word Sense Induction (WSI) to discover senses automatically from data, potentially allowing for more fluid translation models.
Summary updated