Abstract
Disambiguating partial cognates—words that function as cognates in some contexts and false friends in others—is a significant challenge for Machine Translation (MT) and Computer-Assisted Language Learning (CALL). This paper proposes a semi-supervised pipeline that utilizes parallel corpora for automatic seed generation and introduces "bilingual bootstrapping." By translating high-confidence English contexts back into French to enrich the training set, the authors achieve robust sense discrimination with minimal manual annotation.
Introduction: The Challenge of Partial Cognates
In cross-linguistic studies, cognates are words with similar forms and meanings (e.g., nature in French and English), while false friends share forms but differ in meaning (e.g., French main meaning "hand"). Partial cognates represent a more complex middle ground: they behave like cognates in specific contexts but as false friends in others.
For instance, the French word client can translate to the English "client" (cognate sense) or "customer/patron" (false friend sense). Similarly, route might mean a "maritime route" or a common "road." Manually labeling these nuances for machine learning is labor-intensive. This research explores whether parallel text and bootstrapping can automate this disambiguation.
Methodology
The researchers developed a multi-stage pipeline to move from supervised to semi-supervised learning:
- Automatic Seed Collection: Parallel corpora (e.g., Hansard, EuroParl) were used to "seed" the model. If a French sentence containing a target word was paired with an English sentence containing its cognate, it was labeled as a cognate sense. If the English side contained a known false friend synonym, it was labeled accordingly.
- Supervised Baseline: Standard classifiers, including Naïve Bayes (NB-K), Decision Trees, and Support Vector Machines (SVM/SMO), were trained on these seeds using a Bag-of-Words (BOW) approach.
- Monolingual Bootstrapping (MB): Classifiers were applied to large unlabeled French datasets (e.g., Le Monde). High-confidence predictions were iteratively added to the training set.
- Bilingual Bootstrapping (BB): This novel approach involved bootstrapping on English data (BNC) and using an online MT tool to translate those contexts back into French, providing the model with a broader range of linguistic features.
Experimental Setup
The study utilized several datasets:
- Seeds: Extracted from the Canadian Hansard, EuroParl, and the BAF corpus.
- Unlabeled Data: French news from Le Monde and the British National Corpus (BNC).
- Evaluation: A multi-domain corpus of 1.5 million words (fiction, magazines, academic texts) was used to test the model's ability to generalize beyond formal parliamentary language.
Key Results
The supervised baseline achieved an average accuracy of 80.17% using the Naïve Bayes (NB-K) classifier. However, the introduction of bilingual bootstrapping provided significant gains.
As shown in the results, the NB-K classifier remained the most robust, reaching an average accuracy of 83.41% after bilingual bootstrapping. The method proved particularly effective at capturing general linguistic features, allowing models trained on parliamentary data to perform well on diverse text types like fiction and news.
Conclusion and Future Work
This research demonstrates that "distantly supervised" seeds from parallel corpora, combined with bilingual bootstrapping, can effectively disambiguate partial cognates. The reliance on MT to bridge languages allows the model to leverage diverse contexts from both the source and target languages.
Future improvements could include lemmatization to group word roots and the incorporation of syntactic features like Part-of-Speech (POS) tagging to better capture context.
Summary updated