[Paper Review] Addressing Cross-Lingual Word Sense Disambiguation on Low-Density Languages: Application to Persian

Navid Rekabsaz, Mihai Lupu+2 more|arXiv (Cornell University)|Nov 16, 2017

Natural Language Processing Techniques12 references3 citations

TL;DR

This paper proposes an unsupervised cross-lingual word sense disambiguation (CL-WSD) method for low-resource languages using word embeddings and semantic similarity between translations and context terms. It outperforms both the standard baseline and the state-of-the-art CO-Graph system on an English-to-Persian CL-WSD benchmark, achieving an F-measure of 0.502 in the OOF setting and 0.188 in the Best setting.

ABSTRACT

We explore the use of unsupervised methods in Cross-Lingual Word Sense Disambiguation (CL-WSD) with the application of English to Persian. Our proposed approach targets the languages with scarce resources (low-density) by exploiting word embedding and semantic similarity of the words in context. We evaluate the approach on a recent evaluation benchmark and compare it with the state-of-the-art unsupervised system (CO-Graph). The results show that our approach outperforms both the standard baseline and the CO-Graph system in both of the task evaluation metrics (Out-Of-Five and Best result).

Motivation & Objective

• Address the challenge of word sense disambiguation in low-density languages like Persian, which lack sufficient parallel corpora and knowledge resources.
• Develop an unsupervised CL-WSD method that relies only on monolingual corpora and a bilingual lexicon to overcome resource scarcity.
• Evaluate the proposed method on a newly established English-to-Persian CL-WSD benchmark to assess performance in low-resource settings.
• Demonstrate the effectiveness of word embedding-based semantic similarity in improving CL-WSD performance without relying on external knowledge or parallel data.

Proposed method

• Use word embeddings in the target language (Persian) to compute semantic similarity between translation candidates and context terms via cosine similarity of vector representations.
• Define a generalized similarity function that computes the maximum cosine similarity between any word in a multi-word Persian translation and a target word in the context.
• Apply the ContextVec algorithm to aggregate the most similar translation vectors from each context term into a single context vector representation.
• Use two aggregation strategies—RelAgg (vector sum with normalization) and RelGreedy (greedy selection of highest-similarity translations)—to compute context-relatedness scores.
• Incorporate translation frequency as a probability weight $ P(t) $ in the similarity computation to prioritize more common translations.
• Leverage a bilingual lexicon and POS-tagged context terms (nouns and verbs) to generate translation sets for ambiguous terms in English, then map them to Persian.

Experimental results

Research questions

• RQ1 Can an unsupervised CL-WSD approach based on word embeddings and semantic similarity outperform existing state-of-the-art systems in low-resource settings like Persian?
• RQ2 How effective is the use of monolingual corpora and a simple bilingual lexicon in replacing the need for large-scale parallel corpora or knowledge bases in CL-WSD?
• RQ3 To what extent do context-based word embeddings and similarity aggregation strategies improve disambiguation accuracy on a new English-to-Persian benchmark?
• RQ4 What are the limitations of context-based, bag-of-words approaches when the semantic topic of the ambiguous term does not align with the context terms?

Key findings

• The proposed RelAgg method achieved an F-measure of 0.502 on the OOF evaluation setting, significantly outperforming the standard baseline (0.418) and CO-Graph (0.441).
• In the more challenging Best evaluation setting, the RelAgg method achieved an F-measure of 0.188, surpassing the standard baseline (0.158) and CO-Graph (0.174).
• The RelAgg and RelGreedy methods showed comparable performance, with RelAgg slightly outperforming in the OOF setting, indicating the robustness of vector aggregation.
• For terms like 'mood' and 'side', all systems—including the proposed method—struggled due to weak semantic alignment between context terms and ambiguous word senses, revealing inherent limitations of bag-of-words context modeling.
• Despite these limitations, the method demonstrates strong performance in low-resource settings, proving effective where knowledge-based or supervised approaches fail due to lack of resources.
• The results confirm that word embedding-based semantic similarity is a viable and effective alternative for CL-WSD in low-density languages when parallel or knowledge resources are unavailable.