[Paper Review] Utilizing Lexical Similarity between Related, Low-resource Languages for Pivot-based SMT

Anoop Kunchukuttan, Maulik Shah+2 more|arXiv (Cornell University)|Feb 23, 2017

Natural Language Processing Techniques22 references3 citations

TL;DR

This paper proposes a subword-level pivot-based statistical machine translation (SMT) system using related low-resource languages as pivots, leveraging orthographic syllables (OS) and byte pair encoding (BPE) to improve translation quality. It demonstrates that subword-level pivoting outperforms word- and morpheme-level pivoting and rivals direct translation models, especially when combining multiple pivot languages, effectively compensating for the absence of a direct parallel corpus.

ABSTRACT

We investigate pivot-based translation between related languages in a low resource, phrase-based SMT setting. We show that a subword-level pivot-based SMT model using a related pivot language is substantially better than word and morpheme-level pivot models. It is also highly competitive with the best direct translation model, which is encouraging as no direct source-target training corpus is used. We also show that combining multiple related language pivot models can rival a direct translation model. Thus, the use of subwords as translation units coupled with multiple related pivot languages can compensate for the lack of a direct parallel corpus.

Motivation & Objective

• Address the challenge of low-resource machine translation between related languages lacking direct parallel corpora.
• Investigate whether lexical similarity among related languages can be exploited to improve pivot-based translation quality.
• Explore subword units (OS and BPE) as translation units to reduce data sparsity and improve generalization in low-resource settings.
• Evaluate the effectiveness of combining multiple pivot languages to approach or match direct translation system performance.
• Assess robustness of subword-level pivot models under cross-domain translation scenarios.

Proposed method

• Train phrase-based SMT models on subword-level units (orthographic syllables and BPE) between source-pivot and pivot-target language pairs.
• Use monotonic decoding to preserve word order similarity in related languages and employ high-order (10-gram) language models to reduce data sparsity.
• Apply word-level tuning during decoding to optimize BLEU scores and perform desegmentation by reassembling subwords at test time.
• Select a pivot language related to both source and target to enhance lexical similarity and translation consistency.
• Combine multiple pivot models via linear interpolation with equal weights to aggregate translation knowledge from diverse language pairs.
• Use equal-weight interpolation between the direct model and all pivot models to further boost performance.

Experimental results

Research questions

• RQ1 Can subword-level pivot-based SMT outperform word- and morpheme-level pivot models in low-resource settings?
• RQ2 To what extent can a subword-level pivot model rival a direct translation model when no parallel corpus exists between source and target languages?
• RQ3 Does combining multiple pivot models using different related languages improve translation performance and approach direct model quality?
• RQ4 How robust are subword-level pivot models to domain shift compared to morpheme-level or word-level pivot systems?
• RQ5 Can lexical similarity between related languages be effectively leveraged through subword units to enhance translation quality?

Key findings

• Subword-level pivot models (OS and BPE) outperform word-level pivot models by 5–10 BLEU points and achieve ~90% of the BLEU score of the best direct morpheme-level model.
• The subword-level pivot model achieves 95% of the morpheme-level direct model’s BLEU score, demonstrating strong competitiveness without any direct parallel data.
• Combining multiple pivot models using different related pivot languages yields a system (23.69 BLEU for mar-ben) that rivals the direct model (23.53 BLEU), with statistical significance (p < 0.05).
• The combination of all pivot models and the direct model via equal-weight interpolation further improves performance (24.41 BLEU), exceeding the direct model in some cases.
• Subword-level pivot models show significantly less performance drop in cross-domain translation (agriculture domain) compared to morpheme-level models, indicating better robustness.
• The use of subwords mitigates data sparsity and improves feature estimation, enabling multiple pivot languages to enhance translation quality despite limited parallel data.