Hybrid Statistical Machine Translation for English-Myanmar: UTYCC Submission to WAT-2021

Ye Kyaw Thu, Thazin Myint Oo, Hlaing Myat Nwe, Khaing Zar Mon, Nang Aeindray Kyaw, Naing Linn Phyo, Nann Hwan Khun, Hnin Aye ThantACL 2021

TL;DR

This paper introduces a hybrid statistical machine translation (SMT) system for English-to-Myanmar that leverages manually extracted transliteration knowledge. By integrating this knowledge into an SMT decoder using XML markup, the system significantly improved translation quality, achieving the second-highest score in the WAT-2021 shared task.

What Problem Does This Paper Address?

Machine translation (MT) for low-resource languages, particularly those with complex reordering requirements like English-Myanmar, remains a significant challenge. Existing parallel corpora for this language pair are relatively small, making it difficult to train robust MT systems. Furthermore, the presence of transliterated words and phrases, where English words are directly used or adapted into Myanmar script, poses a specific hurdle for traditional SMT models. This paper aims to address these challenges by proposing a novel hybrid approach that explicitly incorporates transliteration knowledge into an SMT framework.

Methods

The core of this work is a hybrid SMT system that augments standard SMT techniques with transliteration information. The authors manually extracted English-Myanmar transliteration word and phrase pairs from the ALT corpus. This extracted knowledge was then integrated into the Moses SMT toolkit using an XML markup scheme. Specifically, they employed the --xml-input flag with two options: exclusive and inclusive. This allows the SMT decoder to treat these marked segments differently, effectively guiding the translation process for transliterated content. The paper explored this hybrid approach with three SMT baselines: Phrase-Based SMT (PBSMT), Operation Sequence Model (OSM), and Hierarchical Phrase-Based SMT (HPBSMT).

Datasets

The primary datasets used in this research are:

  • ALT Corpus: A publicly available English-Myanmar parallel corpus, extracted from English Wikinews. This corpus was used for both training and manual extraction of transliteration pairs.
  • UCSY Corpus: Another publicly available English-Myanmar parallel corpus, used for training.

The combined UCSY and ALT corpora formed the training data for the WAT-2021 shared task. The ALT test set was used for evaluation.

Key Insights and Ideas

  • Transliteration is Key for Low-Resource Languages: The paper strongly suggests that explicitly handling transliterated words and phrases is crucial for improving MT quality in low-resource scenarios, especially when direct translation is difficult or unnatural.
  • Hybrid SMT is Effective and Implementable: The proposed hybrid approach, which combines existing SMT frameworks with manually curated transliteration knowledge, proved to be a practical and effective method for boosting translation performance without requiring entirely new model architectures.
  • XML Markup as a Flexible Integration Tool: The use of XML markup within the Moses decoder provided a flexible mechanism to inject external knowledge (transliteration pairs) into the translation process, demonstrating its utility beyond simple tag insertion.
  • -xml-input exclusive Outperforms -xml-input inclusive: The study found that the exclusive option for XML markup generally yielded slightly better results, suggesting that a more direct application of the transliteration rules might be more beneficial.

Key Results

The hybrid SMT system demonstrated significant improvements across all three SMT approaches.

  • Overall Improvement: The hybrid approach achieved improvements of around 6 BLEU scores compared to the baseline SMT systems.
  • WAT-2021 Performance: The UTYCC submission, using the hybrid PBSMT with XML markup, achieved the second-highest score among all participants in the English-to-Myanmar translation task at WAT-2021, according to BLEU and AMFM metrics.
  • Impact of ALT Corpus: The baseline translation performance showed a difference of approximately 5.0 BLEU scores when training with or without the ALT corpus, highlighting the value of this dataset.
  • HPBSMT Performance: The Hierarchical Phrase-Based SMT (HPBSMT) system, particularly when trained without the ALT corpus, achieved the highest scores among the hybrid systems.

Limitations and Future Work

  • Manual Extraction Effort: The primary limitation acknowledged is the manual effort required for extracting transliteration word/phrase pairs. This process is time-consuming and labor-intensive.
  • Future Work: The authors suggest that further exploration is needed for low-resource distant language pairs. They also imply that automating or semi-automating the transliteration knowledge extraction process could be a valuable direction for future research. The paper also notes that while they explored various XML markup options, they focused on presenting the two that worked best for English-Myanmar hybrid translation, leaving other options for future investigation.

Summary updated

BibTeX
@inproceedings{thu2021hybrid,
  author = {Ye Kyaw Thu and Thazin Myint Oo and Hlaing Myat Nwe and Khaing Zar Mon and Nang Aeindray Kyaw and Naing Linn Phyo and Nann Hwan Khun and Hnin Aye Thant},
  title = {{Hybrid Statistical Machine Translation for English-Myanmar: UTYCC Submission to WAT-2021}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}

Loading related content...

Figure