Parallel Corpora for bi-lingual English-Ethiopian Languages Statistical Machine Translation

Solomon Teferra Abate, Michael Melese, Martha Yifiru Tachbelie, Million Meshesha, Solomon Atinafu, Wondwossen Mulugeta, Yaregal Assabie, Hafte Abera, Binyam Ephrem, Tewodros Abebe, Wondimagegnhue Tsegaye, Amanuel Lemma, Tsegaye Andargie, Seifedin ShifawCOLING 2018

TL;DR

This paper tackles the challenge of building statistical machine translation (SMT) systems for under-resourced Ethiopian languages by creating and releasing parallel corpora. They demonstrate that while English-to-Ethiopian translation is difficult due to morphological complexity and alignment issues, Ethiopian-to-English translation shows promising results, paving the way for future advancements.

What Problem Does This Paper Address?

Natural Language Processing (NLP) applications, including Machine Translation (MT), are increasingly vital for human communication. However, technologically advanced languages have seen significant progress in MT, while under-resourced languages, particularly those in Africa, lag far behind. Ethiopia, with its rich linguistic diversity, faces this challenge acutely. The primary bottleneck for developing MT systems for these languages is the severe lack of parallel corpora – collections of texts and their translations. Existing research is often limited by small, ad-hoc corpora, hindering replicable evaluation and progress. This paper aims to address this by developing and releasing parallel corpora for several Ethiopian languages paired with English, enabling more robust SMT research.

Methods

The authors employed a corpus-based approach to SMT, which is favored for under-resourced languages due to its reduced reliance on deep linguistic analysis. They collected parallel data from the web across religious, historical, and legal domains. For preprocessing, they implemented:

  • Character Normalization: To handle redundant characters in Amharic, reducing vocabulary size and improving consistency.
  • Sentence Tokenization and Alignment: A two-step process involving matching sentence counts and sentence end markers to ensure accurate parallel sentence pairs.
  • Data Preparation: For the English-Ge'ez corpus, they addressed the challenge of multiple verses per line in the source to single verses in the target by merging verses.

For the SMT experiments, they utilized the Moses toolkit for alignment and language modeling, along with the GIZA++ alignment tool. The collected corpora were split into training (80%), tuning (10%), and testing (10%) sets.

Datasets

The paper introduces newly collected parallel corpora for the following language pairs:

  • Amharic-English
  • Tigrigna-English
  • Afan-Oromo-English
  • Wolaytta-English
  • Ge'ez-English

These corpora were gathered from various online sources, including:

  • Religious Domain: Holy Bible, Jehovah's Witnesses materials, Ethiopicbible, Ebible, Ge'ez experience.
  • Historical Domain: "African Almanac" (from admase ethiopia github).
  • Legal Domain: Ethiopian constitution, proclamations, and regulation documents from an Ethiopian legal brief website.

The Amharic-English corpus covers all three domains, while Tigrigna-English and Afan-Oromo-English cover legal and religious domains. Wolaytta-English and Ge'ez-English are primarily from the religious domain.

Key Insights and Ideas

  • Morphological Richness is a Double-Edged Sword: While the morphological complexity of Ethiopian languages is a significant challenge for SMT (especially when they are the target language), it also means that a single word in these languages can carry the meaning of multiple English words, leading to complex alignment issues.
  • Directionality Matters: Ethiopian-to-English translation consistently yields higher BLEU scores than English-to-Ethiopian translation. This is attributed to the "many-to-one" alignment challenge when translating from English to Ethiopian languages, versus the more manageable "one-to-many" alignment in the reverse direction.
  • Domain Impact: The Amharic-English corpus, covering more domains, showed lower BLEU scores than the Tigrigna-English corpus, which covered fewer domains. This suggests that the diversity of domains in the training data can impact translation quality, potentially due to domain-specific vocabulary and phrasing.
  • Data Scarcity Amplified: The significant differences in token and vocabulary counts between English and Ethiopian languages, even for the same number of sentences, highlight how morphological complexity exacerbates data scarcity issues for SMT.

Key Results

The paper reports BLEU scores for bi-directional SMT systems. Key quantitative findings include:

  • Ethiopian-to-English Translation Outperforms:
    • Amharic-English: 22.68 BLEU
    • Tigrigna-English: 27.53 BLEU
    • Afan-Oromo-English: 18.88 BLEU
    • Wolaytta-English: 17.39 BLEU
    • Ge'ez-English: 18.01 BLEU
  • English-to-Ethiopian Translation Challenges:
    • English-Amharic: 13.31 BLEU
    • English-Tigrigna: 17.89 BLEU
    • English-Afan-Oromo: 14.68 BLEU
    • English-Wolaytta: 10.49 BLEU
    • English-Ge'ez: 6.76 BLEU

The Amharic-English translation system showed a vocabulary reduction of 15.78% for the training set after character normalization.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Optimal Alignment: The current alignment strategies may not be optimal for the complex morphological structures of Ethiopian languages. Future work will focus on identifying optimal one-to-many and many-to-one alignment techniques.
  • Understanding English-to-Ethiopian Performance: Further research is needed to pinpoint the exact reasons for the lower performance in English-to-Ethiopian translation systems.
  • Domain Impact Analysis: Investigating the specific effect of different domains on SMT performance is a planned future research direction.
  • Expanding Corpora: While significant, the current corpora could be expanded to include more domains and languages to further improve SMT capabilities.

Summary updated

BibTeX
@inproceedings{abate2018parallel,
  author = {Solomon Teferra Abate and Michael Melese and Martha Yifiru Tachbelie and Million Meshesha and Solomon Atinafu and Wondwossen Mulugeta and Yaregal Assabie and Hafte Abera and Binyam Ephrem and Tewodros Abebe and Wondimagegnhue Tsegaye and Amanuel Lemma and Tsegaye Andargie and Seifedin Shifaw},
  title = {{Parallel Corpora for bi-lingual English-Ethiopian Languages Statistical Machine Translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure