[論文レビュー] A Focus on Neural Machine Translation for African Languages

Laura Martinus, Jade Abbott|arXiv (Cornell University)|Jun 11, 2019

Natural Language Processing Techniques参考文献 19被引用数 37

ひとこと要約

本論文はConvS2SとTransformer NMTアーキテクチャを用いて英語から南部アフリカの公式5言語へ翻訳を行い、データとコードを公開し、アフリカMTの再現性とベンチマークの課題に対処する。

ABSTRACT

African languages are numerous, complex and low-resourced. The datasets required for machine translation are difficult to discover, and existing research is hard to reproduce. Minimal attention has been given to machine translation for African languages so there is scant research regarding the problems that arise when using machine translation techniques. To begin addressing these problems, we trained models to translate English to five of the official South African languages (Afrikaans, isiZulu, Northern Sotho, Setswana, Xitsonga), making use of modern neural machine translation techniques. The results obtained show the promise of using neural machine translation techniques for African languages. By providing reproducible publicly-available data, code and results, this research aims to provide a starting point for other researchers in African machine translation to compare to and build upon.

研究の動機と目的

• Identify key problems hindering MT for African languages (resource scarcity, discoverability, reproducibility, benchmarks).
• Train and evaluate state-of-the-art NMT models (ConvS2S and Transformer) on English-to-five Southern African languages.
• Provide publicly available data, code, and results to establish baselines and benchmarks for future work.

提案手法

• Use public Autshumato parallel corpora aligned at sentence level and clean duplicates to prevent data leakage.
• Train ConvS2S (Word and Best BPE) and Transformer models on each language with default Fairseq and Tensor2Tensor settings respectively.
• Apply beam search during decoding (beam width 5 for ConvS2S, 4 for Transformer).
• Experiment with word-based tokenization and Byte-Pair Encoding (BPE) tokenization, including an ablation study to select optimal BPE token counts per language.
• Evaluate with BLEU scores and perform qualitative analyses including attention visualizations and back-translations.

実験結果

リサーチクエスチョン

• RQ1 What are the achievable BLEU scores for English-to-five South African languages using ConvS2S and Transformer architectures?
• RQ2 Does subword (BPE) tokenization improve translation quality over word-level tokenization for low-resource African languages?
• RQ3 How do data size and language morphology (agglutinative vs non-agglutinative) affect NMT performance in this setting?
• RQ4 Can publicly released data/code establish a reproducible baseline and benchmark for future African MT research?

主な発見

Model	Afrikaans	isiZulu	N. Sotho	Setswana	Xitsonga
ConvS2S (Word)	16.17	0.28	7.41	24.18	36.96
ConvS2S (Best BPE)	25.04 (4k)	1.79 (4k)	12.18 (4k)	26.36 (40k)	37.45 (20k)
Transformer	35.26 (4k)	3.33 (4k)	24.16 (4k)	28.07 (40k)	49.74 (20k)

• Transformer generally outperforms ConvS2S across all languages.
• BPE tokenization consistently yields better performance than word-level tokenization.
• Language performance correlates with dataset size and morphological complexity; isiZulu and Northern Sotho perform worst due to small, low-quality data, while Xitsonga and Setswana perform better with more data.
• Afrikaans (non-agglutinative) achieves reasonable results despite smaller parallel data.
• Maximum BLEU for isiZulu with Transformer is 3.33, indicating severe data quality/size issues.
• Public data/code enables reproducibility and creates a starting benchmark for the five languages.