Morphological and structural complexity analysis of low-resource English-Turkish language pair using neural machine...

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Translation

Background:

• Neural machine translation (NMT) excels in high-resource languages but is under-explored for morphologically rich, low-resource languages like Turkish.
• Turkish's agglutinative nature and limited data present unique challenges for NMT systems.
• Structural differences between Turkish and English necessitate robust evaluation of NMT components.

Purpose of the Study:

• To compare the performance of Transformer and recurrent-based sequence-to-sequence (Seq2Seq) models for English-Turkish and Turkish-English translation.
• To evaluate the impact of different tokenization strategies (BPE vs. Word Tokenization) and attention mechanisms on translation quality.
• To assess the generalizability of NMT models across different architectural depths and translation directions.

Main Methods:

• Comparative analysis of Transformer and Seq2Seq models using attention mechanisms.
• Experimentation with Byte Pair Encoding (BPE) and Word Tokenization strategies.
• Evaluation using standard metrics: BiLingual Evaluation Understudy (BLEU), Metric for Evaluation of Translation with Explicit ORdering (METEOR), and Translation Error Rate (TER).

Main Results:

• The Transformer model with three layers, eight attention heads, and BPE tokenization achieved superior performance.
• English-to-Turkish translation yielded a BLEU score of 47.85 and METEOR score of 44.62.
• Consistent performance trends were observed in the Turkish-to-English direction, demonstrating model generalizability.

Conclusions:

• Optimized Transformer-based NMT systems demonstrate significant potential for morphologically rich, low-resource languages like Turkish.
• BPE tokenization and specific architectural configurations are crucial for enhancing translation quality in challenging linguistic settings.
• The study provides valuable insights for advancing NMT for low-resource languages with complex morphology.