Understanding and Analyzing Model Robustness and Knowledge-Transfer in Multilingual Neural Machine Translation using TX-Ray

TL;DR

This paper explores how knowledge transfer, pruning, and interpretability techniques affect multilingual neural machine translation in extremely low-resource scenarios, revealing that transfer learning improves performance while pruning can harm model robustness.

Contribution

It introduces a novel approach using sequential transfer learning and TX-Ray interpretability to analyze knowledge transfer in low-resource MNMT, highlighting limitations of neuron pruning.

Findings

Sequential transfer learning outperforms baselines in low-resource MNMT.

Pruning neurons degrades translation quality and increases catastrophic forgetting.

TX-Ray effectively quantifies knowledge transfer and interpretability in multilingual models.

Abstract

Neural networks have demonstrated significant advancements in Neural Machine Translation (NMT) compared to conventional phrase-based approaches. However, Multilingual Neural Machine Translation (MNMT) in extremely low-resource settings remains underexplored. This research investigates how knowledge transfer across languages can enhance MNMT in such scenarios. Using the Tatoeba translation challenge dataset from Helsinki NLP, we perform English-German, English-French, and English-Spanish translations, leveraging minimal parallel data to establish cross-lingual mappings. Unlike conventional methods relying on extensive pre-training for specific language pairs, we pre-train our model on English-English translations, setting English as the source language for all tasks. The model is fine-tuned on target language pairs using joint multi-task and sequential transfer learning strategies. Our work addresses three key questions: (1) How can knowledge transfer across languages improve MNMT in extremely low-resource scenarios? (2) How does pruning neuron knowledge affect model generalization, robustness, and catastrophic forgetting? (3) How can TX-Ray interpret and quantify knowledge transfer in trained models? Evaluation using BLEU-4 scores demonstrates that sequential transfer learning outperforms baselines on a 40k parallel sentence corpus, showcasing its efficacy. However, pruning neuron knowledge degrades performance, increases catastrophic forgetting, and fails to improve robustness or generalization. Our findings provide valuable insights into the potential and limitations of knowledge transfer and pruning in MNMT for extremely low-resource settings.