Towards Opening the Black Box of Neural Machine Translation: Source and Target Interpretations of the Transformer

Abstract: In Neural Machine Translation (NMT), each token prediction is conditioned on the source sentence and the target prefix (what has been previously translated at a decoding step). However, previous work on interpretability in NMT has mainly focused solely on source sentence tokens' attributions. Therefore, we lack a full understanding of the influences of every input token (source sentence and target prefix) in the model predictions. In this work, we propose an interpretability method that tracks input tokens' attributions for both contexts. Our method, which can be extended to any encoder-decoder Transformer-based model, allows us to better comprehend the inner workings of current NMT models. We apply the proposed method to both bilingual and multilingual Transformers and present insights into their behaviour.

• The paper introduces ALTI+ to quantify the contribution of both source tokens and target prefixes in transformer-based NMT, enhancing model interpretability.
• It demonstrates that up to 47% of token identities are preserved in multilayered contexts, underscoring the encoder's retention of input-specific information.
• The analysis of residual connections and language resource variations provides insights into safer model debugging and effective bias detection in machine translation.

Exploring the Inner Workings of Neural Machine Translation Models with ALTI+

Understanding Enhancements for Interpretability

Interpretability in the field of Neural Machine Translation (NMT) has evolved significantly with the rise of transformer-based models. Traditional methods have focused heavily on source tokens within the sequence but not as much on the target prefixes — the part of the translation that has already occurred. This oversight leaves some aspects of the model's decision-making process less understood. In response, the paper introduces an extended approach named ALTI+, an adaptation from previous interpretability methodologies designed specifically to handle the intricacies of the encoder-decoder dynamics in transformer models.

Key Contributions of ALTI+

• Methodological Expansion: ALTI+ quantifies how each input token (both from the source sentence and the target prefix) contributes to model predictions, offering a nuanced view of how information is processed throughout a translation task.
• Contextual Information Analysis: The encoder in NMT models preserves a significant portion of the token’s identity, up to 47% even in multilayered contexts, showing a deliberate retention of input-specific information.
• Role of Residual Connections: Insight is provided into residual connections, especially in how they allow for the flow of less informative source tokens like EOS (end of sentence) to manage translation fluidity.
• Behavioral Insight across Resources: By studying models trained with languages of varying resource levels, unique behaviors are unearthed, especially under conditions like hallucinations where models generate coherent but source-unrelated text.

Theoretical and Practical Implications

ALTI+ is not just an advancement in understanding how machine translation works; it's a tool that could refine how models are built. By identifying how different tokens influence translation, developers can fine-tune their models more effectively, potentially enhancing performance and accuracy.

• Model Debugging and Safety: Enhanced interpretability could lead to safer, more reliable model deployments in critical applications like real-time translation for diplomatic or emergency settings where errors or misinterpretations could have serious consequences.
• Resource Allocation: Insights on how models handle high-resource versus low-resource language scenarios can guide dataset compilation and training focus, optimizing computational resources and model training strategies.
• Bias Detection and Correction: Understanding token-level contributions could lead to new methods for detecting and mitigating biases in machine translation models.

Speculations on Future Developments in AI

The adaptability of ALTI+ to various encoder-decoder frameworks suggests potential broad applications in other sequence-to-sequence (seq2seq) tasks beyond translation, such as text summarization or interactive chatbots.

Furthermore, with the advancement of models capable of handling more complex multilingual scenarios, tools like ALTI+ could become essential for model developers aiming to understand and improve the intricate dynamics of language-specific NMT. This might also pave the way for more personalized and context-aware machine translation services.

Conclusion

The ALTI+ methodology extends our understanding of neural machine translation by illuminating the influence of both source and target inputs in model predictions. It offers a granular look at the internal mechanisms of a model, providing valuable insights that could drive future innovations in NMT technology.