Visualizing and Understanding Neural Machine Translation

By Yanzhuo Ding, Yang Liu, Huanbo Luan, Maosong Sun. 2017

185 citations recorded by the source corpus.

While neural machine translation (NMT) has made remarkable progress in recent years, it is hard to interpret its internal workings due to the continuous representations and non-linearity of neural networks. In this work, we propose to use layer-wise relevance propagation (LRP) to compute the contribution of each contextual word to arbitrary hidden states in the attention-based encoder-decoder framework. We show that visualization with LRP helps to interpret the internal workings of NMT and analyze translation errors.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

Citations in the corpus

Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned (2019, methodology): “We start by identifying the most important heads in each encoder layer using layer-wise relevance propagation (Ding et al., 2017).”
Analysis Methods in Neural Language Processing: A Survey (2018, background): “The important or salient features can then be visualized in selected examples (Li et al., 2016a; Aubakirova and Bansal, 2016; Sundararajan et al., 2017; Arras et al., 2017a,b; Ding et al., 2017; Murdoch et al., 2018; Mudrakarta et al., 2018; Montavon et al., 2018; Godin et al., 2018).”
Neural Approaches to Conversational AI (2018, neutral): “sation history, but also some additional input drawn from the user’s environment, such as an image (Das et al., 2017a; Mostafazadeh et al., 2017) or textual information (Ghazvininejad et al., 2018). 6Ding et al. (2017) indeed found that word repetition errors, usually few in machine translation, are”
Is Attention Interpretable? (2019, background): “The existence of so much work on visualizing attention weights is a testament to attention’s popularity in this regard; to name just a few examples of these weights being examined to understand a model, recent work has focused on goals from explaining and debugging the current system’s decision (Lee”
An Analysis of Encoder Representations in Transformer-Based Machine Translation (2018, background): “Traditionally, a common way to inspect neural networks is by visualizing the hidden representation trained for a specific task (Ding et al., 2017; Strobelt et al., 2018a,b), and to evaluate them by assessing the properties through downstream tasks (Chung et al., 2014; Greff et al., 2017).”
ERASER: A Benchmark to Evaluate Rationalized NLP Models (2019, background): “There has also been work on visual evaluations of saliency maps (Li et al., 2016; Ding et al., 2017; Sundararajan et al., 2017).”
Layer-Wise Coordination between Encoder and Decoder for Neural Machine Translation (2018, background): “In NMT, the generation of the target word depends on both source and target contexts, where source contexts affect the adequacy of the generated sentence while target contexts have impact on the ﬂuency [6, 14, 30, 31].”
Explainable Deep Learning: A Field Guide for the Uninitiated (2020, methodology): “LRP methods create a saliency map that, rather than measuring sensitivity, represents the relevance of each input feature to the output of the network (Bach et al., 2015; Lapuschkin et al., 2016; Arras et al., 2016, 2017; Ding et al., 2017; Montavon et al., 2017).”
Exploiting Deep Representations for Neural Machine Translation (2018, background): “Representation Interpretation Several researchers have tried to visualize the representation of each layer to help better understand what information each layer captures (Zeiler and Fergus, 2014; Li et al., 2016; Ding et al., 2017).”
On internal language representations in deep learning: an analysis of machine translation and speech recognition (2018, methodology): “For instance, layer-wise relevance propagation (LRP), which propagates a measure of relevance down the network [37] has been applied to several language processing tasks [10, 11], including neural machine translation [90].”
Explainable Artificial Intelligence: a Systematic Review (2020, methodology): “[353] used LRP for computing the contribution of”
Hallucinations in Neural Machine Translation (2018, methodology): “For example, moving from the original Seq2Seq model [32, 12] to models that utilize attention mechanisms [2], resulted in improved translation quality [23] and better interpretability [15].”
Towards Understanding Neural Machine Translation with Word Importance (2019, unknown): “, 2019) or the hidden units (Ding et al., 2017; Bau et al., 2019).”
Interrogating the Explanatory Power of Attention in Neural Machine Translation (2019, background): “Thus, attention weights should not be used to debug early generation of EOS, and that some other underlying influence in the network (Ding et al., 2017) might be responsible for the model’s decision in this case.”
Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications (2021, methodology): “At the same time, these methods were able to handle different types of data, including images [13], speech [20], text [10], [38], and structured data, such as molecules [162], [165] or genetic sequences [191].”
Saliency-driven Word Alignment Interpretation for Neural Machine Translation (2019, methodology): “ed to computer vision tasks. To the best of our knowledge, Li et al. (2016) presented the only work that directly employs saliency methods to interpret NLP models. Most similar to our work in spirit, Ding et al. (2017) used Layer-wise Relevance Propagation (LRP; Bach et al. 2015), an interpretation ”
Evaluating Explanation Methods for Neural Machine Translation (2020, methodology): “In particular, on structured prediction tasks, many research works design similar methods to understand NMT models (Ding et al., 2017; Alvarez-Melis and Jaakkola, 2017; Ding et al., 2019; He et al., 2019).”
Seq2seq-Vis: A Visual Debugging Tool for Sequence-to-Sequence Models (2018, unknown): “[6] use a layer-wise relevance propagation technique [2] to understand relevance of input with regard to an output in sequenceto-sentence models.”