Analyzing the Source and Target Contributions to Predictions in Neural Machine Translation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1126–1140 August 1–6, 2021. ©2021 Association for Computational Linguistics 1126 Analyzing the Source and Target Contributions to Predictions in Neural Machine Translation Elena Voita1,2 Rico Sennrich3,1 Ivan Titov1,2 1University of Edinburgh, Scotland 2University of Amsterdam, Netherlands 3University of Zurich, Switzerland lena-voita@hotmail.com sennrich@cl.uzh.ch ititov@inf.ed.ac.uk Abstract In Neural Machine Translation (and, more gen- erally, conditional language modeling), the generation of a target token is influenced by two types of context: the source and the pre- fix of the target sequence. While many at- tempts to understand the internal workings of NMT models have been made, none of them explicitly evaluates relative source and target contributions to a generation decision. We ar- gue that this relative contribution can be evalu- ated by adopting a variant of Layerwise Rele- vance Propagation (LRP). Its underlying ‘con- servation principle’ makes relevance propaga- tion unique: differently from other methods, it evaluates not an abstract quantity reflecting to- ken importance, but the proportion of each to- ken’s influence. We extend LRP to the Trans- former and conduct an analysis of NMT mod- els which explicitly evaluates the source and target relative contributions to the generation process. We analyze changes in these contribu- tions when conditioning on different types of prefixes, when varying the training objective or the amount of training data, and during the training process. We find that models trained with more data tend to rely on source informa- tion more and to have more sharp token contri- butions; the training process is non-monotonic with several stages of different nature.1 1 Introduction With the success of neural approaches to natu- ral language processing, analysis of NLP models has become an important and active topic of re- search. In NMT, approaches to analysis include probing for linguistic structure (Belinkov et al., 2017; Conneau et al., 2018), evaluating via con- trastive translation pairs (Sennrich, 2017; Burlot and Yvon, 2017; Rios Gonzales et al., 2017; Tang 1We release the code at https://github.com/ lena-voita/the-story-of-heads. et al., 2018), inspecting model components, such as attention (Ghader and Monz, 2017; Voita et al., 2018; Tang et al., 2018; Raganato and Tiedemann, 2018; Voita et al., 2019) or neurons (Dalvi et al., 2019; Bau et al., 2019), among others. Unfortunately, although a lot of work on model analysis has been done, a question of how the NMT predictions are formed remains largely open. Namely, the generation of a target token is defined by two types of context, source and target, but there is no method which explicitly evaluates the rela- tive contribution of source and target to a given prediction. The ability to measure this relative contribution is important for model understanding since previous work showed that NMT models of- ten fail to effectively control information flow from source and target contexts. For example, adding context gates to dynamically control the influence of source and target leads to improvement for both RNN …

Content truncated to 500 words