The Curious Case of Hallucinations in Neural Machine Translation

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1172–1183 June 6–11, 2021. ©2021 Association for Computational Linguistics 1172 The Curious Case of Hallucinations in Neural Machine Translation Vikas Raunak Arul Menezes Marcin Junczys-Dowmunt Microsoft, USA {viraunak, arulm, marcinjd}@microsoft.com Abstract In this work, we study hallucinations in Neu- ral Machine Translation (NMT), which lie at an extreme end on the spectrum of NMT pathologies. Firstly, we connect the phe- nomenon of hallucinations under source per- turbation to the Long-Tail theory of Feldman (2020), and present an empirically validated hypothesis that explains hallucinations under source perturbation. Secondly, we consider hallucinations under corpus-level noise (with- out any source perturbation) and demonstrate that two prominent types of natural hallucina- tions (detached and oscillatory outputs) could be generated and explained through specific corpus-level noise patterns. Finally, we eluci- date the phenomenon of hallucination ampli- fication in popular data-generation processes such as Backtranslation and sequence-level Knowledge Distillation. We have released the datasets and code to replicate our re- sults at https://github.com/vyraun/ hallucinations. 1 Introduction Neural Machine Translation (NMT) enjoys tremen- dous success, far surpassing the performance of pre- vious statistical approaches in high-to-moderate re- source settings (Koehn and Knowles, 2017). How- ever, NMT suffers from well known pathologies such as coverage (Tu et al., 2016), mistranslation of named entities (Ugawa et al., 2018), etc. In terms of adequacy of the generated output (Martindale et al., 2019), hallucinations are egregious mistakes that lie at the extreme end of NMT pathologies. Such hallucinated outputs are characterized as be- ing decoupled from the source sequence, despite being (fully or moderately) fluent in the target lan- guage (Müller et al., 2020). Two main hallucina- tion phenomena have been reported in the existing literature: 1. NMT models tend to generate hallucinated outputs under certain cases of source pertur- bation (Lee et al., 2018). 2. NMT models have a propensity to hallucinate more frequently under out-of-domain inputs (Müller et al., 2020). However, a plausible theory to explain the genera- tion of different types of hallucinations, including the above two results is still lacking in the NMT literature. Lee et al. (2018) posited that hallucina- tions could be happening due to decoder instability, however, their experiments to engineer solutions based on this proved inconclusive. In this work, we present a systematic study of different kinds of hallucinations, studying them through the lens of generalization, memorization and optimization in sequence to sequence models. Our key contribu- tions are as follows: 1. We extend the Memorization Value Estimator proposed in Feldman and Zhang (2020) to the sequence to sequence setting and demonstrate that hallucinations under source-side perturba- tions could be explained through the long-tail theory they propose. 2. We introduce corpus-level noise into NMT parallel corpora and show that specific noise patterns interact with sequence to sequence training dynamics in different ways to gen- erate the prominent hallucination patterns re- ported in the literature (Lee et al., 2018). 3. We demonstrate the phenomenon …

Content truncated to 500 words