Neural sequence generation models are known to “hallucinate”, by producing outputs that are unrelated to the source text. These hallucinations are potentially harmful, yet it remains unclear in what conditions they arise and how to mitigate their impact. In this work, we first identify internal model symptoms of hallucinations by analyzing the relative token contributions to the generation in contrastive hallucinated vs. non-hallucinated outputs generated via source perturbations. We then show that these symptoms are reliable indicators of natural hallucinations, by using them to design a lightweight hallucination detector which outperforms both model-free baselines and strong classifiers based on quality estimation or large pre-trained models on manually annotated English-Chinese and German-English translation test beds.

Successful Machine Translation (MT) deployment requires understanding not only the intrinsic qualities of MT output, such as fluency and adequacy, but also user perceptions. Users who do not understand the source language respond to MT output based on their perception of the likelihood that the meaning of the MT output matches the meaning of the source text. We refer to this as believability. Output that is not believable may be off-putting to users, but believable MT output with incorrect meaning may mislead them. In this work, we study the relationship of believability to fluency and adequacy by applying traditional MT direct assessment protocols to annotate all three features on the output of neural MT systems. Quantitative analysis of these annotations shows that believability is closely related to but distinct from fluency, and initial qualitative analysis suggests that semantic features may account for the difference.

Stylistic variations of language, such as formality, carry speakers’ intention beyond literal meaning and should be conveyed adequately in translation. We propose to use lexical formality models to control the formality level of machine translation output. We demonstrate the effectiveness of our approach in empirical evaluations, as measured by automatic metrics and human assessments.

Statistical post-editing has been shown in several studies to increase BLEU score for rule-based MT systems. However, previous studies have relied solely on BLEU and have not conducted further study to determine whether those gains indicated an increase in quality or in score alone. In this work we conduct a human evaluation of statistical post-edited output from a weak rule-based MT system, comparing the results with the output of the original rule-based system and a phrase-based statistical MT system trained on the same data. We show that for this weak rule-based system, despite significant BLEU score increases, human evaluators prefer the output of the original system. While this is not a generally conclusive condemnation of statistical post-editing, this result does cast doubt on the efficacy of statistical post-editing for weak MT systems and on the reliability of BLEU score for comparison between weak rule-based and hybrid systems built from them.