Controlled Hallucinations: Learning to Generate Faithfully from Noisy Data

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 864–870 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 864 Controlled Hallucinations: Learning to Generate Faithfully from Noisy Data Katja Fillippova Google Research, Berlin, Germany katjaf@google.com Abstract Neural text generation (data- or text-to-text) demonstrates remarkable performance when training data is abundant which for many ap- plications is not the case. To collect a large corpus of parallel data, heuristic rules are often used but they inevitably let noise into the data, such as phrases in the output which cannot be explained by the input. Consequently, mod- els pick up on the noise and may hallucinate– generate fluent but unsupported text. Our con- tribution is a simple but powerful technique to treat such hallucinations as a controllable as- pect of the generated text, without dismissing any input and without modifying the model architecture. On the WikiBio corpus (Lebret et al., 2016), a particularly noisy dataset, we demonstrate the efficacy of the technique both in an automatic and in a human evaluation. 1 Introduction Deep neural network-based (DNN) models have demonstrated remarkable performance on a multi- tude of text-to-text (Bahdanau et al., 2015; Rothe et al., 2019; Narayan et al., 2018; Rush et al., 2015, inter alia) as well as data-to-text generation tasks (Wiseman et al., 2017; Puduppully et al., 2019, in- ter alia). To reach high performance, DNN models require a large training corpus which is normally not readily available. Indeed, it is rare to have a sufficiently large human-curated corpus of parallel data (Koehn, 2005), and researchers have come up with heuristic rules to mine input-output pairs on a large scale (Hermann et al., 2015; Rush et al., 2015; Narayan et al., 2018). No matter how power- ful, DNN models are known to be sensitive to data artifacts (Kaushik and Lipton, 2018) and pick on the noise in the training data. While hallucinations have not been defined for- mally, the term is standardly used to refer to the generated content which is either unfaithful to the Figure 1: Two WikiBio sources and targets with exam- ple attributes: tense and length can be read-off the tar- get directly. When added to the input, the model gets a knob to control for length and tense. We propose to estimate the noise degree by comparing the source with the target thus obtaining a hallucination knob. input, or nonsensical (Maynez et al., 2020). In our work we are concerned with the former hallucina- tion kind which is primarily caused by imperfect quality of the training data. If the data are noisy, how can one reduce the chances of hallucinating? One may try to improve the quality of a dataset and clean it from phrases for which a clear support in the input is missing, or augment the input with information found only in the output. The former path is risky as it easily results in ungrammatical targets. The latter approach of enforcing a stronger alignment between inputs and outputs has been tried …

Content truncated to 500 words