How Does Context Matter? On the Robustness of Event Detection with Context-Selective Mask Generalization

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 223–235 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 223 How Decoding Strategies Affect the Verifiability of Generated Text Luca Massarelli1∗† Fabio Petroni2† Aleksandra Piktus2† Myle Ott2 Tim Rockt¨aschel2,3 Vassilis Plachouras2 Fabrizio Silvestri2 Sebastian Riedel2,3 1 Sapienza University of Rome massarelli@diag.uniroma1.it 2Facebook AI {fabiopetroni, piktus, myleott, rockt, vplachouras, fsilvestri, sriedel}@fb.com 3University College London Abstract Recent progress in pre-trained language mod- els led to systems that are able to generate text of an increasingly high quality. While sev- eral works have investigated the fluency and grammatical correctness of such models, it is still unclear to which extent the generated text is consistent with factual world knowledge. Here, we go beyond fluency and also investi- gate the verifiability of text generated by state- of-the-art pre-trained language models. A gen- erated sentence is verifiable if it can be corrob- orated or disproved by Wikipedia, and we find that the verifiability of generated text strongly depends on the decoding strategy. In particular, we discover a tradeoff between factuality (i.e., the ability of generating Wikipedia corrobo- rated text) and repetitiveness. While decoding strategies such as top-k and nucleus sampling lead to less repetitive generations, they also produce less verifiable text. Based on these finding, we introduce a simple and effective de- coding strategy which, in comparison to previ- ously used decoding strategies, produces less repetitive and more verifiable text. 1 Introduction Recent years have led to a considerable surge of interest in and capabilities of pre-trained language models (LMs). Today, they play a critical role in many NLP tasks, such as text classification, ma- chine comprehension and natural language infer- ence (Peters et al., 2018; Devlin et al., 2018; Liu et al., 2019a; Yang et al., 2019), to name just a few. They serve as a pre-training objective for down- stream applications and they have been used to showcase and measure the general progress in NLP (Yu et al., 2017; Liu et al., 2019b). Several works (Radford et al., 2019b; Keskar et al., 2019) show the remarkable fluency and gram- ∗Work done during internship with Facebook. †Equal contribution. matical correctness of text decoded from modern LMs. Additionally, recent works (Petroni et al., 2019; Logan et al., 2019; Broscheit, 2019; Roberts et al., 2020) demonstrate that beyond general lin- guistic capabilities, language models can also pick up factual knowledge present in the training data. However, it is unclear if LMs are able to convey such knowledge at decoding time when producing long sequences—do they generate fluent, grammat- ical but “babbler-level” text or can they produce utterances that reflect factual world knowledge? Understanding this behaviour becomes crucially important as the downstream adoption of auto- matically generated text increases. Already to- day LMs face growing scrutiny from the media and the broader society, as well as from the re- searchers themselves. For example, Radford et al. (2019b) initially argued against releasing their mod- els in order to prevent automatic generation of fake news (Radford et al., …

Content truncated to 500 words