Reflective Decoding: Beyond Unidirectional Generation with Off-the-Shelf Language Models

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1435–1450 August 1–6, 2021. ©2021 Association for Computational Linguistics 1435 Reflective Decoding: Beyond Unidirectional Generation with Off-the-Shelf Language Models Peter West1,2 Ximing Lu1,2 Ari Holtzman1 Chandra Bhagavatula2 Jena Hwang2 Yejin Choi1,2 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Allen Institute for Artificial Intelligence {pawest, ahai, yejin}@cs.washington.edu {ximinglu, chandrab, jenah}@allenai.org Abstract Publicly available, large pretrained Language Models (LMs) generate text with remarkable quality, but only sequentially from left to right. As a result, they are not immediately appli- cable to generation tasks that break the unidi- rectional assumption, such as paraphrasing or text-infilling, necessitating task-specific super- vision. In this paper, we present REFLECTIVE DE- CODING, a novel unsupervised algorithm that allows for direct application of unidirectional LMs to non-sequential tasks. Our 2-step ap- proach requires no supervision or even paral- lel corpora, only two off-the-shelf pretrained LMs in opposite directions: forward and back- ward. First, in the contextualization step, we use LMs to generate ensembles of past and future contexts which collectively capture the input (e.g. the source sentence for paraphras- ing). Second, in the reflection step, we condi- tion on these “context ensembles”, generating outputs that are compatible with them. Com- prehensive empirical results1 demonstrate that REFLECTIVE DECODING outperforms strong unsupervised baselines on both paraphrasing and abductive text infilling, significantly nar- rowing the gap between unsupervised and su- pervised methods. REFLECTIVE DECODING surpasses multiple supervised baselines on var- ious metrics including human evaluation. 1 Introduction Language Models (LMs) like GPT-2 (Radford et al., 2019), trained over vast unstructured data, can leverage enhanced generation methods (Holtzman et al., 2020; Martins et al., 2020; Welleck et al., 2019) to give fluent and coherent continuations to given input text—e.g. news articles or stories. 1Further results and resource are available at https://homes.cs.washington.edu/˜pawest/ ReflectiveDecoding.html !!"#: How are circulatory system tissues formed? "!: How do circulatory systems form? input paraphrase generated contexts Paraphrasing 1 Sample contexts #$~%& (#|)%&') 2 Sample "! from +, +,()%&') c1: This is a medical question best answered by a doctor… … c2: As with all tissue in the body, this begins with cell division … c3: is one of many key questions about the circulatory system … /: I picked her up and took her to San Francisco General hospital. input hypothesis 1 Sample contexts #$~%& (#|2*, 2+) 2 Sample "! from +, … c1: The day after her discharge she told me she was a lot better … 1(: Amy had heart palpitations after a lot of caffiene 1): By the time she arrived her heart felt much better +,(2*, 2+) -NLG generated contexts Figure 1: Illustration of REFLECTIVE DECODING ap- plied to paraphrasing and abductive infilling (αNLG Bhagavatula et al., 2020). Only the right-context is shown, although both right- and left-contexts are used in practice. First the contextualization step (1) captures aspects of an input by generating many representative contexts for …

Content truncated to 500 words