Large-Scale, Diverse, Paraphrastic Bitexts via Sampling and Clustering

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 44–54 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 44 Large-scale, Diverse, Paraphrastic Bitexts via Sampling and Clustering J. Edward Hu Abhinav Singh Nils Holzenberger Matt Post Benjamin Van Durme Johns Hopkins University {edward.hu,asingh78,nholzen1}@jhu.edu; {post,vandurme}@cs.jhu.edu Abstract Producing diverse paraphrases of a sentence is a challenging task. Natural paraphrase cor- pora are scarce and limited, while existing large-scale resources are automatically gen- erated via back-translation and rely on beam search, which tends to lack diversity. We de- scribe PARABANK 2, a new resource that con- tains multiple diverse sentential paraphrases, produced from a bilingual corpus using nega- tive constraints, inference sampling, and clus- tering. We show that PARABANK 2 signif- icantly surpasses prior work in both lexical and syntactic diversity while being meaning- preserving, as measured by human judgments and standardized metrics. Further, we illus- trate how such paraphrastic resources may be used to refine contextualized encoders, leading to improvements in downstream tasks. 1 Introduction The ability to understand and produce paraphrases is a basic competency task, one that is often used as a teaching aid to validate if a student under- stands a statement or a concept. Current deep learning systems struggle with this task, exhibit- ing brittleness to both understanding and produc- ing paraphrastic expressions (Iyyer et al., 2018). One crucial factor behind this incompetence is the dearth of sentential paraphrastic data. Many works have sought to leverage the relative abun- dance of sub-sentential paraphrastic resources in paraphrase detection or generation (Napoles et al., 2016). Yet, they fail to capture contextualized word choices or syntactical variations, as word- or phrase-level resources cannot incorporate infor- mation from the whole input sentence. Recent works have focused on leveraging bilin- gual resources to create large sentence-level para- phrastic collections using translation-based meth- ods (Wieting and Gimpel, 2018; Hu et al., 2019). I took this by mistake. I took it by mistake. vzal jsem ho omylem. I took this by mistake. I took it by accident. vzal jsem ho omylem. Source Target (Reference) Paraphrase I took this by mistake. I took it by mistake. … … I took this by mistake. I picked it up accidentally. Constrained Unconstrained ⊕ ⊖ ⊖ ⊖ I took this by mistake. I took it by accident. vzal jsem ho omylem. I took it by mistake. I picked it up accidentally. Clustered I picked up accidentally. I picked it accidentally. I took mistake. I took it. I took by accident. Figure 1: Contrived example paraphrases from previ- ous work (unconstrained and constrained—used with permission) and ours (clustered). However, these works are confined to using beam search in decoding, which tend not to produce di- verse candidates. One approach to force diverse translations is the use of hard lexical constraints at inference time (Hu et al., 2019). While effective in some cases, current approaches to automatic selec- tion of such constraints is based on heuristics and task-oriented trial-and-error. We present a novel resource with accurate …

Content truncated to 500 words