We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Break, Perturb, Build: Automatic Perturbation of Reasoning Paths Through Question Decomposition

Abstract: Recent efforts to create challenge benchmarks that test the abilities of natural language understanding models have largely depended on human annotations. In this work, we introduce the "Break, Perturb, Build" (BPB) framework for automatic reasoning-oriented perturbation of question-answer pairs. BPB represents a question by decomposing it into the reasoning steps that are required to answer it, symbolically perturbs the decomposition, and then generates new question-answer pairs. We demonstrate the effectiveness of BPB by creating evaluation sets for three reading comprehension (RC) benchmarks, generating thousands of high-quality examples without human intervention. We evaluate a range of RC models on our evaluation sets, which reveals large performance gaps on generated examples compared to the original data. Moreover, symbolic perturbations enable fine-grained analysis of the strengths and limitations of models. Last, augmenting the training data with examples generated by BPB helps close the performance gaps, without any drop on the original data distribution.
Comments: Accepted for publication in Transactions of the Association for Computational Linguistics (TACL), 2021. Author's final version
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2107.13935 [cs.CL]
  (or arXiv:2107.13935v2 [cs.CL] for this version)

Submission history

From: Mor Geva [view email]
[v1] Thu, 29 Jul 2021 12:49:03 GMT (698kb,D)
[v2] Mon, 18 Oct 2021 18:23:10 GMT (705kb,D)

Link back to: arXiv, form interface, contact.