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: On the Robustness of Reading Comprehension Models to Entity Renaming

Abstract: We study the robustness of machine reading comprehension (MRC) models to entity renaming -- do models make more wrong predictions when answer entities have different names? Such failures would indicate that models are overly reliant on entity knowledge to answer questions, and therefore may generalize poorly when facts about the world change or questions are asked about novel entities. To systematically audit model robustness, we propose a general and scalable method to replace person names with names from a variety of sources, ranging from common English names to names from other languages to arbitrary strings. Across four datasets and three pretrained model architectures, MRC models consistently perform worse when entities are renamed, with particularly large accuracy drops on datasets constructed via distant supervision. We also find large differences between models: SpanBERT, which is pretrained with span-level masking, is more robust than RoBERTa, despite having similar accuracy on unperturbed test data. Inspired by this, we experiment with span-level and entity-level masking as a continual pretraining objective and find that they can further improve the robustness of MRC models.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2110.08555 [cs.CL]
  (or arXiv:2110.08555v1 [cs.CL] for this version)

Submission history

From: Jun Yan [view email]
[v1] Sat, 16 Oct 2021 11:46:32 GMT (348kb,D)

Link back to: arXiv, form interface, contact.