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

Download:

Current browse context:

cs.CL

Change to browse by:

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: R2-D2: A Modular Baseline for Open-Domain Question Answering

Abstract: This work presents a novel four-stage open-domain QA pipeline R2-D2 (Rank twice, reaD twice). The pipeline is composed of a retriever, passage reranker, extractive reader, generative reader and a mechanism that aggregates the final prediction from all system's components. We demonstrate its strength across three open-domain QA datasets: NaturalQuestions, TriviaQA and EfficientQA, surpassing state-of-the-art on the first two. Our analysis demonstrates that: (i) combining extractive and generative reader yields absolute improvements up to 5 exact match and it is at least twice as effective as the posterior averaging ensemble of the same models with different parameters, (ii) the extractive reader with fewer parameters can match the performance of the generative reader on extractive QA datasets.
Comments: Accepted to Findings of EMNLP'21. arXiv admin note: substantial text overlap with arXiv:2102.10697
Subjects: Computation and Language (cs.CL); Information Retrieval (cs.IR); Machine Learning (cs.LG)
Cite as: arXiv:2109.03502 [cs.CL]
  (or arXiv:2109.03502v1 [cs.CL] for this version)

Submission history

From: Martin Fajčík [view email]
[v1] Wed, 8 Sep 2021 08:51:27 GMT (585kb,D)

Link back to: arXiv, form interface, contact.