Multi-passage BERT: A Globally Normalized BERT Model for Open-domain Question Answering (1908.08167v2)

Abstract: BERT model has been successfully applied to open-domain QA tasks. However, previous work trains BERT by viewing passages corresponding to the same question as independent training instances, which may cause incomparable scores for answers from different passages. To tackle this issue, we propose a multi-passage BERT model to globally normalize answer scores across all passages of the same question, and this change enables our QA model find better answers by utilizing more passages. In addition, we find that splitting articles into passages with the length of 100 words by sliding window improves performance by 4%. By leveraging a passage ranker to select high-quality passages, multi-passage BERT gains additional 2%. Experiments on four standard benchmarks showed that our multi-passage BERT outperforms all state-of-the-art models on all benchmarks. In particular, on the OpenSQuAD dataset, our model gains 21.4% EM and 21.5% $F_1$ over all non-BERT models, and 5.8% EM and 6.5% $F_1$ over BERT-based models.

• The paper introduces a novel globally normalized approach that integrates evidence from multiple passages to improve open-domain question answering.
• It leverages the BERT architecture to jointly process and rank passages, effectively mitigating errors from passage-level decisions.
• Empirical results show significant gains in accuracy, highlighting the model’s potential for real-world QA applications.

Overview of the ACL 2019 Proceedings Instruction Paper

This document serves as a guideline for preparing a camera-ready manuscript for the Association for Computational Linguistics (ACL) 2019 proceedings. It addresses authors who are contributing papers for review or submitting final versions of accepted papers. The guidelines outlined in this document ensure the uniformity and consistency of the proceedings in terms of structure and presentation.

General Formatting Instructions

The document stipulates that all manuscripts must be formatted in a two-column layout. Exceptions are made for the title, author names, affiliations, and full-width figures or tables. Manuscripts should be single-spaced, conform to A4 paper dimensions, and respect a defined page limit: eight pages for long papers and four pages for short ones, excluding references. Review submissions must remain anonymous through a double-blind process, necessitating a lack of author information until the final versions are submitted.

Styles and Electronic Resources

The paper offers practical advice for authors using LaTeX to prepare their papers. It emphasizes the importance of using the ACL-specific LaTeX style files (acl2019.sty), which are engineered to handle the formatting requirements accurately. The document provides instructions for generating PDFs from LaTeX using software tools like pdflatex or Adobe Distiller.

Page Layout and Typography

Specific layout dimensions are given for margins and column sizes to maintain a uniform appearance across all ACL 2019 submissions. The recommended typefaces include Times Roman or, if unavailable, Computer Modern Roman. The document also mandates particular font sizes for diverse sections of a paper, such as titles, section headings, and regular text, ensuring readability and consistency.

Structuring and Citing Content

The guidelines meticulously detail the structure of a typical paper, including the order and formatting of titles, abstracts, body text, and references. It integrates recommendations on the use of section headings, footnotes, and figures or tables, urging the usage of grayscale for visuals to improve accessibility.

The paper offers comprehensive guidance on citation styles and methods, suggesting the use of the provided BibTeX style files that comply with the format's requirements, thereby facilitating appropriate referencing of DOI or ACL Anthology Identifiers.

Addressing Non-English Terms and Supplementary Material

Authors are advised to provide transliterations or translations of non-English terms to enhance comprehensibility. Furthermore, guidelines for supplementing the main paper with appendices or additional material, such as data or code necessary for replicating experiments, are presented. This additional material must adhere to same anonymity and formatting rules as the main document and be submitted as separate files.

Implications and Future Developments

Compliance with these guidelines is imperative to maintain consistency and quality in conference publications, which reflects the broader efforts of ACL in streamlining peer-reviewed scholarly communication. As the conference continues to adapt to new technologies and publication standards, future guidelines may evolve to incorporate changes in digital publishing practices.

In conclusion, this instructional paper provides a detailed, systematic approach to preparing manuscripts for ACL 2019, ensuring all submissions maintain a high standard of presentation and accessibility. Authors are encouraged to meticulously follow these directives to facilitate a seamless review and publication process.