Enhancing Extractive Question Answering through Generative Models: Addressing Label Sparsity and Multi-Span Answers

Registration ID: IJNRD_226466

Published ID: IJNRD2408165

DOI: Click Here to Get

Rajendra Arakh , Alok Gupta , Ruchi Patel

Extractive Question Answering (EQA), Synthetic Data Generation, Fine-tuning, Longformer-based Generative Augmentation (LFG-AUG), Exact Match (EM)

Extractive Question Answering (EQA) models have become integral components of modern natural language processing, providing automated information retrieval capabilities. However, they often face challenges with label sparsity and identifying answers that span multiple text segments (multi-span answers), limiting their effectiveness and applicability. This research proposes a novel approach, LFG-Aug, which combines the Longformer architecture from Hugging Face Transformers with a generative augmentation strategy to address these limitations. By employing synthetic data generation and fine-tuning, LFG-Aug enhances the performance of EQA models, particularly in handling multi-span answers and mitigating label sparsity issues. The evaluation of our approach utilizes the AllenAI/longformer-large-4096-finetuned-triviaQA model, a strong baseline, and demonstrates the effectiveness of LFG-Aug. The integration of the Longformer architecture enables the model to process long sequences effectively, making it well-suited for complex, real-world scenarios. The generative augmentation strategy addresses label sparsity by creating diverse training examples, enhancing the model's ability to generalize and understand multi-span answers. The evaluation results showcase superior performance, with significant improvements in Exact Match (EM) and F1 scores compared to baseline models. This research contributes to the advancement of EQA techniques, broadening their applicability and robustness, and offering promising directions for future work in this field. The proposed LFG-Aug model, combining the strengths of the Longformer architecture and generative modeling, highlights the potential of hybrid approaches in addressing longstanding challenges in EQA. By generating synthetic training data and fine-tuning the model, LFG-Aug enhances the understanding of complex, multi-span answers and improves EQA accuracy. This abstract provides a comprehensive overview of our research, detailing the challenges, proposed approach, methodology, and the significance of our findings in enhancing EQA through generative models.

Published Paper   E-Certificate

Rajendra Arakh , Alok Gupta, & Ruchi Patel (August-2024). Enhancing Extractive Question Answering through Generative Models: Addressing Label Sparsity and Multi-Span Answers . INTERNATIONAL JOURNAL OF NOVEL RESEARCH AND DEVELOPMENT, 9(8), b702-b709. https://ijnrd.org/papers/IJNRD2408165.pdf

Volume 9 Issue 8, August-2024

Pages : b702-b709

Paper Reg. ID: IJNRD_226466

Published Paper Id: IJNRD2408165

Downloads: 000121983

Research Area: Computer Science & Technology 

Country: Jabalpur , Madhya Pradesh , India

Published Paper PDF: https://ijnrd.org/papers/IJNRD2408165.pdf

Published Paper URL: https://ijnrd.org/viewpaperforall?paper=IJNRD2408165

Journal Name: INTERNATIONAL JOURNAL OF NOVEL RESEARCH AND DEVELOPMENT(IJNRD)

ISSN: 2456-4184 | IMPACT FACTOR: 8.76 Calculated By Google Scholar | ESTD YEAR: 2016

An International Scholarly Open Access Journal, Peer-Reviewed, Refereed Journal Impact Factor 8.76 Calculate by Google Scholar and Semantic Scholar | AI-Powered Research Tool, Multidisciplinary, Monthly, Multilanguage Journal Indexing in All Major Database & Metadata, Citation Generator

Publisher: IJNRD (IJ Publication) Janvi Wave | IJNRD.ORG | IJNRD.COM | IJPUB.ORG

This work is licensed under a Creative Commons Attribution 4.0 International License and The Open Definition

Call For Paper - Volume 10 | Issue 10 | October 2025

IJNRD is a Scholarly Open Access, Peer-reviewed, and Refereed Journal with a High Impact Factor of 8.76 (calculated by Google Scholar & Semantic Scholar | AI-Powered Research Tool). It is a Multidisciplinary, Monthly, Low-Cost Journal that follows UGC CARE 2025 Peer-Reviewed Journal Policy norms, Scopus journal standards, and Transparent Peer Review practices to ensure quality and credibility. IJNRD provides indexing in all major databases & metadata repositories, a citation generator, and Digital Object Identifier (DOI) for every published article with full open-access visibility.

The INTERNATIONAL JOURNAL OF NOVEL RESEARCH AND DEVELOPMENT (IJNRD) aims to advance applied, theoretical, and experimental research across diverse fields. Its goal is to promote global scientific information exchange among researchers, developers, engineers, academicians, and practitioners. IJNRD serves as a platform where educators and professionals can share research evidence, models of best practice, and innovative ideas, contributing to academic growth and industry relevance.

Indexing Coverage includes Google Scholar, SSRN, ResearcherID-Publons, Semantic Scholar (AI-Powered Research Tool), Microsoft Academic, Academia.edu, arXiv.org, ResearchGate, CiteSeerX, ResearcherID (Thomson Reuters), Mendeley, DocStoc, ISSUU, Scribd, and many more recognized academic repositories.

How to submit the paper?

By Our website

Click Here to Submit Paper Online

By Mail

Send your paper to editor@ijnrd.org

Important Dates for Current issue

Paper Submission Open For: October 2025

Current Issue: Volume 10 | Issue 10 | October 2025

Impact Factor: 8.76

Last Date for Paper Submission: Till 31-Oct-2025

Notification of Review Result: Within 1-2 Days after Submitting paper.

Publication of Paper: Within 01-02 Days after Submititng documents.

Frequency: Monthly (12 issue Annually).

Journal Type: IJNRD is an International Peer-reviewed, Refereed, and Open Access Journal with Transparent Peer Review as per the new UGC CARE 2025 guidelines, offering low-cost multidisciplinary publication with Crossref DOI and global indexing.

Subject Category: Research Area

Call for Paper: More Details