Question Generation Capabilities of “Small" Large Language Models

Publikation: Beitrag in Buch/Bericht/Sammelwerk/KonferenzbandAufsatz in KonferenzbandForschungPeer-Review

Autoren

  • Joshua Berger
  • Jonathan Koß
  • Markos Stamatakis
  • Anett Hoppe
  • Ralph Ewerth
  • Christian Wartena

Organisationseinheiten

Externe Organisationen

  • Hochschule Hannover (HsH)
  • Technische Informationsbibliothek (TIB) Leibniz-Informationszentrum Technik und Naturwissenschaften und Universitätsbibliothek
Forschungs-netzwerk anzeigen

Details

OriginalspracheEnglisch
Titel des SammelwerksNatural Language Processing and Information Systems
Untertitel29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings
Herausgeber/-innenAmon Rapp, Luigi Di Caro, Farid Meziane, Vijayan Sugumaran
Herausgeber (Verlag)Springer Science and Business Media Deutschland GmbH
Seiten183-194
Seitenumfang12
ISBN (elektronisch)978-3-031-70242-6
ISBN (Print)9783031702419
PublikationsstatusVeröffentlicht - 20 Sept. 2024
Veranstaltung29th International Conference on Natural Language and Information Systems, NLDB 2024 - Turin, Italien
Dauer: 25 Juni 202427 Juni 2024

Publikationsreihe

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Band14763 LNCS
ISSN (Print)0302-9743
ISSN (elektronisch)1611-3349

Abstract

Questions are an integral part of test formats in education. Also online learning platforms like Coursera or Udemy use questions to check learners’ understanding. However, the manual creation of questions can be very time-intensive. This problem can be mitigated through automatic question generation. In this paper, we present a comparison of fine-tuned text-generating transformers for question generation. Our methods include (i) a comparison of multiple fine-tuned transformers to identify differences in the generated output, (ii) a comparison of multiple token search strategies evaluated on each model to find differences in generated questions across different strategies and (iii) a newly developed manual evaluation metric that evaluates generated questions regarding aspects of naturalness and suitability. Our experiments show a difference in question length, structure and quality depending on the used transformer architecture, which indicates a correlation between transformer architecture and question structure. Furthermore, different search strategies for the same model architecture do not greatly impact structure or quality.

ASJC Scopus Sachgebiete

Zitieren

Question Generation Capabilities of “Small" Large Language Models. / Berger, Joshua; Koß, Jonathan; Stamatakis, Markos et al.
Natural Language Processing and Information Systems : 29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings. Hrsg. / Amon Rapp; Luigi Di Caro; Farid Meziane; Vijayan Sugumaran. Springer Science and Business Media Deutschland GmbH, 2024. S. 183-194 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Band 14763 LNCS).

Publikation: Beitrag in Buch/Bericht/Sammelwerk/KonferenzbandAufsatz in KonferenzbandForschungPeer-Review

Berger, J, Koß, J, Stamatakis, M, Hoppe, A, Ewerth, R & Wartena, C 2024, Question Generation Capabilities of “Small" Large Language Models. in A Rapp, L Di Caro, F Meziane & V Sugumaran (Hrsg.), Natural Language Processing and Information Systems : 29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bd. 14763 LNCS, Springer Science and Business Media Deutschland GmbH, S. 183-194, 29th International Conference on Natural Language and Information Systems, NLDB 2024, Turin, Italien, 25 Juni 2024. https://doi.org/10.1007/978-3-031-70242-6_18
Berger, J., Koß, J., Stamatakis, M., Hoppe, A., Ewerth, R., & Wartena, C. (2024). Question Generation Capabilities of “Small" Large Language Models. In A. Rapp, L. Di Caro, F. Meziane, & V. Sugumaran (Hrsg.), Natural Language Processing and Information Systems : 29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings (S. 183-194). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Band 14763 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-70242-6_18
Berger J, Koß J, Stamatakis M, Hoppe A, Ewerth R, Wartena C. Question Generation Capabilities of “Small" Large Language Models. in Rapp A, Di Caro L, Meziane F, Sugumaran V, Hrsg., Natural Language Processing and Information Systems : 29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings. Springer Science and Business Media Deutschland GmbH. 2024. S. 183-194. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-031-70242-6_18
Berger, Joshua ; Koß, Jonathan ; Stamatakis, Markos et al. / Question Generation Capabilities of “Small" Large Language Models. Natural Language Processing and Information Systems : 29th International Conference on Applications of Natural Language to Information Systems, NLDB 2024, Proceedings. Hrsg. / Amon Rapp ; Luigi Di Caro ; Farid Meziane ; Vijayan Sugumaran. Springer Science and Business Media Deutschland GmbH, 2024. S. 183-194 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
Download
@inproceedings{2c41e6fa34954e20afabdb8a335b3d0d,
title = "Question Generation Capabilities of “Small{"} Large Language Models",
abstract = "Questions are an integral part of test formats in education. Also online learning platforms like Coursera or Udemy use questions to check learners{\textquoteright} understanding. However, the manual creation of questions can be very time-intensive. This problem can be mitigated through automatic question generation. In this paper, we present a comparison of fine-tuned text-generating transformers for question generation. Our methods include (i) a comparison of multiple fine-tuned transformers to identify differences in the generated output, (ii) a comparison of multiple token search strategies evaluated on each model to find differences in generated questions across different strategies and (iii) a newly developed manual evaluation metric that evaluates generated questions regarding aspects of naturalness and suitability. Our experiments show a difference in question length, structure and quality depending on the used transformer architecture, which indicates a correlation between transformer architecture and question structure. Furthermore, different search strategies for the same model architecture do not greatly impact structure or quality.",
keywords = "Automatic Question Generation, Pre-trained Transformer, Transformer Architecture",
author = "Joshua Berger and Jonathan Ko{\ss} and Markos Stamatakis and Anett Hoppe and Ralph Ewerth and Christian Wartena",
note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.; 29th International Conference on Natural Language and Information Systems, NLDB 2024 ; Conference date: 25-06-2024 Through 27-06-2024",
year = "2024",
month = sep,
day = "20",
doi = "10.1007/978-3-031-70242-6_18",
language = "English",
isbn = "9783031702419",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Science and Business Media Deutschland GmbH",
pages = "183--194",
editor = "Amon Rapp and {Di Caro}, Luigi and Farid Meziane and Vijayan Sugumaran",
booktitle = "Natural Language Processing and Information Systems",
address = "Germany",

}

Download

TY - GEN

T1 - Question Generation Capabilities of “Small" Large Language Models

AU - Berger, Joshua

AU - Koß, Jonathan

AU - Stamatakis, Markos

AU - Hoppe, Anett

AU - Ewerth, Ralph

AU - Wartena, Christian

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.

PY - 2024/9/20

Y1 - 2024/9/20

N2 - Questions are an integral part of test formats in education. Also online learning platforms like Coursera or Udemy use questions to check learners’ understanding. However, the manual creation of questions can be very time-intensive. This problem can be mitigated through automatic question generation. In this paper, we present a comparison of fine-tuned text-generating transformers for question generation. Our methods include (i) a comparison of multiple fine-tuned transformers to identify differences in the generated output, (ii) a comparison of multiple token search strategies evaluated on each model to find differences in generated questions across different strategies and (iii) a newly developed manual evaluation metric that evaluates generated questions regarding aspects of naturalness and suitability. Our experiments show a difference in question length, structure and quality depending on the used transformer architecture, which indicates a correlation between transformer architecture and question structure. Furthermore, different search strategies for the same model architecture do not greatly impact structure or quality.

AB - Questions are an integral part of test formats in education. Also online learning platforms like Coursera or Udemy use questions to check learners’ understanding. However, the manual creation of questions can be very time-intensive. This problem can be mitigated through automatic question generation. In this paper, we present a comparison of fine-tuned text-generating transformers for question generation. Our methods include (i) a comparison of multiple fine-tuned transformers to identify differences in the generated output, (ii) a comparison of multiple token search strategies evaluated on each model to find differences in generated questions across different strategies and (iii) a newly developed manual evaluation metric that evaluates generated questions regarding aspects of naturalness and suitability. Our experiments show a difference in question length, structure and quality depending on the used transformer architecture, which indicates a correlation between transformer architecture and question structure. Furthermore, different search strategies for the same model architecture do not greatly impact structure or quality.

KW - Automatic Question Generation

KW - Pre-trained Transformer

KW - Transformer Architecture

UR - http://www.scopus.com/inward/record.url?scp=85205514065&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-70242-6_18

DO - 10.1007/978-3-031-70242-6_18

M3 - Conference contribution

AN - SCOPUS:85205514065

SN - 9783031702419

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 183

EP - 194

BT - Natural Language Processing and Information Systems

A2 - Rapp, Amon

A2 - Di Caro, Luigi

A2 - Meziane, Farid

A2 - Sugumaran, Vijayan

PB - Springer Science and Business Media Deutschland GmbH

T2 - 29th International Conference on Natural Language and Information Systems, NLDB 2024

Y2 - 25 June 2024 through 27 June 2024

ER -