A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness

Enguerrand Boitel; Alaa Mohasseb; Ella Haig

doi:10.1007/978-3-031-47508-5_44

A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness

Enguerrand Boitel, Alaa Mohasseb, Ella Haig

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

This paper presents a comparative analysis between the GPT model and a BERT model for emotion recognition. While GPT has shown remarkable performance across various tasks, the question arises as to whether its results can be compared to models specifically trained for particular tasks. The paper explores the presentations of GPT, including its models (Davinci, Curie, Babbage and Ada) as well as the fine-tuning approach. Similarly, the paper discusses the presentation of BERT, specifically DeBERTa v3, and its fine-tuning methodology. A comparison is then drawn between the two models, considering the general vs. task-specific nature of AI. Furthermore, the paper investigates the reasons behind GPT's exceptional performance and examines why the obtained results for emotion recognition may not be significantly better than those of task-specific models. Ultimately, the conclusion reflects on the better accuracy often exhibited by BERT, but highlights the potential for future advancements, such as the new GPT-4, that may surpass existing models and offer unparalleled versatility in answering various questions.

Original language	English
Title of host publication	Advances in Computational Intelligence Systems
Subtitle of host publication	Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK
Editors	Nitin Naik, Paul Jenkins, Paul Grace, Longzhi Yang, Shaligram Prajapat
Publisher	Springer
Pages	567–579
ISBN (Electronic)	9783031475085
ISBN (Print)	9783031475078
DOIs	https://doi.org/10.1007/978-3-031-47508-5_44
Publication status	Published - 1 Feb 2024
Event	UKCI 2023: The 22nd Workshop on Computational Intelligence - Aston University, Birmingham, United Kingdom Duration: 6 Sept 2023 → 8 Sept 2023 https://www.uk-ci.org/home

Publication series

Name	Advances in Intelligent Systems and Computing
Publisher	Springer
Volume	1453
ISSN (Print)	2194-5357
ISSN (Electronic)	2194-5365

Workshop

Workshop	UKCI 2023: The 22nd Workshop on Computational Intelligence
Country/Territory	United Kingdom
City	Birmingham
Period	6/09/23 → 8/09/23
Internet address	https://www.uk-ci.org/home

Keywords

GPT model
BERT model
emotion recognition
general vs. task specific AI

Access to Document

10.1007/978-3-031-47508-5_44

Embargoed Document

A Comparative Analysis of GPT-3 and BERT Models for Text-based Emotion Recognition: Performance, Efficiency, and Robustness
Accepted author manuscript (Post-print), 782 KB
Due to publisher’s copyright restrictions, this document is not freely available to download from this website until: 1/02/25

Best Paper Award
Boitel, E. (Recipient), Mohasseb, A. (Recipient) & Haig, E. (Recipient), Sept 2023
Prize: Prize (including medals and awards)

Cite this

Boitel, E., Mohasseb, A., & Haig, E. (2024). A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness. In N. Naik, P. Jenkins, P. Grace, L. Yang, & S. Prajapat (Eds.), Advances in Computational Intelligence Systems: Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK (pp. 567–579). (Advances in Intelligent Systems and Computing; Vol. 1453). Springer. https://doi.org/10.1007/978-3-031-47508-5_44

Boitel, Enguerrand ; Mohasseb, Alaa ; Haig, Ella. / A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness. Advances in Computational Intelligence Systems: Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK. editor / Nitin Naik ; Paul Jenkins ; Paul Grace ; Longzhi Yang ; Shaligram Prajapat. Springer, 2024. pp. 567–579 (Advances in Intelligent Systems and Computing).

@inproceedings{07f1c15911ad451eb9936ad0ac51a3a4,

title = "A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness",

abstract = "This paper presents a comparative analysis between the GPT model and a BERT model for emotion recognition. While GPT has shown remarkable performance across various tasks, the question arises as to whether its results can be compared to models specifically trained for particular tasks. The paper explores the presentations of GPT, including its models (Davinci, Curie, Babbage and Ada) as well as the fine-tuning approach. Similarly, the paper discusses the presentation of BERT, specifically DeBERTa v3, and its fine-tuning methodology. A comparison is then drawn between the two models, considering the general vs. task-specific nature of AI. Furthermore, the paper investigates the reasons behind GPT's exceptional performance and examines why the obtained results for emotion recognition may not be significantly better than those of task-specific models. Ultimately, the conclusion reflects on the better accuracy often exhibited by BERT, but highlights the potential for future advancements, such as the new GPT-4, that may surpass existing models and offer unparalleled versatility in answering various questions.",

keywords = "GPT model, BERT model, emotion recognition, general vs. task specific AI",

author = "Enguerrand Boitel and Alaa Mohasseb and Ella Haig",

note = "12 month embargo. This version of the article has been accepted for publication, after peer review (when applicable) and is subject to Springer Nature{\textquoteright}s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: http://dx.doi.org/[insert DOI]; UKCI 2023: The 22nd Workshop on Computational Intelligence ; Conference date: 06-09-2023 Through 08-09-2023",

year = "2024",

month = feb,

day = "1",

doi = "10.1007/978-3-031-47508-5_44",

language = "English",

isbn = "9783031475078",

series = "Advances in Intelligent Systems and Computing",

publisher = "Springer",

pages = "567–579",

editor = "Naik, {Nitin } and Jenkins, {Paul } and Grace, {Paul } and Yang, {Longzhi } and Prajapat, {Shaligram }",

booktitle = "Advances in Computational Intelligence Systems",

url = "https://www.uk-ci.org/home",

}

Boitel, E, Mohasseb, A & Haig, E 2024, A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness. in N Naik, P Jenkins, P Grace, L Yang & S Prajapat (eds), Advances in Computational Intelligence Systems: Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK. Advances in Intelligent Systems and Computing, vol. 1453, Springer, pp. 567–579, UKCI 2023: The 22nd Workshop on Computational Intelligence, Birmingham, United Kingdom, 6/09/23. https://doi.org/10.1007/978-3-031-47508-5_44

A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness. / Boitel, Enguerrand; Mohasseb, Alaa ; Haig, Ella.
Advances in Computational Intelligence Systems: Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK. ed. / Nitin Naik; Paul Jenkins; Paul Grace; Longzhi Yang; Shaligram Prajapat. Springer, 2024. p. 567–579 (Advances in Intelligent Systems and Computing; Vol. 1453).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness

AU - Boitel, Enguerrand

AU - Mohasseb, Alaa

AU - Haig, Ella

N1 - 12 month embargo. This version of the article has been accepted for publication, after peer review (when applicable) and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: http://dx.doi.org/[insert DOI]

PY - 2024/2/1

Y1 - 2024/2/1

N2 - This paper presents a comparative analysis between the GPT model and a BERT model for emotion recognition. While GPT has shown remarkable performance across various tasks, the question arises as to whether its results can be compared to models specifically trained for particular tasks. The paper explores the presentations of GPT, including its models (Davinci, Curie, Babbage and Ada) as well as the fine-tuning approach. Similarly, the paper discusses the presentation of BERT, specifically DeBERTa v3, and its fine-tuning methodology. A comparison is then drawn between the two models, considering the general vs. task-specific nature of AI. Furthermore, the paper investigates the reasons behind GPT's exceptional performance and examines why the obtained results for emotion recognition may not be significantly better than those of task-specific models. Ultimately, the conclusion reflects on the better accuracy often exhibited by BERT, but highlights the potential for future advancements, such as the new GPT-4, that may surpass existing models and offer unparalleled versatility in answering various questions.

AB - This paper presents a comparative analysis between the GPT model and a BERT model for emotion recognition. While GPT has shown remarkable performance across various tasks, the question arises as to whether its results can be compared to models specifically trained for particular tasks. The paper explores the presentations of GPT, including its models (Davinci, Curie, Babbage and Ada) as well as the fine-tuning approach. Similarly, the paper discusses the presentation of BERT, specifically DeBERTa v3, and its fine-tuning methodology. A comparison is then drawn between the two models, considering the general vs. task-specific nature of AI. Furthermore, the paper investigates the reasons behind GPT's exceptional performance and examines why the obtained results for emotion recognition may not be significantly better than those of task-specific models. Ultimately, the conclusion reflects on the better accuracy often exhibited by BERT, but highlights the potential for future advancements, such as the new GPT-4, that may surpass existing models and offer unparalleled versatility in answering various questions.

KW - GPT model

KW - BERT model

KW - emotion recognition

KW - general vs. task specific AI

UR - https://link.springer.com/book/9783031475078

U2 - 10.1007/978-3-031-47508-5_44

DO - 10.1007/978-3-031-47508-5_44

M3 - Conference contribution

SN - 9783031475078

T3 - Advances in Intelligent Systems and Computing

SP - 567

EP - 579

BT - Advances in Computational Intelligence Systems

A2 - Naik, Nitin

A2 - Jenkins, Paul

A2 - Grace, Paul

A2 - Yang, Longzhi

A2 - Prajapat, Shaligram

PB - Springer

T2 - UKCI 2023: The 22nd Workshop on Computational Intelligence

Y2 - 6 September 2023 through 8 September 2023

ER -

Boitel E, Mohasseb A , Haig E. A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness. In Naik N, Jenkins P, Grace P, Yang L, Prajapat S, editors, Advances in Computational Intelligence Systems: Contributions Presented at the 22nd UK Workshop on Computational Intelligence (UKCI 2023), September 6–8, 2023, Birmingham, UK. Springer. 2024. p. 567–579. (Advances in Intelligent Systems and Computing). doi: 10.1007/978-3-031-47508-5_44

A comparative analysis of GPT-3 and BERT models for text-based emotion recognition: performance, efficiency, and robustness

Abstract

Publication series

Workshop

Keywords

Access to Document

Embargoed Document

Fingerprint

Prizes

Best Paper Award

Cite this