Survey of Large Language Models on the Text Generation Task

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F62156489%3A43110%2F24%3A43925828" target="_blank" >RIV/62156489:43110/24:43925828 - isvavai.cz</a>

Result on the web

<a href="https://doi.org/10.11118/978-80-7509-990-7-0195" target="_blank" >https://doi.org/10.11118/978-80-7509-990-7-0195</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.11118/978-80-7509-990-7-0195" target="_blank" >10.11118/978-80-7509-990-7-0195</a>

Result language

angličtina

Original language name

Survey of Large Language Models on the Text Generation Task

Original language description

This paper focuses on the comparison of GPT, GPT-2, XLNet, T5 models on text generation tasks. None of the autoencoder models are included in the comparison ranking due to their unsuitability for text generation tasks. The comparison of the models was performed using the BERT-score metric, which calculates precision, recall and F1 values for each sentence. The median was used to obtain the final results from this metric. A preprocessed dataset of empathetic dialogues was used to test the models, which is presented in this paper and compared with other datasets containing dialogues in English. The tested models were only pre-trained and there was no fine-tune on the dataset used for testing. The transformers library from Hugging face and the Python language were used to test the models. The research showed on the pre-trained dataset empathic dialogues has the highest precision model T5, recall and F1 has the highest precision model GPT-2.

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

<a href="/en/project/EF16_017%2F0002334" target="_blank" >EF16_017/0002334: Research Infrastructure for Young Scientists</a><br>

Continuities

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)<br>S - Specificky vyzkum na vysokych skolach

Publication year

2024

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

26th International Conference Economic Competitiveness and Sustainability: Proceedings

ISBN

978-80-7509-990-7

ISSN

—

e-ISSN

—

Number of pages

6

Pages from-to

195-200

Publisher name

Mendelova univerzita v Brně

Place of publication

Brno

Event location

Brno

Event date

Mar 21, 2024

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

—