Mgr. Lukáš Mikula

Master's thesis

Think Twice Before You Answer: Mitigating Biases of Question Answering Models

Think Twice Before You Answer: Mitigating Biases of Question Answering Models
Abstract:
Veľké jazykové modely (z angl. Large Language Models) založené na architektúre Transformerov predstavujú najlepšie modely pre vačšinu problémov spracovania prirodzeného jazyka (z angl. Natural Language Modeling). Napriek tomu majú tieto modely tendenciu učiť sa skreslenia a systematické chyby z trénovacích datasetov. Toto učenie im síce môže pomôcť na datasetoch s rovnakou distribúciou, lenže znižuje …more
Abstract:
Large Language Models based on Transformer architecture hold state-of-the-art in a majority of Natural Language Modeling tasks. Nevertheless, these models tend to learn biases from training dataset, which help them on the training dataset, but hurt their out-of-domain accuracy as a result. This work focuses on obtaining more robust BERT models for the Extractive Question Answering task. We explore …more
 
 
Language used: English
Date on which the thesis was submitted / produced: 17. 5. 2022

Thesis defence

  • Date of defence: 22. 6. 2022
  • Supervisor: Mgr. Michal Štefánik
  • Reader: RNDr. Vít Suchomel, Ph.D.

Citation record

Full text of thesis

Contents of on-line thesis archive
Published in Theses:
  • světu
Other ways of accessing the text
Institution archiving the thesis and making it accessible: Masarykova univerzita, Fakulta informatiky

Masaryk University

Faculty of Informatics

Master programme / field:
Artificial intelligence and data processing / Big data