Izboljšava obravnave negacije v velikih jezikovnih modelih

Kranjec, Matej (author), Robnik Šikonja, Marko (mentor)

V diplomski nalogi smo preizkusili metodo za izboljšavo klasifikacije globokih nevronskih mrež s predznanjem o negaciji. Najuspešnejši jezikovni modeli, kot na primer BERT ali ELMo, so uspešni pri klasifikaciji besedil, a odpovejo pri negaciji. Prednaučene jezikovne modele smo prilagodili, da tudi v slovenščini bolje delujejo z negacijo. To smo dosegli z spreminjanjem funkcije izgube nevronske mreže ter prilagajanjem obstoječih modelov. Metodo smo preizkusili na prilagojenem korpusu z dodanimi negacijami osnovnih stavkov. Metoda je uspešno zmanjšala delež napačnih napovedi v negiranih stavkih pri maskiranem jezikovnem modelu, točnost na nalogah iz slovenske zbirke SuperGLUE pa je ponekod izboljšala, drugje pa poslabšala.

• Language: Slovenian
• Keywords: globoke nevronske mreže, klasifikacija, obravnava negacije, veliki vnaprej naučeni jezikovni modeli
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-139701
• COBISS.SI-ID: 121785859
• Publication date in RUL: 06.09.2022

KRANJEC, Matej, 2022, Izboljšava obravnave negacije v velikih jezikovnih modelih [online]. Bachelor’s thesis. [Accessed 17 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=139701

Secondary language

• Language: English
• Title: Improving negation handling in large language models

Abstract

In the thesis we have tested a method for improved classification of deep neural networks with prior knowledge of negation. State of the art language models, such as ELMo and BERT, are successful at text classification, but fail when there is negation involved. We adjusted pre-trained language models to work better with negation in Slovene. We modified the loss function of the neural networks and retrained the models. We have tested the method on a modified corpus with added negations of original sentences. The method successfully reduced the error in the negated sentences for masked language models, and it increased the accuracy for some tasks from the Slovene version of the SuperGLUE benchmark but decreased for others.

• Keywords: deep neural networks, classification, negation modeling, large pretrained language models