Sklepanje v naravnem jeziku s pomočjo zdravorazumske baze

Gospodarič, Jakob (author), Robnik Šikonja, Marko (mentor)

V tej diplomski nalogi bomo raziskovali izboljšanje zmožnosti zdravorazumskega sklepanja jezikovnih modelov ChatGPT in SloBERTa z integracijo zdravorazumskega znanja iz baze SloATOMIC. Na začetku bomo predstavili problemsko domeno in opisali področje. Nato bomo predstavili uporabljeno tehnologijo in postopek priprave podatkov. Primerjali bomo rezultate na podatkovni bazi SI-NLI brez in z dodatnimi stavki iz SloATOMIC. Dodatno bomo še primerjali rezultate na manjši podmnožici ročno popravljenih podatkov. Na koncu bomo opisali težave in predstavili možne nadaljnje izboljšave.

• Language: Slovenian
• Keywords: Zdravorazumsko sklepanje, jezikovni modeli, SloATOMIC, SI-NLI, grafi znanj
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-158181
• COBISS.SI-ID: 197359619
• Publication date in RUL: 28.05.2024

Secondary language

• Language: English
• Title: Natural language inference using commonsense reasoning database

Abstract

In this thesis, we will explore the enhancement of the commonsense reasoning capabilities of the language models ChatGPT and SloBERTa by integrating commonsense knowledge from the SloATOMIC database. We will begin by presenting the problem domain and describing the field. Then, we will introduce the technology used and the data preparation process. We will compare the results on the SI-NLI dataset with and without the additional sentences from SloATOMIC. Additionally, we will compare the results on a smaller subset of manually corrected data. Finally, we will describe the challenges encountered and present possible further improvements.

• Keywords: Commonsense reasoning, Language models, SloATOMIC, SINLI, Knowledge graphs