Using Large Language Models for Aspect Based Sentiment Analysis
Author
Baral, Sadiksha
Term
4. semester
Publication year
2024
Abstract
This thesis examines how Large Language Models (LLMs) - advanced AI systems that work with text - can support aspect-based sentiment analysis (ABSA), a method that breaks a text into topics (aspects) and judges the sentiment for each. The work starts with traditional machine learning and progresses to advanced deep learning models. We evaluate models on three tasks: category classification (which aspect is being discussed), sentiment polarity (whether the tone is positive, negative, or neutral), and a joint task that combines both. Fine-tuning OpenAI's GPT-3.5 Turbo led to clear performance gains, highlighting the benefits of adapting a model to the task. Despite these advances, the study identifies limits around data availability, high computational demands, and model hallucinations (producing plausible but incorrect content). It recommends expanding datasets, optimizing model architectures, and exploring more advanced fine-tuning techniques to improve the robustness and accuracy of LLMs for ABSA in real-world applications.
Dette speciale undersøger, hvordan store sprogmodeller (Large Language Models, LLM'er) kan bruges til aspektbaseret sentimentanalyse (ABSA) - en metode, der opdeler en tekst i konkrete emner (aspekter) og vurderer holdningen til hvert af dem. Arbejdet begynder med traditionelle maskinlæringsmetoder og bevæger sig videre til avancerede deep learning-modeller. Modellerne vurderes på tre opgaver: kategoriklassifikation (hvilket aspekt omtales), sentimentpolaritet (om tonen er positiv, negativ eller neutral) og en fælles opgave, der kombinerer begge dele. Finjustering af OpenAI's GPT-3.5 Turbo gav markante forbedringer i resultaterne og viser værdien af at tilpasse modellen til opgaven. På trods af fremskridt peger studiet på begrænsninger som begrænset datatilgængelighed, store krav til beregningsressourcer og hallucinationer (når modellen opfinder plausible men forkerte oplysninger). Fremtidigt arbejde bør udvide datasæt, optimere modelarkitekturer og undersøge mere avancerede finjusteringsteknikker for at gøre LLM'er til ABSA mere robuste og præcise i brug i virkelige sammenhænge.
[This apstract has been rewritten with the help of AI based on the project's original abstract]