Domain Specific Artificial Intelligence for Small Dateset

ACL ARR 2025 February Submission2641 Authors

Abstract: In the field of economics, analyzing market news for commodities like oil is crucial for forecasting trends and making informed decisions. The sheer volume of news data requires efficient methods for sentiment analysis. This thesis explores the use of language models for sentiment analysis within the oil commodity market, focusing on extracting information related to price, supply, and demand dynamics from daily news. The study investigates the efficacy of zero-shot and few-shot learning, along with the use of adapters for continuous training, in both small and large language models. It is hypothesized that few-shot prompt engineering offers a cost-effective and efficient solution for sentiment analysis in this context. The research examines the performance of various models, including those trained on domain-specific datasets and those continuously trained with adapters. The findings contribute to developing more accurate and efficient tools for economic analysis and forecasting, while also considering the environmental impact of different techniques.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Small dataset, energy use, sentiment analysis

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 2641