Large Language Models as Data Preprocessors

In the realm of artificial intelligence, Large Language Models (LLMs), like OpenAI's GPT series and Meta's LLaMA variants, have been making significant strides. These models, trained on vast amounts of text data, are capable of understanding and generating human-like text across a myriad of topics. A recent study by Haochen Zhang, Yuyang Dong, Chuan Xiao, and Masafumi Oyamada has delved into further applications of LLMs, specifically in data preprocessing - a crucial stage in data mining and analytics applications.

The researchers explored the applicability of state-of-the-art LLMs such as GPT-3.5, GPT-4, and Vicuna-13B for tasks like error detection, data imputation, schema matching, and entity matching. To enhance the performance and efficiency of these models, they proposed an LLM-based framework for data preprocessing, which integrates advanced prompt engineering techniques with traditional methods like contextualization and feature selection.

The effectiveness of LLMs in data preprocessing was evaluated through an experimental study involving 12 datasets. GPT-4 stood out, achieving 100% accuracy or F1 score on 4 datasets, indicating the immense potential of LLMs in these tasks. However, the study also highlighted certain limitations of LLMs, particularly their computational expense and inefficiency.

Despite these challenges, the study underscores the promise of LLMs in this domain and anticipates future developments to overcome current hurdles. This research not only widens the horizon of LLM applications but also provides a roadmap for future research and development in this area, by integrating LLMs with traditional data preprocessing techniques.

Read the whole article here: http://arxiv.org/abs/2308.16361v1