Are Large Language Models Good Data Preprocessors?

TL;DR

This paper evaluates the effectiveness of various large language models in cleaning and refining noisy image caption data, assessing their impact on downstream multimodal tasks.

Contribution

It provides an empirical comparison of multiple LLMs for data preprocessing, highlighting their potential and limitations in improving noisy textual data for complex tasks.

Findings

LLMs can improve data quality but with limited statistical significance

Effectiveness varies depending on dataset complexity and noise level

Further research needed to optimize LLMs for data cleaning

Abstract

High-quality textual training data is essential for the success of multimodal data processing tasks, yet outputs from image captioning models like BLIP and GIT often contain errors and anomalies that are difficult to rectify using rule-based methods. While recent work addressing this issue has predominantly focused on using GPT models for data preprocessing on relatively simple public datasets, there is a need to explore a broader range of Large Language Models (LLMs) and tackle more challenging and diverse datasets. In this study, we investigate the use of multiple LLMs, including LLaMA 3.1 70B, GPT-4 Turbo, and Sonnet 3.5 v2, to refine and clean the textual outputs of BLIP and GIT. We assess the impact of LLM-assisted data cleaning by comparing downstream-task (SemEval 2024 Subtask "Multilabel Persuasion Detection in Memes") models trained on cleaned versus non-cleaned data. While our experimental results show improvements when using LLM-cleaned captions, statistical tests reveal that most of these improvements are not significant. This suggests that while LLMs have the potential to enhance data cleaning and repairing, their effectiveness may be limited depending on the context they are applied to, the complexity of the task, and the level of noise in the text. Our findings highlight the need for further research into the capabilities and limitations of LLMs in data preprocessing pipelines, especially when dealing with challenging datasets, contributing empirical evidence to the ongoing discussion about integrating LLMs into data preprocessing pipelines.