Leveraging Structured and Unstructured Data for Tabular Data Cleaning

The quality of data is essential for the success of any data pipeline, as even small inconsistencies can lead to significant errors in downstream tasks. However, real-world data is often incomplete and noisy, requiring effective data cleaning strategies to ensure the accuracy of downstream operations consuming this data (e.g., training ML models, visualization). A large class of data-cleaning systems relies on user input and/or predefined integrity constraints (e.g., Functional Dependencies) to detect and repair errors in tables. As a result, user expertise is usually needed to specify integrity constraints or provide curated training datasets used to build data cleaning models for specific datasets and use cases. Furthermore, existing approaches do not leverage other data modalities (e.g., text documents), which may contain a wealth of information that helps fix erroneous cells in input tables. In light of these limitations, we propose our work-in-progress system, Beaver, that can clean input tables by learning "good repairs" from a large collection of tables and text documents that are used to fine-tune a Large Language Model (LLM). Unlike traditional methods, which rely on user expertise, Beaver benefits from the capabilities of LLMs to identify and learn complex patterns and relationships between different attribute values from structured and unstructured data. Beaver repairs an input table by performing the following steps: (1) Generating question and answer pairs based on the input text documents and tables provided by the user. (2) Fine-tuning the LLM model (e.g., T5) using these generated question-and-answer pairs. (3) Once the model has been trained, Beaver detects erroneous cells in an input dirty table. (4) Invoking the fine-tuned LLM to generate predictions on the erroneous cell values in the input table.