Product Entity Matching by Leveraging Tabular Data

Abstract

Product Entity Matching (PEM) is a challenging subfield of record linkage that involves linking records referring to the same real-world product. Despite recent transformer models showing near-perfect performance scores on various datasets, they struggle the most when dealing with PEM datasets. In this thesis, we study PEM under the common setting where the information is spread over text and tables. To facilitate our research, we introduce a new dataset based on existing Amazon, Walmart and Google datasets, where each product contains a mix of both textual and tabular details. Our hypothesis is that leveraging detail tables alongside textual data can effectively tackle complex entity matching tasks where textual information alone falls short. However, existing models have proven to be inefficient and ineffective in utilizing such tabular data. We propose TATEM and TATTOO models, which offer an effective solution by harnessing pre-trained language models along with a novel serialization technique to encode tabular product data. Our models incorporate a novel attribute ranking module to make our model more data-efficient. Our experiments on both current benchmark datasets and our proposed datasets show significant improvements compared to state-of-the-art methods, including large language models in zero-shot and few-shot settings. Moreover, in out-of-domain and few-shot experiments, the TATTOO model showcases its superiority by outperforming strong baselines by a substantial margin.