Evaluating the Impact of Content Deletion on Tabular Data Similarity and Retrieval Using Contextual Word Embeddings

Table retrieval involves providing a ranked list of relevant tables in response to a search query. A critical aspect of this process is computing the similarity between tables. Recent Transformer-based language models have been effectively employed to generate word embedding representations of tables for assessing their semantic similarity. However, generating such representations for large tables comprising thousands or even millions of rows can be computationally intensive. This study presents the hypothesis that a significant portion of a table’s content (i.e., rows) can be removed without substantially impacting its word embedding representation, thereby reducing computational costs while maintaining system performance. To test this hypothesis, two distinct evaluations were conducted. Firstly, an intrinsic evaluation was carried out using two different datasets and five state-of-the-art contextual and not-contextual language models. The findings indicate that, for large tables, retaining just 5% of the content results in a word embedding representation that is 90% similar to the original one. Secondly, an extrinsic evaluation was performed to assess how three different reduction techniques proposed affects the overall performance of the table-based query retrieval system, as measured by MAP, precision, and nDCG. The results demonstrate that these techniques can not only decrease data volume but also improve the performance of the table retrieval system.