StruBERT: The Power of Combining Textual and Structural Information for Table Retrieval and Classification
In today's world of big data, tables are often used to store a vast amount of information. Retrieval of such data tables has always been of utmost importance, especially in cases where users want to find tables that are relevant to their queries. However, previous methods only treated each source of information independently. This resulted in the neglect of the essential connection between textual and structural information when it comes to table retrieval. To improve the accuracy of data table retrieval, the researchers have developed StruBERT, a structure-aware BERT model that fuses the contextual and structural information of a data table to produce context-aware representations for both textual and tabular content of a data table.
Understanding Data Tables
Data tables are used to store structured information in an organized manner. Rows and columns of a table form the primary structural information. These rows and columns comprise data values that the table stores information about. However, textual information related to the table, such as a caption, page title, etc., also play an imperative role in understanding the table better. Thus, understanding the relationship between the textual and structural information is essential when retrieving and classifying data tables.
The Need for StruBERT
Previous methods for retrieving data tables only considered textual information or structural information when retrieving the data table. This led to the neglect of the connection between these two aspects, leading to counterparts coupling irrelevant data tables.
With its structure-aware BERT model, StruBERT solves the above problem by fusing both the textual and structural information of a data table. It produces context-aware representations of both textual and tabular content of a data table. This allows the system to understand and retrieve the complex relationship between the textual content and the structural organization of the data table.
The Advancement Delivered by StruBERT
StruBERT provides numerous advantages over its predecessors. Firstly, it improves the accuracy of text-based and structure-based table retrieval up to 20% by fusing both the textual and structural information. Secondly, StruBERT can efficiently retrieve data tables similar to the one in question, improving table similarity up to 30%. Lastly, StruBERT provides a tool for classifying data tables without supervision, making it useful for databases with a large amount of unlabeled data tables.
StruBERT delivers a state-of-the-art data table retrieval and classification solution. By combining textual and structural information, StruBERT delivers higher accuracy, improved efficiency, and better retrieval of data tables. StruBERT provides an optimal solution for people who want to retrieve data tables and understand the relationship between textual and structural information.
