TabVec: Table Vectors for Classification of Web Tables

Abstract

There are hundreds of millions of tables in Web pages that contain useful information for many applications. Leveraging data within these tables is di cult because of the wide variety of structures, formats and data encoded in these tables. TabVec is an unsupervised method to embed tables into a vector space to support classi cation of tables into categories (entity, relational, matrix, list, and nondata) with minimal user intervention. TabVec deploys syntax and semantics of table cells, and embeds the structure of tables in a table vector space. This enables superior classi cation of tables even in the absence of domain annotations. Our evaluations in four real world domains show that TabVec improves classi cation accuracy by more than 20% compared to three state of the art systems, and that those systems require signi cant in domain training to achieve good results.

A framework for information extraction from tables in biomedical literature

“…By using cost-sensitive classification and assign larger costs to positive class than to negative, it is possible to improve these results. We managed to improve the F1-scores by almost 10% (see Tables 11,12,13).…”

“…Vector representations that would involve both context and structure of the article element should be explored in the future. The first attempt to do so was performed by [13]; however, this approach is limited to classification of table clusters, with no relation to other tasks where vector representation may be helpful. Also, the performance of information extraction using recurrent neural networks in combination with the mentioned representation model should be further explored in the future.…”

Automated Mining of Leaderboards for Empirical AI Research

“…Milosevic et al [39] tested methods for extracting numerical (number of patients, age, gender distribution) and textual (adverse reactions) information from tables identified by the < table > tag in the clinical literature as XML articles. Further, another line of work examined the classification of tables from HTML pages as entity, relational, matrix, list, and nondata leveraging specialized table embeddings called TabVec [21]. Wei et al [46] defined a question answering task with data in Table cells as the answers over two different datasets, i.e.…”