gTBLS: Generating Tables from Text by Conditional Question Answering

Anonymous

TL;DR: This paper proposes an approach to automatically generate tabular content from textual paragraphs.

Abstract: Distilling large, unstructured text into a structured, condensed form such as tables is an open research problem in Natural Language Processing. Prior approaches address this task through additional parameters in the Transformer's attention mechanism. This paper presents Generative Tables (gTBLS), a two-stage, parameter-efficient solution to automatically construct structured tables from text. The first stage infers table structure (row and column headers) from the text, and the second stage formulates questions and fine-tunes a causal language model to answer them. gTBLS improves prior approaches by up to 21% in BERTScore on the table content generation task of the E2E, WikiTableText, WikiBio, and Rotowire datasets with 66% fewer parameters.

Paper Type: short

Research Area: Information Extraction

Contribution Types: NLP engineering experiment, Data analysis

Languages Studied: English