An Information Extraction Study: Take In Mind the Tokenization!

TL;DR

This study investigates how tokenization affects information extraction from biomedical texts, comparing subword and character-based models, and finds that tokenization patterns influence performance while token-free models are promising.

Contribution

It provides a comparative analysis of tokenization's impact on IE, highlighting the potential of token-free models in biomedical information extraction.

Findings

Tokenization patterns can introduce beneficial inductive bias.

Character-based models achieve promising results.

Token-free models are feasible for IE tasks.

Abstract

Current research on the advantages and trade-offs of using characters, instead of tokenized text, as input for deep learning models, has evolved substantially. New token-free models remove the traditional tokenization step; however, their efficiency remains unclear. Moreover, the effect of tokenization is relatively unexplored in sequence tagging tasks. To this end, we investigate the impact of tokenization when extracting information from documents and present a comparative study and analysis of subword-based and character-based models. Specifically, we study Information Extraction (IE) from biomedical texts. The main outcome is twofold: tokenization patterns can introduce inductive bias that results in state-of-the-art performance, and the character-based models produce promising results; thus, transitioning to token-free IE models is feasible.