Extracting Measured Properties for Numerical Data with SciBERT model and Question Answering

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2022-01-24

Department

Major/Subject

Human-Computer Interaction and Design

Mcode

SCI3020

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

50+0

Series

Abstract

Quantity is a measurement (e.g. 18g), which usually consists of numerical data and units. Quantity is crucial and very frequently mentioned in scientific publications. At Elsevier, there are good solutions for searching quantities or their components, like numerical data or units. In these products, you can search all the papers which contain, for example, ”< 2mm”, in their full text. However, what the observed measurements represent is still unclear. For example, when you search ”< 2mm”, does ”2mm” represent the length or diameter of a tube? The ambiguity causes many irrelevant results in their search engines. The property behind the quantity is called measured property. To solve this ambiguity and enhance the search capability, extracting what measured property a quantity represents is the next step of Elsevier. When users can search both quantity and measured property at the same time, they can definitely get more accurate results. In this paper, we propose a Question-Answering architecture for joint measured property and relationship extraction based on the numerical data extraction model. The Question-Answering architecture enables a named entity recognition model to extract entity and relationship jointly. We train a SciBERT model to extract quantity in the corpus and another SciBERT model to extract corresponding measured property for each quantity. Meanwhile, we annotate a dataset with the publications from the engineering domain, MeasPro, for our model training. It proves that our approach has excellent accuracy and it is better than the state-of-art models on MeasEval dataset.

Description

Supervisor

Theune, Mariët

Thesis advisor

Keulen, Maurice
Doornenbal, Marius

Keywords

NLP, entity extraction, relationship extraction, measurement, scientific publications

Other note

Citation