Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Ensembling for Entity Linking in the Medical Domain

IP.com Disclosure Number: IPCOM000249596D
Publication Date: 2017-Mar-07
Document File: 1 page(s) / 31K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method to leverage the existing entity linking systems such that the resulting ensemble is better than each individual entity-linking component.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 73% of the total text.

1

Ensembling for Entity Linking in the Medical Domain

Entity linking is the task of associating a text mention (i.e., span of text) with one or more existing entities in a Knowledge Base (KB). Multiple diverse systems are available for performing this task in various domains, such as medical. Each system takes as input the same span of text but can produce associations with different entities in the KB. The associations between the text mention and the KB entities vary in quality. Some systems have better precision, while some systems have better recall. Moreover, some systems can have better performance with certain types of entities.

The problem addressed herein is ensembling for entity linking. Ensembling is a technique for automatically identifying the strengths and weaknesses of component systems and combining those to isolate high performance of each.

The novel contribution is a method to leverage the existing entity linking systems such that the resulting ensemble is better than each individual entity-linking component.

The components and process for implementing the method follow: 1. Given: a domain specific knowledge base (KB) 2. Given: a mention (span) in a textual context 3. Given: N entity linking systems where each system associates the mention with one or more entities in the KB,

optionally producing an association score 4. Given: a training set of mention-to-entity pairs that are correctly linked (positive training examples) and

mention-to-entity pairs that ar...