A semantic-based workflow for biomedical literature annotation

Abstract

Computational annotation of textual information has taken on an important role in knowledge extraction from the biomedical literature, since most of the relevant information from scientific findings is still maintained in text format. In this endeavour, annotation tools can assist in the identification of biomedical concepts and their relationships, providing faster reading and curation processes, with reduced costs. However, the separate usage of distinct annotation systems results in highly heterogeneous data, as it is difficult to efficiently combine and exchange this valuable asset. Moreover, despite the existence of several annotation formats, there is no unified way to integrate miscellaneous annotation outcomes into a reusable, sharable and searchable structure. Taking up this challenge, we present a modular architecture for textual information integration using semantic web features and services. The solution described allows the migration of curation data into a common model, providing a suitable transition process in which multiple annotation data can be integrated and enriched, with the possibility of being shared, compared and reused across semantic knowledge bases.

Figure 1.

Semantic-based architecture for scientific information integration.

Figure 2.

Annotation model: sample extraction of the integration and representation of an annotation related to the ‘Alzheimer disease’.

Figure 3.

Relation model: sample extraction of the integration and representation of a relatedTo annotation relationship.

Figure 4.

Validation workflow overview. (1) Dataset is extracted from the NCBI database. (2) Neji and cTAKES API services were used for information extraction, generating diverse outputs and formats. Additional annotation services can be used. (3) Annotations are forwarded and integrated into a unified model and stored in an accessible knowledge base.

Figure 5.

Knowledge base sample annotation model. The annotators involved share concept attributions (i.e. umls: C0027061), increasing the likelihood of being correctly identified.