ARAGORN: Autonomous Relay Agent for Generation Of Ranked Networks

Univ Of North Carolina Chapel Hill, Chapel Hill NC

Investigators

Linked publications, trials & patents

Abstract

We propose an Autonomous Relay Agent for Generation of Ranked Networks (ARAGORN), which will query Knowledge Providers (KPs) and synthesize answers relevant to user-specified questions, building upon algorithms and components developed as part of the ROBOKOP [1,2] application during the feasibility phase of Translator. The ARAGORN services represent the next generation of ROBOKOP component services, iterating and innovating in response to challenges exposed in the Translator feasibility phase. Based on that work, we have identified overarching issues that must be addressed to truly unleash the power of Translator. 1. ARAs must be able to operate in a federated knowledge environment effectively and efficiently. First-generation Translator tools assembled full data sets from which to extract answers, which were subsequently ranked. Second-generation tools must be able to efficiently operate on massive, distributed data, demanding a new approach. ARAGORN will act asynchronously, interleaving KP queries with partial scoring of answers, prioritizing search directions on-the-fly, and delivering early results that are updated over time in response to newly explored paths 2. ARAs must bridge the precision mismatch between data representations and algorithms that require specificity, and users who pose questions and prefer answers at a more abstract level. Biomedical scientists do not pose questions as database queries. Furthermore, even expert users of current biomedical databases such as ROBOKOP KG or RTX require exploration and experimentation to craft queries to express their intent. ARAGORN will employ multiple strategies to remove this barrier to asking questions effectively, from basic maintenance of a question library, to node generalization, query rewriting, and machine learning techniques such as capsule graph networks. ARAGORN will further use elements of specific answers to create gestalt explanations, clustering, and combining answers with similar content, revealing the commonalities and contradictions in answers. 3. ARAs must be able to generalize answer ranking to address a broader range of question formulations and data types, and to account for counterevidence. In the Translator implementation phase, we anticipate having access to many varied KPs and ARAs that provide diverse quantitative metadata regarding the confidence in assertions or strength of associations. There will be a pressing need to synthesize such data into scores for arbitrarily-shaped answer graphs, in order to filter and prioritize answers for further analysis or user digestion. ARAGORN will address this need by providing a novel scoring algorithm capable of (a) scoring arbitrary directed multi-hypergraphs, (b) accounting for heterogeneous quantitative metadata; and (c) leveraging relationship polarity to incorporate counterevidence. ARAGORN will provide access to this functionality, and connect to KPs using community-defined APIs and data models. The ARAGORN team has contributed to these community efforts during the Translator feasibility phase, and if funded will continue to work with the Standards and Reference Implementations (SRI) group, NCATS staff, and other awardees to continue to define and refine methods and models for data sharing and collaboration. The ARAGORN services will be created with collaboration in mind, such that they can be plugged into larger pipelining and architectural efforts. Most of the risks to the ARAGORN strategy are shared by the entire program; as standardization evolves, the ARAGORN team and other members of the Translator consortium will be required to spend effort updating components. ARAGORN will require access to ontology and similarity tools that we anticipate will be provided by KPs or shared infrastructure; if these do not materialize, the ARAGORN team will create the tools that it needs to accomplish its goals. Additionally, we are assuming the existence of fully translator-compliant KPs from which to draw data; if the program collectively decides that compliance is enforced in ARAs instead, we will draw on our work in ROBOKOP to implement the necessary normalization components in ARAGORN.