- Services & Solutions
- Clients & Partners
Quadri: Intelligent Question Answering in the HIV Domain
SRI, Stanford University, and PARC co-developed an intelligent system for answering English questions about HIV treatment.
Clinical researchers and practitioners frequently need to access information from diverse data sources whose structure and interface may be unfamiliar. To facilitate this, SRI, Stanford, and PARC collaborated on the prototype Quadri, which uses a natural English interface and subject-domain reasoning to reveal clinical information.
Although many people understand English, a computer has difficulty because of ambiguities in the question, discrepancies in vocabulary between the user and the available sources, and heterogeneity in the sources. Further, the answer to the question may not be present explicitly in any one source, but may need to be deduced from information distributed among many sources. If the machine has an understanding of the subject domain of the query and the capabilities of the various sources, these difficulties can be mitigated.
By restricting attention to a particular topic area, domain knowledge may be brought to bear in understanding the question and constructing the appropriate answers. Although the approach can be applied to any area, the team selected HIV drug treatment as its initial subject domain.
Here is how Quadri is used:
- An English question is parsed and translated into a logical form by PARC’s natural language technology with domain-specific transformations.
- The logical form is expanded and decomposed by the SRI AI Center’s deductive machinery, informed by an axiomatic theory of the subject domain and an HIV ontology developed with the Stanford Center for BioMedical Informatics Research.
- HIV databases, principally those developed by the Stanford Division of Infectious Diseases, are invoked.
- An answer, with English explanation and visualization, is produced.
Quadri can deal with a sequence of questions that can refer back to the statements and results of previous questions.
The project was funded under an NIH Challenge Grant, Award Number RC1LM010583 from the National Library of Medicine. This content is solely the responsibility of the authors and does not necessarily represent the official views of the National Library of Medicine or the National Institutes of Health.