The Problem: Organizations accumulate knowledge in reports, studies, and other documents, but understanding what that record means collectively requires people to connect it by hand. Findings relevant to the same question may use different terminology, come from different disciplines, or arrive years apart. Our work with the National Institutes of Health (NIH) addressed this problem in scientific research, where the volume and diversity of published findings make integration particularly difficult.
Our Approach: We created BRAINWORKS, a platform that organizes scientific literature and its context as an interactive knowledge graph; the platform achived this through a three-step process:
The Impact: BRAINWORKS gave researchers a way to examine connections across publications and relate them to the people and programs behind the work. We used the platform to investigate research trends and compare collaboration patterns across funded groups. The underlying tools provide a reusable approach to analyzing knowledge distributed across large document collections.
An organization may have extensive records of what its teams have learned and still struggle to answer questions that span those records. A relevant observation might appear in one report, an explanation in another, and supporting evidence in work produced years earlier. Connecting them requires familiarity with the documents and enough subject knowledge to recognize when different terms describe the same concept. Scientific research makes this difficulty particularly clear: around two million peer-reviewed papers were added in 2020 alone, to a literature already estimated at more than sixty million works. Studies examine related questions at different scales, and findings that could inform one another may be published in separate disciplines and described using different terminology. Searching for papers helps locate relevant material, but the reader still has to determine how their findings relate. The context matters as well; a research program needs to understand which work its funding supported, who collaborated on it, and how the resulting publications relate to the wider field. Answering these questions requires connecting the content of the literature to records of the activity that produced it.
We connected publications to their context, converted statements into structured relationships, and built interactive tools for exploring the resulting knowledge graph.
1. Connect the knowledge to its context: We built a common data foundation linking publications with their authors, institutions, grants, and citations. PubMed supplied publication records and abstracts, while NIH funding records connected papers to the projects that supported them. These links let a researcher move between a publication, the work it cited, and the people and programs responsible for it.
2. Turn statements into a connected knowledge graph: We developed a scalable language-processing pipeline that extracted statements from scientific abstracts as a subject, a relationship, and an object. We then matched the subject and object to concepts in the Unified Medical Language System (UMLS), which brings together biomedical terminology. This allowed different expressions for a recognized concept to be represented consistently across papers.
3. Explore the relationships and how they change: We built interactive visualizations that let researchers examine different relationships within the same collection. Citation networks connect papers, topic networks show which concepts appear together, and knowledge graphs display the relationships extracted from text. An individual-publication view lets the reader inspect a paper and its extracted statements. The views support searching, selecting related nodes, and examining connections at different levels of detail. We also developed analyses across time periods, allowing researchers to investigate how interest in a topic changed and how collaboration networks developed.
Our work made it possible to explore relationships across the literature and examine the research activity behind them.
Researchers could examine related concepts across papers, investigate the development of research topics, and connect publications to the programs that supported them. These capabilities brought together information that otherwise required separate searches and manual reconstruction. While our application focused on scientific research, the method applies wherever useful knowledge is spread across documents: connect the source records, represent their content consistently, and make the relationships available for people to examine.