Introduction

The rapid expansion of scientific literature necessitates a reevaluation of their information retrieval (IR) engines [1,2]. Traditional keyword-based approaches are inadequate for tracking fast-paced scientific advancements. There is a growing demand for structured scientific content representations [3,4] and advanced machine learning algorithms [5,6] to enhance retrieval accuracy. Initiatives like the Open Research Knowledge Graph (ORKG) [7] drive this paradigm shift towards structured knowledge representations, enabling intelligent views and comparisons of research facets [8,9]. Our goal is to simplify access to scientific articles and reduce cognitive load for researchers using information technology (IT). We propose dashboards as visual tools to represent structured scientific knowledge, enhancing research filtering and discovery processes [10]. Dashboards have been widely used, including during the Covid-19 pandemic, where they helped track cases, analyze trends, and support decision-making with data from sources like the WHO and Johns Hopkins [11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36]. In contrast, our approach focuses on applying IT to structure scientific knowledge itself, using information extraction (IE) mechanisms and large language models (LLMs) to power next-generation information systems.

In pursuit of our vision, this poster paper presents a proof of concept (POC) using the ORKG-R0 semantic model [37] to structure articles on the "reproductive number estimate of infectious diseases" theme [38]. The model captures essential properties like disease name, study location, date, 𝑅0 value, % confidence interval values, and computation method, enabling effective comparison across studies. Four research questions (RQs) guide article search and exploration: RQ1 identifies maximum 𝑅0 estimates, RQ2 examines study counts by disease and location, RQ3 analyzes 𝑅0 value ranges by location for selected diseases, and RQ4 maps study locations globally on the world map. These RQs are visualized in a dashboard to enhance article filtering and provide researchers with concise insights into research progress.

Next-Generation Scientific Information Retrieval (IR)

We introduce a next-generation IR platform for "reproductive number estimates of infectious diseases, " enhancing scientific article access with IT and four visual charted summaries tailored to four specific RQs alluded to earlier. In the following subsections, we will detail the LLM-based IE method, article collection, and platform workflow.

The Scientific Information Extraction (IE) Large Language Model (LLM)

We employ the ORKG-FLAN-T5 R0 LLM [39]. This model is an instruction fine-tuned variant of FLAN-T5 Large (780 M) using the instruction-tuning paradigm introduced as FLAN (Finetuned Language Net) [40,41,42,43]. It processes a paper's title and abstract to produce structured summaries based on six key properties: disease name, location, date, 𝑅0 value, % confidence interval (CI) values, and method, related to the 𝑅0 estimate [39].

Table 1

The top 20 infectious disease names (and number of papers) in our initial dataset. covid-19 (1002) mers-cov (21) measles (15) hepatitis c (8) dengue (41) cholera (18) hepatitis b ( 12) tuberculosis ( 8) influenza (29) zika (18) zika virus ( 12) monkeypox ( 8) hiv (23) african swine fever ( 17) ebola (11) west nile virus (7) sars (22) ebola (17) hand, foot, and mouth disease (8) malaria ( 7)

The Scholarly Articles Collection

The targeting papers with any synonyms of 𝑅0 in the title or abstract. This yielded 7,127 articles. We leveraged the ORKG-FLAN-T5 R0 LLM [39] to filter articles that did not report an 𝑅0 value as unanswerable; or otherwise provide structured JSON descriptions for articles with 𝑅0 estimates. After filtering out unanswerables, 2,051 articles remained, yielding 2,736 structured summaries. The processed data was imported to a PostgreSQL 16 database, serving as the backend storage. The top 20 most represented infectious diseases in our initial database is shown in Table 1. Notably, the LLM's high precision confirmed that the top reported diseases are indeed ascertained infectious diseases. Our database covers studies from all seven continents.

The Information Retrieval (IR) Platform Workflow

The platform is accessible as a web application at the following URL: https://orkg.org/usecases/ r0-estimates. The visualization dashboard widget and underlying workflow are displayed in Figure 1. In this workflow, the frontend communicates with the backend through a Web API for database queries and data retrieval. A Python script scheduler, programmed to run monthly, periodically updates the database with new articles querying PubMed and following the LLM processing cycle before updating the database with structured summaries. Our workflow maximizes the use of cutting-edge technology, including an optimized next-generation LLM.

Charting the data: collating, summarizing, and reporting

Our IR platform includes three main components: 1) a statistics snapshot showing total papers, structured knowledge, infectious diseases, and locations, 2) a standard paper listing in a keywordbased table, filtered as needed, built with the ag-grid JavaScript library, and 3) a visual analytical dashboard with four charts addressing our research questions. This process involves collating relevant properties, selecting the best chart from the React chart library to summarize the response, and creating a query to report the visual summary. Each RQ is represented by a visual chart. E.g., RQ1, "What are the maximum 𝑅0 estimates reported for diseases in our database?" is illustrated with a bar chart that plots diseases on the x-axis against their maximum 𝑅0 values on the y-axis. Hovering over a bar displays the disease and its max 𝑅0. This interactive chart, which can be adjusted for specific 𝑅0 ranges, simplifies the comparison of 𝑅0 estimates across numerous studies. Clicking on a bar provides a direct link to the contributing article on PubMed, thereby enhancing scholarly information retrieval significantly beyond traditional methods.

Conclusion

In this poster paper, we present a POC for a new scholarly IR engine that enhances access and reduces the cognitive load of traditional, keyword-based searches. We address the inefficiencies of manual paper filtering in traditional IR systems, exacerbated by rapidly increasing publication volumes. Our approach models key research aspects for machine processing, paving the way for next-generation visual assistants that streamline scholarly research access.