This repository is a grant project supported by https://github.com/OriginTrail/ChatDKG. It contains source code for AI-powered GPT (copilot) for scientific research that combines the OriginTrail Decentralized Knowledge Graph with LLMs. The copilot is designed to help retrieve information on scientific papers, analyze their results, and manage citations.
The copilot should allows for more effective and efficient research by providing access to information that could be verified and analyzed utilizing the OriginTrail Decentralized Knowledge Graph (DKG). Key benefits of this copilot include ensuring data verifiability and trust by using the DKG, as well as performing machine learning algorithms over trusted data.
In the sections below, there are various use cases and further research opportunities for the copilot.
Example: https://chat.openai.com/share/cbfb7da5-1afa-4fbe-8d71-3c8c75a8806e
The copilot assists in creating a comprehensive and up-to-date literature review. It analyzes the user's research topic and abstracts retrieved from the DKG. Then, it summarizes key findings, highlights trends, and identifies gaps in current research. Additionally, it helps researchers generate new hypotheses based on existing literature.
Example: https://chat.openai.com/share/2c223cc6-9b8c-4623-a562-083368b4bd3c
Copilot can use clustering algorithms to analyze trends in scientific research and group them based on vectors embedding similarity. This can help in identifying emerging fields, popular methodologies, or collaboration patterns. Use topic modeling to identify under-researched areas within a specific field. This can guide researchers towards topics that are ripe for exploration.
Copilot uses clustering algorithms to analyze trends in scientific research and groups them based on abstracts embedding similarity. This helps in identifying emerging fields and under-researched areas within a specific field. This guides researchers towards topics that have wide range of opportunities for further research.
Example: https://chat.openai.com/share/06844f72-fa7f-4b39-a174-7835b50c7449
Copilot creates a model that predicts the future citation count of papers or fields, which can be an indicator of their impact and relevance in the field. This serves as a valuable indicator of the potential impact and relevance of these papers or fields within the academic landscape. By forecasting citation counts, researchers gain valuable insights into the long-term influence and significance of their/other work.
The copilot could help researchers identify potential collaborators. It could analyze authorship data from papers in the database to suggest researchers with similar interests or complementary expertise.
The copilot could assist researchers in identifying potential funding sources. By analyzing the topics, content, and results of their work, the copilot can suggest relevant grants, scholarships, or funding parties.
Copilot could perform sentiment analysis on citations and references to evaluate the impact of research papers in the academic community. This means it could assess the overall sentiment associated with citations and references, and it could also track how the sentiment around a paper changes over time, giving researchers valuable insights into how their work is perceived as industry evolves.
A detailed project specification with architecture and sequence diagrams is available here.
The project has been executed in two phases: beta and release version. The beta version includes setting up the environment and verifying the approach with Semantic Scholar data and the ScholarlyArticle ontology, without AI Cortex. The beta version will be released and made public for early feedback.
The release version will improve accuracy, introduce AI Cortex, and update the DKG with ArXiv data and ontology.
Upon the completion of this project, a list of findings will be presented and the copilot will be available at the OpenAI marketplace.
The next steps involve implementing the future use cases mentioned above. This includes retrieving blockchain proofs for assets along with the requested data. As a fallback option, we plan to introduce an embedding-based RAG if generated sparql query is incorrect.
Additionally, our goal is to test smaller models and refine them using reinforcement learning based on human feedback to achieve better results. Also, we want to test the OT World API, and integrate the Twitter API with the DKG copilot.
Document version: 1.0.0-beta
Last changed: 26th Dec 2023