Speaker
Description
Scientific research publication abstracts are often highly advanced and primarily accessible to scientific communities. To extend accessibility and broaden the audience beyond this community, an audience-specific summarisation of abstracts called Tailored Abstract Translator (TAT) was implemented using open-source Large Language Models (LLMs).
This implementation makes scientific knowledge accessible to a wider group of users, thereby increasing the value and societal impact of the information. The module is open source and built on top of the InvenioRDM framework, making it suitable for adoption in scientific publication repositories.
The European Spallation Source (ESS) operates its own publication portal (https://publications.ess.eu/
), powered by the open-source InvenioRDM platform. As part of the HALRIC funding initiative, an LLM-based module has been implemented to generate simplified summaries of publication abstracts. This module is pluggable and can be configured and enabled within the existing publication portal. A scheduled daily job has been configured to process publications missing an abstract summary and generate new ones automatically. The LLM models are hosted on an internal Azure instance and consumed by these background jobs.
With the on-premise Kubernetes cluster, the implementation of these jobs is scalable, less complex, and provides improved observability. In the future, it is planned to introduce an administrative validation step, allowing generated summaries to be reviewed and, if necessary, regenerated through refined prompting. The quality of the summaries can further be improved through more targeted prompt engineering.
Two levels of summaries are currently generated: “Simple” and “Very Simple.” The “Very Simple” level provides highly accessible explanations, making it easier for non-specialist audiences to understand the abstract and encouraging further exploration of the subject matter.
