This is a follow up project based on SCDs.
Intelligent Subjective Agents
Subjective Content Descriptions (SCDs) are a solid foundation for NLP based AI systems. The vision of an Intelligent Subjective Agent (ISA) is to use SCDs for bringing subjectivity into, e.g., state-of-the-art ChatBot-Systems using LLMs. In comparison to Retrieval Augmented Generation (RAG), SCDs provide additional subjective context to prompts, documents, and results. SCDs can also be used to create subjective comments relevant for the principals of an agent.
Principals shall be able to provide their personal SCDs to an ISA and get results aligned to the provided SCDs. Furthermore, different principals are able to exchange their SCDs and get results with different points of view on the same prompt, because different SCDs represent different persons and views.
Recap of SCDs
See Project: SCDs for more details.
SCDs are like comments attached with text documents in the corpus. These comments might have no direct grounding in the base text, but indeed involve the context of principals. If comments are formulated in natural language, a human principal can easily be pointed to relevant documents and, in those documents, to paragraphs or sentences that are relevant for the context, e.g., a new scientific problem the human is working on.
The comments attached with text documents are considered to be subjective because it is the context, not only the base text, that matters for the interpretation of comments. Thus, we call the above-mentioned comments SCDs.
SCDs provide a framework to model perceptions humans have when working with corpora of text documents. These perceptions are subjective to each human and an intelligent artificial system is required to anticipate such perceptions of principals. SCDs can be used together with LLMs, e.g., for contributing additional context, mitigating hallucinations, or attaching citations.
Current Results
Enhancement of SCDs by using Human Feedback
The result is on automating enhancement of SCDs based on feedback from human users, i.e., the principals of an ISA. The paper presents an approach to decide how to incorporate the feedback and when to update the SCDs. Altogether, SCDs can be updated with human feedback, allowing users to create even more specific SCDs for their needs.
- Magnus Bender, Tanya Braun, Ralf Möller, Marcel Gehrke
Enhancement of Subjective Content Descriptions by using Human Feedback
to be published in International Journal of Semantic Computing, 2025
PDF ArXiv
Humanities Aligned ChatBot
The Humanities Aligned ChatBot (ChatHA) provides some of the mentioned features of an ISA. In the context of humanities, ChatHA does not only take subjectivity into account, but also sustainable data management in compliance with the FAIR principles: The aim is to provide a platform for scholars which makes the obtained, and often textual, research results permanently available for both human and machine processing. Then ChatHA enables the scholars to communicate with the(ir) data—subjectively in the sense of SCDs.
Related publications:
- Thomas Asselborn, Sylvia Melzer, Simon Schiff, Magnus Bender, Florian Andreas Marwitz, Said Aljoumani, Stefan Thiemann, Konrad Hirschler, Ralf Möller
Building Sustainable Information Systems and Transformer Models on Demand
to be published in Humanities and Social Sciences Communications, Nature, 2025 - Hagen Peukert, Lucas F. Voges, Thomas Asselborn, Magnus Bender, Ralf Möller, Sylvia Melzer
Humanities in the Center of Data Usability: Data Visualization in Institutional Research Repositories
in Proceedings of the Humanities-Centred AI (CHAI) Workshop at KI2024, 47th German Conference on Artificial Intelligence, 2024
DOI PDF URL - Thomas Asselborn, Sylvia Melzer, Said Aljoumani, Magnus Bender, Florian Andreas Marwitz, Konrad Hirschler and Ralf Möller
Fine-tuning BERT Models on Demand for Information Systems Explained Using Training Data from Pre-modern Arabic
in Proceedings of the Humanities-Centred AI (CHAI) Workshop at KI2023, 46th German Conference on Artificial Intelligence, 2023
DOI PDF URL