Host institution: Aalborg University (AAU), Copenhagen 🇩🇰
Supervisors: Prof. Jannick Sørensen (AAU) and Prof. Tim Raats (VUB)
Academic secondment: Vrije Universiteit Brussel (VUB) 🇧🇪
Industrial secondment: European Broadcasting Union (EBU), Geneva 🇨🇭
PhD duration: 3 years
This PhD project is titled: Quality news both for Public Service Media and is part of WP4 Algorithms and value-based curation.
AI-powered chatbots have become a standard feature for all types of services but, when used in quality-oriented Public Interest Media, the correctness of the answers given by the chatbot is pivotal. The reputation of Public Interest Media is based on unbiased, quality-checked trustworthiness. The European Broadcasting Union (EBU) in Geneva has developed, for its Public Service Media members, a state-of-the-Art RAG-based conversational chatbot: NEO, which is based on four million news articles, growing daily by 3000. NEO can be used both by journalists and end-users.
NEO is a very relevant case for this PhD position not only as the quality of its replies must adhere to the highest journalistic standard, but also because it must integrate in the editorial profile of the news organisations where it is implemented. The current version of NEO only searches articles published by public media journalists but if a ‘deep search’ mode should be implemented – to provide better answers on more complex questions – external trusted sources must be carefully integrated and answers must be automatically quality-controlled, e.g., via a novel ‘agentic workflow’ to be developed in the PhD project where elements of information in the chatbot answer are examined by a network of independent software agents. How to prevent NEO from using AI-generated content? In short: How to ensure the highest journalistic quality in the output from NEO?
Another possible topic for the PhD candidate is designing NEO for different user needs. Journalists have different needs, so have end-users. Furthermore, the use context for NEO as a plug-in for social media is radically different from NEO implemented within a news website: How to speak to the user, which pictures to show? Should the style of NEO’s answers only satisfy user needs (e.g.: Entertain me!) or should NEO also try to nudge peoples’ fields of interest?
A third possible topic is performance improvement of using a graph-based analysis and/or infrastructure. Typical RAG systems use a semantic search based on embeddings. NEO uses a hybrid approach with both embeddings-based and keyword-based retrieval. While the hybrid approach has clearly shown its supremacy, retrieval could still be better. A graph-based storage of the article content (or a sample of it) could better model the complex links between people, facts, organisations, and concepts, to name a few, described in the articles. On simple questions like ‘what are the latest (major) news today?’ or ‘give me a timeline of the COVID pandemic’, NEO might struggle. A graph-based approach might help to identify the broad concepts first, and then move on to the details.