Edgar J. Meij

This is the website of Edgar Meij. I lead several groups of researchers and engineers at Bloomberg, responsible for all AI Platforms at the company. Before this, I led all company-wide initiatives on improving search and discovery, including question answering and autocomplete using the Bloomberg Knowledge Graph together with its value-add machine learning-based Analytics. I am also a keynote speaker and senior scientist specializing in AI, LLMs, FinTech, Information Retrieval, Natural Language Processing, and Machine Learning and still active as senior program committee member for the main conferences in AI. I continue to be interested in advancing the state of the art in AI both in industry and at scale.

My research in the past has mainly focused on semantic search, namely, designing entity-oriented search systems that employ advanced NLP and machine learning techniques to improve user models, search, recommender systems, and content matching. Before Bloomberg, I was a research scientist at Yahoo Labs, working on various semantic search, query understanding, and recommender systems applications, employing knowledge graphs, query log analysis, machine learning, and distributed computing. Before that I did a post-doc at the Information and Language Processing Systems (ILPS) group of the Intelligent Systems Lab (ISLA) of the Informatics Institute of the University of Amsterdam. Research projects I have been involved there with include VL-e, CCCT, and Daeso. Most recently, I was working on DutchSemCor and LiMoSINe, two EU projects that center around semantic search, semantic annotations, and semantic information access. For the latter I was a workpackage leader.

In 2010 I finished my PhD under supervision of Maarten de Rijke. The topic of my PhD was leveraging conceptual knowledge from ontologies, thesauri, tags, annotations, or any other (structured) knowledge source to enhance information access. Information access – in this sense – entails retrieval and navigation of both documents and knowledge. To this end I am using statistical language modeling techniques, which are naturally capable of capturing language use and which I employ to bridge the semantic gap between (a priori defined) knowledge and (observed) language. ((Or, as Ludwig Wittgenstein said: “Meaning is use”. http://en.wikipedia.org/wiki/Ludwig_Wittgenstein)) Using this framework I am able to compare queries, documents, concepts, and relations on a conceptual level using language observations. More information can be found at http://phdthes.is/. In 2008 I spent some time in Barcelona, where I worked with Hugo Zaragoza and Peter Mika at Yahoo Labs. My research interests include, but are not limited by: (Semantic) Information Retrieval, the Semantic Web, Language Modeling, Big data, and Data and Text Mining.

I’m passionate about semantic search, information retrieval, search engines, semantic web, machine learning, information visualization, and mathematics and this website is my digital business card as well as my personal blog. I write on information retrieval, semantic web technologies, research in general, and, on occasion, stuff that doesn’t fit neatly into one of these categories. I also occasionally write about resources I discover or find interesting.

selected publications

A new era of AI-assisted journalism at Bloomberg

Claudia Quinonez, and Edgar Meij

AI Magazine, 2024

Abs Website

Artificial intelligence (AI) is impacting and has the potential to upend entire business models and structures. The adoption of such new technologies to support newsgathering processes is established practice for newsrooms. For AI specifically, we are seeing a new era of AI-assisted journalism emerge with trust in the AI-driven analyses and accuracy of results as core tenets. In Part I of this position paper, we discuss the contributions of six recently published research papers co-authored by Bloomberg’s Artificial Intelligence Engineering team that show the intricacies of training AI models for reliable newsgathering processes. The papers investigate (a) the creation of models for updated headline generation, showing that headline generation models benefit from access to the past state of the article, (b) sequentially controlled text generation, which is a novel task and we show that in general, more structured awareness results in higher control accuracy and grammatical coherence, (c) chart summarization, which looks into identifying the key message and generating sentences that describe salient information in the multimodal documents, (d) a semistructured natural language inference task to develop a framework for data augmentation for tabular inference, (e) the introduction of a human-annotated dataset (ENTSUM) for controllable summarization with a focus on named entities as the aspect to control, and (f) a novel defense mechanism against adversarial attacks (ATINTER). We also examine Bloomberg’s research work, building its own internal, not-for-commercial-use large language model, BloombergGPT, and training it with the goal of demonstrating support for a wide range of tasks within the financial industry. In Part II, we analyze the evolution of automation tasks in the Bloomberg newsroom that led to the creation of Bloomberg’s News Innovation Lab. Technology-assisted content creation has been a reality at Bloomberg News for nearly a decade and has evolved from rules-based headline generation from structured files to the constant exploration of potential ways to assist story creation and storytelling in the financial domain. The Lab now oversees the operation of hundreds of software bots that create semi- and fully automated stories of financial relevance, providing journalists with depth in terms of data and analysis, speed in terms of reacting to breaking news, and transparency to corners of the financial world where data investigation is a gigantic undertaking. The Lab recently introduced new tools that provide journalists with the ability to explore automation on demand while it continues to experiment with ways to assist story production. In Part III, we conceptually discuss the transformative impact that generative AI can have in any newsroom, along with considerations about the technology’s shortcomings in its current state of development. As with any revolutionary new technology, as well as with exciting research opportunities, part of the challenge is balancing any potential positive and negative impacts on society. We offer our principles and guidelines used to inform our approach to experimenting with the new generative AI technologies. Bloomberg News’ style guide reminds us that our “journalism is aimed at possibly the most sophisticated audience in the world, for whom accuracy is essential.”
ECIR 23 tutorial: Neuro-symbolic approaches for information retrieval

Laura Dietz, Hannah Bast, Shubham Chatterjee, and 3 more authors

In European Conference on Information Retrieval, 2023

Abs Website

This tutorial will provide an overview of recent advances on neuro-symbolic approaches for information retrieval. A decade ago, knowledge graphs and semantic annotations technology led to active research on how to best leverage symbolic knowledge. At the same time, neural methods have demonstrated to be versatile and highly effective. From a neural network perspective, the same representation approach can service document ranking or knowledge graph reasoning. End-to-end training allows to optimize complex methods for downstream tasks. We are at the point where both the symbolic and the neural research advances are coalescing into neuro-symbolic approaches. The underlying research questions are how to best combine symbolic and neural approaches, what kind of symbolic/neural approaches are most suitable for which use case, and how to best integrate both ideas to advance the state of the art in information retrieval.
Knowledge graphs: An information retrieval perspective

Ridho Reinanda, Edgar Meij, Maarten Rijke, and 1 more author

Foundations and Trends in Information Retrieval, 2020

Website