Michael Völske

The annual lecture on the fundamentals of machine learning is part of the curricula at Bauhaus-Universität Weimar and Universität Leipzig.

My responsibilities included teaching in-person and holding online classes, designing lecture materials and lab exercises, and supervising teaching assistants. To support the associated programming labs, I rolled out a Kubernetes-based Jupyterhub deployment servicing up to a hundred students at a time. I significantly expanded the teaching materials on neural networks and deep learning.

To better understand how web users ask questions and how search engines can serve their information needs, I analyzed a query log of more than one billion individual entries from a major commercial search engine. Due to the amount of data involved, I leveraged big data technologies, including Apache Spark deployed on a 100-node Hadoop cluster, to implement a comprehensive data cleaning and analysis pipeline. Using user-labeled community question answering data as a training set, I developed query classification models based on different feature sets—such as unigrams and topic models—and evaluated their effectiveness and efficiency. The resulting publication was accepted at CIKM 2015 and served as the basis for varied downstream research on question classification.

Axiomatic IR deals with formal properties that a good relevance scoring function should fulfill.

This research project investigated strategies to make retrieval axioms directly usable to benefit real-world search engines. A pilot study published at CIKM 2016 showed how retrieval axioms could directly modify result rankings and thus improve search result quality. Follow-up work in ICTIR 2021 showed how retrieval axioms could generate explanations for arbitrary rankings, making complex relevance scoring functions such as those based on deep neural networks more interpretable. The axiomatic re-ranking pipeline I implemented has contributed to several further publications.

I was primarily responsible for the planning, procurement, installation, and maintenance of research computing infrastructure spanning nearly 300 individual servers over more than five years. To provision, maintain, and monitor these systems, my team leveraged infrastructure-as-code technologies such as SaltStack, and developed custom solutions for user authentication and access control on top of Gitlab, Kubernetes, and Slurm. Altogether, the infrastructure has enabled hundreds of successful publications at high-ranking venues across our research group.

An abstractive summary distills the core content of a source text using novel words and phrases not necessarily found in the source. This yields a shorter summary and more fluent reading experience than a mere selection of relevant phrases from the source could achieve. To learn automatic abstractive summarization, machine learning models require a large number of source-summary pairs as training data.

I led an effort to mine more than four million human-written source-summary pairs from the social media posts made to the Reddit platform, where users frequently summarize long messages as a courtesy to their readers, prefixing the summary with "TL;DR" ("too long; didn't read") or similar. To handle the scale of the input data (all Reddit posts ever made up to the year 2017), I leveraged technologies like Hadoop and Spark. The resulting Webis-TLDR-17 dataset formed the basis for a shared-task competition on abstractive summarization organized by a mixed team of researchers from the industry and academia at INLG 2019. Our dataset was subsequently included in the Huggingface and TensorFlow datasets libraries and has been cited in more than 40 publications so far.