Homepage of Benjamin Matthias Ruppik

Hello and welcome! You have landed on the website of a mathematician who is working on Topological Deep Learning and Representation Learning in Data Science. My name is Ben Ruppik (he/him) and I hold a PhD in Low-dimensional Topology from the Max Planck Institute for Mathematics and University of Bonn. Currently, I am part of the Dialog Systems and Machine Learning lab led by Prof. Milica Gašić at Heinrich-Heine-University Düsseldorf, exploring the use of Topological Data Analysis (TDA) in Natural Language Processing and Task-oriented Dialogue Systems. On this website, you will find some of my projects, publications, and other resources related to my research interests. You are welcome to reach out to me via e-mail or LinkedIn in German or English!

Selected Papers and Preprints

For a complete list, follow the links to arXiv; ORCID iD icon ORCID; Google Scholar; ResearchGate.

Topological Deep Learning

Image related to Local Topology Measures of Contextual Language Model Latent Spaces With Applications to Dialogue Term Extraction

Local Topology Measures of Contextual Language Model Latent Spaces With Applications to Dialogue Term Extraction

[Benjamin Ruppik, Michael Heck, Carel van Niekerk, Renato Vukovic, Hsien-chin Lin, Shutong Feng, Marcus Zibrowius, Milica Gašić]

We show that contextual topological features derived from a given text corpus can be used to enhance the performance of a tagging task on dialogue data.

Nominated for Best Paper Award at SIGDIAL 2024Proceedings of the 25th Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL 2024), Kyoto University, Japandoi:10.18653/v1/2024.sigdial-1.31arXiv:2408.03706Public Codebase

Image related to Dialogue Term Extraction using Transfer Learning and Topological Data Analysis

Dialogue Term Extraction using Transfer Learning and Topological Data Analysis

[Renato Vukovic, Michael Heck, Benjamin Ruppik, Carel van Niekerk, Marcus Zibrowius, Milica Gašić]

We demonstrate that topological features derived from neighborhoods in a static word embedding can be used to extract dialogue terms. In particular, the Wasserstein norm and Persistence Image Vectorization of the persistent homology of the neighborhood appear to be generalizable features which are effective in a transfer learning approach.

Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL 2022), pages 564-581, Edinburgh, UKdoi:10.18653/v1/2022.sigdial-1.53arXiv:2208.10448

Task-oriented Dialogue Systems

Image related to Dialogue Ontology Relation Extraction via Constrained Chain-of-Thought Decoding

Dialogue Ontology Relation Extraction via Constrained Chain-of-Thought Decoding

[Renato Vukovic, David Arps, Carel van Niekerk, Benjamin Matthias Ruppik, Hsien-Chin Lin, Michael Heck, Milica Gašić]

We present a novel method that enhances relation extraction in task-oriented dialogue systems by adapting Chain-of-Thought (CoT) decoding to generative models, improving their generalization capabilities. By constraining the decoding process to ontology terms and relations, our approach reduces hallucinations and demonstrates improved performance on two widely used datasets.

Proceedings of the 25th Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL 2024), Kyoto University, Japandoi:10.18653/v1/2024.sigdial-1.33arXiv:2408.02361

Image related to ChatGPT for Zero-shot Dialogue State Tracking: A Solution or an Opportunity?

ChatGPT for Zero-shot Dialogue State Tracking: A Solution or an Opportunity?

[Michael Heck, Nurul Lubis, Benjamin Ruppik, Renato Vukovic, Shutong Feng, Christian Geishauser, Hsien-Chin Lin, Carel van Niekerk, Milica Gašić]

We demonstrate that ChatGPT achieves state-of-the-art zero-shot dialogue state tracking (DST) without task-specific training, outperforming methods that rely on heavy data augmentation and fine-tuning. While general-purpose models like ChatGPT have limitations that prevent them from fully replacing specialized systems, their in-context learning abilities can support the development of dynamic and dedicated DST systems.

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 936--950, Toronto, Canadadoi:10.18653/v1/2023.acl-short.81arXiv:2306.01386

Low-dimensional Topology

Image related to A group-theoretic framework for low-dimensional topology

A group-theoretic framework for low-dimensional topology

[Sarah Blackwell, Rob Kirby, Michael Klug, Vincent Longo, Benjamin Ruppik]

A correspondence, by way of Heegaard splittings, between closed oriented 3-manifolds and pairs of surjections from a surface group to a free group has been studied by Stallings, Jaco, and Hempel. This correspondence, by way of trisections, was recently extended by Abrams, Gay, and Kirby to the case of smooth, closed, connected, oriented 4-manifolds. We unify these perspectives and generalize this correspondence to the case of links in closed oriented 3-manifolds and links of knotted surfaces in smooth, closed, connected, oriented 4-manifolds. The algebraic manifestations of these four subfields of low-dimensional topology (3-manifolds, 4-manifolds, knot theory, and knotted surface theory) are all strikingly similar, and this correspondence perhaps elucidates some unique character of low-dimensional topology.

To appear in Algebr. Geom. Topol.arXiv:2301.05685

Image related to Casson-Whitney Unknotting Numbers of 2-spheres in the 4-sphere

Casson-Whitney Unknotting Numbers of 2-spheres in the 4-sphere

[Jason Joseph, Michael Klug, Benjamin Ruppik, Hannah Schwartz]

Inspired by Schneiderman-Teichner's perspective on Gabai's 4-dimensional light bulb theorem we define the Casson-Whitney unknotting number of knotted 2-spheres in S4: Since every 2-knot K in S4 can be obtained by first performing a number of (trivial) finger moves on the unknot, and in a next step removing the resulting intersection points in pairs via Whitney moves (along possibly complicated discs), we can define ucw(K) as the minimal number of finger moves needed in such a process to arrive at K. We relate this to the 1-handle stabilization number and look specifically at examples of ribbon 2-knots and twist-spins of classical knotted arcs.

J. Topology 14.4 (2021), 1321-1350doi:10.1112/topo.12209arXiv:2007.13244

Finger-Whitney Unknotting Numbers - Part 1 (.pdf, handwritten slides)Finger-Whitney Unknotting Numbers - Part 2 [Michael] (.pdf, handwritten slides)

Thesis

Iterated Whitehead double of the dual knot in the Poincare homology sphere

Casson-Whitney unknotting, Deep slice knots and Group trisections of knotted surface type

Advisors: Arunima Ray and Peter Teichner.

PhD thesis published at University of Bonn.

Main new results:
  • An explicit construction of regular homotopies between rim surgeries (which subsumes some of the previous applications to twist-spun knots).
  • A generalization of the fusion number upper bound for the Casson-Whitney number to higher genus ribbon surfaces in the 4-sphere.
  • The stabilization number of Suciu's ribbon 2-knots Rk is equal to 1.
  • The Casson-Whitney distance of the standard unknotted RP2 and Suciu's associated RP2-knots Rk#RP2 is equal to 1.
  • Partial results of the algebraic Casson-Whitney number of Suciu's knots.
  • The (+1)-trace on the right handed trefoil knot contains infinitely many non-local knots in its boundary which exhibit the difference between shallow topological sliceness and smooth deep sliceness. This is an interplay between Freedman's construction of a topological slice for 3-fold iterated positively clasped Whitehead doubles on the one hand, and the Heegaard-Floer τ-invariant giving a smooth obstruction for sliceness in X×[0,1].
  • Group trisections of knotted surface type of the trefoil group, which comes from a bridge trisection of Suciu's ribbon 2-knots.

Resources

Activities

Summer Term 2023

Teaching
Talks
Conferences

Winter Term 2022/2023

Teaching
Talks
Conferences

Summer Term 2022

Teaching
Conferences

Winter Term 2021/2022

Talks

Summer Term 2021

Talks
Conferences

Winter Term 2020/2021

Talks
Conferences

Summer Term 2020

Talks
Conferences/Workshops

Winter Term 2019/2020

Talks
Conferences
  • Winter Braids X, Pisa, February 17 - 21, 2020
  • Low-dimensional topology workshop, Regensburg, October 21 - 23, 2019.

Summer Term 2019

Talks
Conferences

Winter Term 2018/2019

Summer Term 2018

  • Material for the rep 'Introduction to Geometry and Topology' in the summer term 2018.

Winter Term 2017/2018

Summer Term 2017

  • Material for the rep 'Introduction to Geometry and Topology' in the summer term 2017.
  • Handout (.pdf) for a talk on the First Lie Theorem.

Summer Term 2016

Other

Links: Github, MathStackExchange

Dr. Benjamin Matthias Ruppik