Berlin ML Group - Bringing Deep Learning to poetry analysis.

Talk 1: Deep Learning as a Challenge for Digital Humanities

Speaker: PD Dr. Burkhard Meyer-Sickendiek and Dr. Hussein Hussein

Abstract: Our project "rhythmicalizer" – funded by Volkswagen – applies deep learning methods to literary and cultural studies. This project classifies poetic texts from the website "lyrikline" according to rhythmic patterns, which is important for our idea of modern poetry and its adequate translation. In this talk, we will explain both feature- and neural networks-based classifications, mainly with regard to character-embeddings. A special focus will be on the combination of audio and text data, which is also of interest for other areas such as film-dubbings.

Bio: Burkhard Meyer-Sickendiek
Burkhard is head of the VW research group "Rhythmicalizer -- A Digital Tool to Identify Free Verse Prosody" and leads a project in the field of digital humanities in cooperation with the Berlin portal lyrikline. He studied at the Universities of Bielefeld and Münster, received his doctorate at the University of Tübingen with a thesis on the problem of epigonism and then habilitated as a postdoc at the LMU Munich with a study on German-Jewish satire. In 2009 he came to Freie Universität Berlin as a visiting professor for Modern German Literature, and from 2010 to 2015 he was a Heisenberg fellow of the German Research Foundation (DFG). His research interests include computational humanities, E-learning, 17th- and 18th-century theatre, poetry of modernism/postmodernism, and the literature of German-Jewish modernism. Burkhard is the author of several monographs covering comparative literature from the 17th up to the 21st century.

Bio: Hussein Hussein
Hussein studied at the Dresden University of Technology (TUD), where he received a PhD in the field of acoustic and speech communication. During his PhD studies, Hussein worked as a research assistant at the Laboratory of Acoustics and Speech Communication (TUD) and at the Beuth University of Applied Sciences (BHT) in Berlin. He is presently a research assistant at the Free University Berlin.

Talk 2: OCR as Ultra-Dense Object Detection

Speaker: Dr. Marco Spinaci

Abstract: In this talk I will present a simple single-stage end-to-end deep learning approach for optical character recognition (OCR). Following established practices in the computer vision community, this approach is based on an encoder-decoder architecture; the network predicts both a semantic segmentation (encoding character information) and single-stage object detection (to distinguish multiple occurrences of the same character). In comparison with current state-of-the-art models (based on multiple preprocessing steps such as splitting the page into lines or blocks of text) our approach is easier to train, and it provides superior results to common open source OCR solutions (such as Tesseract). This is based on joint work with Johannes Höhne, Anoop Katti, and Christian Reisswig.

Bio: Marco is a Senior Data Scientist at SAP. He holds a Ph.D. in mathematics from Université Joseph Fourier (Grenoble). Motivated by the successes of deep learning, he left academia four years ago and first applied machine learning techniques to risk management. In 2018 he joined SAP's Deep Learning Center of Excellence, where we build machine learning prototypes for problems in natural language processing and computer vision.