Learning representations with self-attention
13:30-13:45 January 28 · with Jakob Uszkoreit
How well can text representations address lexical composition?
13:45-14:00 January 28 · with Vered Shwartz
Out-of-distribution detection for neural NLP models
14:00-14:10 January 28 · with Hrant Khachatrian
Chargrid: towards understanding 2D documents
14:10-14:20 January 28 · with Christian Reisswig
Panel: representations, interpretability & visualisation
14:20-14:40 January 28 · with Jakob Uszkoreit, Vered Shwartz, Hrant Khachatrian, Christian Reisswig
Explore: research suggestions at your fingertips
14:40-14:50 January 28 · with Richard Zens
Practical transfer learning for NLP with spaCy and Prodigy
14:50-15:00 January 28 · with Ines Montani
15:00-15:20 January 28
Large contexts in neural machine translation
15:20-15:30 January 28 · with Andrei Popescu-Belis
Interactive and adaptive translation for professionals
15:30-15:40 January 28 · with Joern Wuebker
Near real-time multilingual customer service
15:40-15:50 January 28 · with João Graça
Panel: machine translation
15:50-16:00 January 28 · with Andrei Popescu-Belis, João Graça, Joern Wuebker
16:00-16:10 January 28
Towards breaking the closed-world assumption in deep neural networks
16:10-16:20 January 28 · with Michele Sama
Enabling speech-to-meaning with acoustic language processing
16:20-16:30 January 28 · with Nicolas Perony
Building a live recommendation agent for communication in healthcare
16:30-16:40 January 28 · with Lars Maaløe
Panel: products & startups
16:40-17:00 January 28 · with Nicolas Perony, Lars Maaløe, João Graça, Michele Sama, Joern Wuebker, Ines Montani
Professor, HEIG-VD / HES-SO
Senior Data Scientist, SAP SE
Data Scientist, IntelinAir - Researcher, YerevaNN
Founder, Explosion AI
Lead, Google Brain Berlin
Director of Research, Lilt
Staff Research Scientist, Google
PhD Student, Bar-Ilan University
The AI & Language track includes researchers and engineers at top tech companies, research groups and startups using machine learning for language tasks and features four major themes.
Representations and interpretability are especially challenging for language tasks because language is sparse and, unlike with vision, there are no inherently visualisable intermediate representations. Representations, interpretability, evaluation and confidence are intriguing open research questions.
The tools and datasets for building and deploying natural language processing systems are evolving, while online learning pipelines introduce new demands. We will hear from engineering teams behind highly scaled systems on their stacks and processes.
This seminal task, which combines the challenges of natural language understanding, natural language generation and multi-lingual methods, has inspired new approaches like sequence-to-sequence models and attention mechanisms.
The progress in research and engineering has born a wave of startups competing with the established players. Founders of top startups in the space will share how they are building growing platforms and businesses around speech, translation and conversation.