Prof. Ido Dagan
Department of Computer Science, Bar-Ilan University
Israel
Ido Dagan is a Professor at the Department of Computer Science at Bar-Ilan University, Israel, the founder of the Natural Language Processing (NLP) Lab at Bar-Ilan, a Fellow of the Association for Computational Linguistics (ACL), and was the founding Director of the nationally funded Data Science Institute of Bar-Ilan University. His interests are in applied semantic processing, focusing on textual inference, natural open semantic representations, consolidation and summarization of multi-text information, and interactive text summarization and exploration. Dagan and colleagues initiated and promoted the prominent empirical task of textual entailment recognition (RTE, later aka NLI). He was the President of the ACL in 2010 and served on its Executive Committee during 2008-2011, where he led the establishment of the journal Transactions of the Association for Computational Linguistics, currently one of two premiere journals in NLP. During 1998-2003 he was co-founder and CTO of FocusEngine and VP of Technology of LingoMotors, and later has been regularly consulting in the industry. His academic research involved extensive industrial collaboration, including funds from IBM, Google, Thomson-Reuters, Bloomberg, Intel and Facebook.
Abstract Summary: In favor of modularity and decomposition: tasks and representations Deep learning has pushed us to develop monolithic models and representations, which we understand and control to a much lesser degree than earlier methods. In this talk I suggest revisiting the pursuit for decomposed models and representations, aiming to obtain better modularity, transparency and control over our systems while leveraging deep learning power. With respect to modeling, I will discuss an approach where a complex task is decomposed to inherent subtasks, each addressed by a distinguished model. The approach will be illustrated for challenging use cases in (multi-) text summarization, yielding appealing advantages with respect to explainability, reusability and human-AI collaboration. With respect to representations, I will advocate an approach that decomposes textual information into a systematic and comprehensive set of “minimal” question-answer pairs. This provides an appealing semi-structured natural-language based representation, which may be viewed as a midpoint between fully-structured representations and opaque neural representations. The utility of this approach will be illustrated through an ongoing project that aims to identify unfaithful pieces of generated text in a fine-grained localized manner. Overall, I will suggest that by leveraging modularity and decomposition along with available powerful language models we should move on to setting higher bars for computational text understanding and collaborative AI.