◆  Claire Cardie, Cornell University

Keynote Topic: Information Extraction in the Age of Neural Networks

Short Bio: Claire Cardie is the John C. Ford Professor of Engineering in the Departments of Computer Science and Information Science at Cornell Uni­versity. She has worked since the early 1990’s on application of machine learning methods to problems in Natural Language Processing – on topics ranging from information extraction, noun phrase coreference resolution, text summariza­tion and question answering to the automatic analysis of opinions, argumentation, and decep­tion in text. She has served on the executive committees of the ACL and AAAI and twice as secretargy of NAACL. She has been Program Chair for ACL/COLING, EMNLP and CoNLL, and General Chair for ACL in 2018. Cardie was named a Fellow of the ACL in 2015 and a Fellow of the Association for Computing Machinery (ACM) in 2019. At Cornell, she led the development of the university’s academic programs in Information Science and was the founding Chair of its Information Science Department.

Homepage: https://www.cs.cornell.edu/home/cardie/.

Abstract: In this talk, I’ll examine the state of the NLP subfield of information extraction from its inception almost 30 years ago to its current realization in neural network models. Which aspects of the original formulation of the task do state­-of­-the-­art methods handle well? Where are current methods still falling short?

◆  Edward Grefenstette, Facebook AI Research

Keynote Topic: Exploiting External Knowledge in Reinforcement Learning

Short Bio: Edward Grefenstette is a Research Scientist at Facebook AI Research, and Honorary Associate Professor at UCL. He previously was a Staff Re­search Scientist at DeepMind, and as a Junior Re­search Fellow within Oxford’s Department of Computer Science and Somerville College. He completed his DPhil (PhD) at the University of Oxford in 2013 under the supervision of Profs Coecke and Pulman, and Dr Sadrzadeh, working on applying category­-theoretic tools–initially de­veloped to model quantum information flow–to model compositionality of distributed represen­ tations in natural language semantics. His recent research has covered topics at the intersection of deep learning and machine reasoning, addressing questions such as how neural networks can model or understand logic and mathematics, infer implicit or human­-readable programs, or learn to understand instructions from simulation.

Homepage: https://www.egrefen.com.

Abstract: Humans typically do not approach new tasks tabula rasa. Rather, we face new problems armed not only with (possibly) relevant experience, but also knowledge of a process by which we inform ourselves of the nature of the task, seek clarification, and attempt to learn from others rather than just through trial and error. As Reinforcement Learning is applied to in­creasingly complex tasks, the need for a corresponding meta­-learning process of learning to condition on external information is required. In this talk, we introduce a new environment for RL research for which a large amount of human written external knowledge has been produced (and is very much needed to complete the task), and discuss some of the methods we have explored and challenges we face when it comes to training RL agents to condition on external knowledge for better generalization.

◆  Danqi Chen, Princeton University

Keynote Topic: Open-domain Question Answering: State of the Art and Future Perspectives

Short Bio: Danqi Chen is an Assistant Professor of Computer Science at Princeton University and co­leads the Princeton NLP Group. Her research focuses on deep learning for natural language processing, espe­cially in the intersection of text understanding and knowledge representation & reasoning and applica­tions in question answering, information extraction, and conversational systems. Before joining Prince­ton, Danqi worked as a visiting scientist at Face­book AI Research in Seattle. She received her Ph.D. from Stanford University (2018) and B.E. from Ts­inghua University (2012), both in Computer Science. In the past, she was a recipient of the 2019 Arthur Samuel Best Doctoral Thesis Award at Stanford Uni­versity, a Facebook Fellowship, a Microsoft Research Women’s Fellowship, and paper awards at ACL’16 and EMNLP’17.

Homepage: https://www.cs.princeton.edu/~danqic/.

Abstract: Question answering (QA) is one of the earliest and core topics in natural language processing and has played a central role in many real-­world applications such as search engines and personal assistants. The problem of open­-domain QA, which aims to automatically answer questions posed by humans in a natural language form, usually based on a large collection of unstructured documents, has (re­-)gained a lot of popularity in the last couple of years. This talk will review some of the exciting advances in the field, including some of my earlier and recent experiences in building neural QA systems. In particular, I would like to discuss the role of pre­-training in question answering, learning dense representations for retrieval in place of sparse models, the interplay with structured knowledge, as well as the trade­-off between open­-book and closed­-book models. I will conclude with current limitations and future directions.

◆  Ido Dagan, Bar-Ilan University

Keynote Topic: Modeling, consolidating and exploring open multi-text information

Short Bio: Prof. 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, the founder and head of the nationally­-funded Bar­-Ilan University Data Science Institute and a Fellow of the As­sociation for Computational Linguistics (ACL). His interests are in applied semantic processing, focusing on textual inference, natural open se­mantic representations, consolidation and sum­marization of multi­text information, and inter­active text summarization. Dagan and colleagues initiated textual entailment recognition (RTE, aka NLI) as a generic empirical task. He was the President of the ACL in 2010 and served on its Executive Committee during 2008­2011. In that capacity, he led the establishment of the journal Transactions of the Association for Com­putational Linguistics, which became one of two premiere journals in NLP. Dagan received his B.A. summa cum laude and his Ph.D. (1992) in Computer Science from the Technion. He was a research fellow at the IBM Haifa Scientific Center (1991) and a Member of Tech­nical Staff at AT&T Bell Laboratories (1992­-1994). During 1998­-2003 he was co­-founder and CTO of FocusEngine and VP of Technology of LingoMotors, and has been regularly consulting in the industry since then. His academic research has been involving extensive industrial collaboration and funding, including funds from IBM, Google, Thomson­-Reuters, Bloomberg, Intel and Facebook, as well as collaboration with local companies under funded projects of the Israel Innovation Authority.

Homepage: https://u.cs.biu.ac.il/~dagan/.

Abstract: For most information needs, relevant information is spread across multiple texts. To allow effective human consumption, as well as automatic analysis, of multi­-text information we need mechanisms for modeling information units, consolidating them, and dynamically ex­ploring the consolidated information. In this talk I will outline a research program that targets these goals. First, I will point at two shortcomings of current NLP technology when address­ing multi­-text information: the need for interactive exploration, not addressed by current “static” multi­-document summarization, and the possible inadequacy of purely end­-to-­end methods for supporting such exploration. To that end, we first propose decomposing tex­tual information to explicit sub­-sentential information units. Unlike traditional semantic ap­proaches, which are based on formal predefined schemata, our information units are based on natural language fragments and allow for laymen crowdsourced annotation. We present two such approaches, based on either full proposition spans or on minimal question­-answer pairs, which roughly correspond to predicate­-argument relations (in the spirit of the QA­-SRL paradigm). Next, we suggest approaches for aligning corresponding information units across texts, which extends the notion of cross­-document coreference resolution to the proposition level. Such alignments may facilitate creating consolidated information structures, loosely seen as “open” unsupervised analogs of knowledge graphs. Finally, to facilitate research on interactive exploration of multiple texts, we propose a systematic and replicable evaluation framework suitable for this interactive setting. In the talk I will provide an overview of the research program and illustrate several of its evolving tasks.