BEGIN:VCALENDAR VERSION:2.0 PRODID:-//128.220.36.25//NONSGML kigkonsult.se iCalcreator 2.26.9// CALSCALE:GREGORIAN METHOD:PUBLISH X-FROM-URL:https://www.clsp.jhu.edu X-WR-TIMEZONE:America/New_York BEGIN:VTIMEZONE TZID:America/New_York X-LIC-LOCATION:America/New_York BEGIN:STANDARD DTSTART:20231105T020000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RDATE:20241103T020000 TZNAME:EST END:STANDARD BEGIN:DAYLIGHT DTSTART:20240310T020000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RDATE:20250309T020000 TZNAME:EDT END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:ai1ec-20716@www.clsp.jhu.edu DTSTAMP:20240328T082055Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:
Abstract
\nOver the last few years\, deep neural models have taken over the field of natural language processin g (NLP)\, brandishing great improvements on many of its sequence-level tas ks. But the end-to-end nature of these models makes it hard to figure out whether the way they represent individual words aligns with how language b uilds itself from the bottom up\, or how lexical changes in register and d omain can affect the untested aspects of such representations.
\nIn this talk\, I will present NYTWIT\, a dataset created to challenge large l anguage models at the lexical level\, tasking them with identification of processes leading to the formation of novel English words\, as well as wit h segmentation and recovery of the specific subclass of novel blends. I wi ll then present XRayEmb\, a method which alleviates the hardships of proce ssing these novelties by fitting a character-level encoder to the existing models’ subword tokenizers\; and conclude with a discussion of the drawba cks of current tokenizers’ vocabulary creation schemes.
\nBi ography
\nYuval Pinter is a Senior Lecturer in the Department of Comp uter Science at Ben-Gurion University of the Negev\, focusing on natural l anguage processing. Yuval got his PhD at t he Georgia Institute of Technology School of Interactive Computing as a Bl oomberg Data Science PhD Fellow. Before that\, he worked as a Research Eng ineer at Yahoo Labs and as a Computational Linguist at Ginger Software\, a nd obtained an MA in Linguistics and a BSc in CS and Mathematics\, both fr om Tel Aviv University. Yuval blogs (in He brew) about language matters on Dagesh Kal.
DTSTART;TZID=America/New_York:20210910T120000 DTEND;TZID=America/New_York:20210910T131500 LOCATION:Hackerman Hall B17 @ 3400 N. Charles Street\, Baltimore\, MD SEQUENCE:0 SUMMARY:Yuval Pinter (Ben-Gurion University – Virtual Visit) “Challenging a nd Adapting NLP Models to Lexical Phenomena” URL:https://www.clsp.jhu.edu/events/yuval-pinter/ X-COST-TYPE:free X-TAGS;LANGUAGE=en-US:2021\,Pinter\,September END:VEVENT BEGIN:VEVENT UID:ai1ec-21259@www.clsp.jhu.edu DTSTAMP:20240328T082055Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract
\nNatural language processin g has been revolutionized by neural networks\, which perform impressively well in applications such as machine translation and question answering. D espite their success\, neural networks still have some substantial shortco mings: Their internal workings are poorly understood\, and they are notori ously brittle\, failing on example types that are rare in their training d ata. In this talk\, I will use the unifying thread of hierarchical syntact ic structure to discuss approaches for addressing these shortcomings. Firs t\, I will argue for a new evaluation paradigm based on targeted\, hypothe sis-driven tests that better illuminate what models have learned\; using t his paradigm\, I will show that even state-of-the-art models sometimes fai l to recognize the hierarchical structure of language (e.g.\, to conclude that “The book on the table is blue” implies “The table is blue.”) Second\ , I will show how these behavioral failings can be explained through analy sis of models’ inductive biases and internal representations\, focusing on the puzzle of how neural networks represent discrete symbolic structure i n continuous vector space. I will close by showing how insights from these analyses can be used to make models more robust through approaches based on meta-learning\, structured architectures\, and data augmentation.
\nBiography
\nTom McCoy is a PhD candidate in the Department of Cognitive Science at Johns Hopkins University. As an undergr aduate\, he studied computational linguistics at Yale. His research combin es natural language processing\, cognitive science\, and machine learning to study how we can achieve robust generalization in models of language\, as this remains one of the main areas where current AI systems fall short. In particular\, he focuses on inductive biases and representations of lin guistic structure\, since these are two of the major components that deter mine how learners generalize to novel types of input.
DTSTART;TZID=America/New_York:20220131T120000 DTEND;TZID=America/New_York:20220131T131500 LOCATION:Ames Hall 234 @ 3400 N. Charles Street\, Baltimore\, MD 21218 SEQUENCE:0 SUMMARY:Tom McCoy (Johns Hopkins University) “Opening the Black Box of Deep Learning: Representations\, Inductive Biases\, and Robustness” URL:https://www.clsp.jhu.edu/events/tom-mccoy-johns-hopkins-university-open ing-the-black-box-of-deep-learning-representations-inductive-biases-and-ro bustness/ X-COST-TYPE:free X-TAGS;LANGUAGE=en-US:2022\,January\,McCoy END:VEVENT BEGIN:VEVENT UID:ai1ec-22394@www.clsp.jhu.edu DTSTAMP:20240328T082055Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract
\nModel robustness and spurious correlations have received increasing atten tion in the NLP community\, both in methods and evaluation. The term “spur ious correlation” is overloaded though and can refer to any undesirable sh ortcuts learned by the model\, as judged by domain experts.
\nWhen designing mitigation algorithms\, we oft en (implicitly) assume that a spurious feature is irrelevant for predictio n. However\, many features in NLP (e.g. word overlap and negation) are not spurious in the sense that the background is spurious for classifying obj ects in an image. In contrast\, they carry important information that’s ne eded to make predictions by humans. In this talk\, we argue that it is mor e productive to characterize features in terms of their necessity and suff iciency for prediction. We then discuss the implications of this categoriz ation in representation\, learning\, and evaluation.
\nBiogr aphy
\nHe He is an Assistant Professor in the Department of Computer Science and the Center for Data Science at New York University. She obtained her PhD in Computer Science at the University of Maryland\, C ollege Park. Before joining NYU\, she spent a year at AWS AI and was a pos t-doc at Stanford University before that. She is interested in building ro bust and trustworthy NLP systems in human-centered settings. Her recent re search focus includes robust language understanding\, collaborative text g eneration\, and understanding capabilities and issues of large language mo dels.
\n DTSTART;TZID=America/New_York:20221014T120000 DTEND;TZID=America/New_York:20221014T131500 LOCATION:Hackerman Hall B17 @ 3400 N. Charles Street\, Baltimore\, MD 21218 SEQUENCE:0 SUMMARY:He He (New York University) “What We Talk about When We Talk about Spurious Correlations in NLP” URL:https://www.clsp.jhu.edu/events/he-he-new-york-university/ X-COST-TYPE:free X-TAGS;LANGUAGE=en-US:2022\,He\,October END:VEVENT BEGIN:VEVENT UID:ai1ec-23304@www.clsp.jhu.edu DTSTAMP:20240328T082055Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract
\nTransformers are essential to pretraining. As we approach 5 years of BERT\, the connection between a ttention as architecture and transfer learning remains key to this central thread in NLP. Other architectures such as CNNs and RNNs have been used t o replicate pretraining results\, but these either fail to reach the same accuracy or require supplemental attention layers. This work revisits the semanal BERT result and considers pretraining without attention. We consid er replacing self-attention layers with recently developed approach for lo ng-range sequence modeling and transformer architecture variants. Specific ally\, inspired by recent papers like the structured space space sequence model (S4)\, we use simple routing layers based on state-space models (SSM ) and a bidirectional model architecture based on multiplicative gating. W e discuss the results of the proposed Bidirectional Gated SSM (BiGS) and p resent a range of analysis into its properties. Results show that architec ture does seem to have a notable impact on downstream performance and a di fferent inductive bias that is worth exploring further.
\nBi ography
\n