Abstract

Transformers are essential to pretraining. As we approach 5 years of BERT, the connection between attention as architecture and transfer learning remains key to this central thread in NLP. Other architectures such as CNNs and RNNs have been used to 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 consider replacing self-attention layers with recently developed approach for long-range sequence modeling and transformer architecture variants. Specifically, 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. We discuss the results of the proposed Bidirectional Gated SSM (BiGS) and present a range of analysis into its properties. Results show that architecture does seem to have a notable impact on downstream performance and a different inductive bias that is worth exploring further.

Biography

Abstract

While large language models have advanced the state-of-the-art in natural language processing, these models are trained on large-scale datasets, which may include harmful information. Studies have shown that as a result, the models exhibit social biases and generate misinformation after training. In this talk, I will discuss my work on analyzing and interpreting the risks of large language models across the areas of fairness, trustworthiness, and safety. I will first describe my research in the detection of dialect bias between African American English (AAE) vs. Standard American English (SAE). The second part investigates the trustworthiness of models through the memorization and subsequent generation of conspiracy theories. I will end my talk with recent work in AI safety regarding text that may lead to physical harm.

Biography

Sharon is a 5th-year Ph.D. candidate at the University of California, Santa Barbara, where she is advised by Professor William Wang. Her research interests lie in natural language processing, with a focus on Responsible AI. Sharon’s research spans the subareas of fairness, trustworthiness, and safety, with publications in ACL, EMNLP, WWW, and LREC. She has spent summers interning at AWS, Meta, and Pinterest. Sharon is a 2022 EECS Rising Star and a current recipient of the Amazon Alexa AI Fellowship for Responsible AI.

Abstract

Biases in datasets, or unintentionally introduced spurious cues, are a common source of misspecification in machine learning. Performant models trained on such data can gender stereotype or be brittle under distribution shift.  In this talk, we present several results in multimodal and question answering applications studying sources of dataset bias, and several mitigation methods.  We propose approaches where known dimensions of dataset bias are explicitly factored out of a model during learning, without needing to modify data. Finally, we ask whether dataset biases can be attributable to annotator behavior during annotation. Drawing inspiration from work in psychology on cognitive biases, we show certain behavioral patterns are highly indicative of the creation of problematic (but valid) data instances in question answering. We give evidence that many existing observations around how dataset bias propagates to models can be attributed to data samples created by annotators we identify.

Biography

Mark Yatskar is an Assistant Professor at University of Pennsylvania in the department of Computer and Information Science. He did his PhD at University of Washington co-advised by Luke Zettlemoyer and Ali Farhadi. He was a Young Investigator at the Allen Institute for Artificial Intelligence for several years working with their computer vision team, Prior. His work spans Natural Language Processing, Computer Vision, and Fairness in Machine Learning. He received a Best Paper Award at EMNLP for work on gender bias amplification, and his work has been featured in Wired and the New York Times.

Abstract

While GPT models have shown impressive performance on summarization and open-ended text generation, it’s important to assess their abilities on more constrained text generation tasks that require significant and diverse rewritings. In this talk, I will discuss the challenges of evaluating systems that are highly competitive and perform close to humans on two such tasks: (i) paraphrase generation and (ii) text simplification. To address these challenges, we introduce an interactive Rank-and-Rate evaluation framework. Our results show that GPT-3.5 has made a major step up from fine-tuned T5 in paraphrase generation, but still lacks the diversity and creativity of humans who spontaneously produce large quantities of paraphrases.

Additionally, we demonstrate that GPT-3.5 performs similarly to a single human in text simplification, which makes it difficult for existing automatic evaluation metrics to distinguish between the two. To overcome this shortcoming, we propose LENS, a learnable evaluation metric that outperforms SARI, BERTScore, and other existing methods in both automatic evaluation and minimal risk decoding for text generation.

Biography

Wei Xu is an assistant professor in the School of Interactive Computing at the Georgia Institute of Technology, where she is also affiliated with the new NSF AI CARING Institute and Machine Learning Center. She received her Ph.D. in Computer Science from New York University and her B.S. and M.S. from Tsinghua University. Xu’s research interests are in natural language processing, machine learning, and social media, with a focus on text generation, stylistics, robustness and controllability of machine learning models, and reading and writing assistive technology. She is a recipient of the NSF CAREER Award, CrowdFlower AI for Everyone Award, Criteo Faculty Research Award, and Best Paper Award at COLING’18. She has also received funds from DARPA and IARPA. She is an elected member of the NAACL executive board and regularly serves as a senior area chair for AI/NLP conferences.

Abstract

Understanding the implications underlying a text is critical to assessing its impact, in particular the social dynamics that may result from a reading of the text. This requires endowing artificial intelligence (AI) systems with pragmatic reasoning, for example to correctly conclude that the statement “Epidemics and cases of disease in the 21st century are “staged”” relates to unfounded conspiracy theories. In this talk, I discuss how shortcomings in the ability of current AI systems to reason about pragmatics present challenges to equitable detection of false or harmful language. I demonstrate how these shortcomings can be addressed by imposing human-interpretable structure on deep learning architectures using insights from linguistics.

In the first part of the talk, I describe how adversarial text generation algorithms can be used to improve robustness of content moderation systems. I then introduce a pragmatic formalism for reasoning about harmful implications conveyed by social media text. I show how this pragmatic approach can be combined with generative neural language models to uncover implications of news headlines. I also address the bottleneck to progress in text generation posed by gaps in evaluation of factuality. I conclude by showing how context-aware content moderation can be used to ensure safe interactions with conversational agents.

Biography

Saadia Gabriel is a PhD candidate in the Paul G. Allen School of Computer Science & Engineering at the University of Washington, advised by Prof. Yejin Choi and Prof. Franziska Roesner. Her research
revolves around natural language processing and machine learning, with a particular focus on building systems for understanding how social commonsense manifests in text (i.e. how do people typically behave in social scenarios), as well as mitigating spread of false or harmful text (e.g. Covid-19 misinformation). Her work has been covered by a wide range of media outlets like Forbes and TechCrunch. It has also received a 2019 ACL best short paper nomination, a 2019 IROS RoboCup best paper nomination and won a best paper award at the 2020 WeCNLP summit. Prior to her PhD, Saadia received a BA summa cum laude from Mount Holyoke College in Computer Science and Mathematics.

Abstract

Advanced neural language models have grown ever larger and more complex, pushing forward the limits of language understanding and generation, while diminishing interpretability. The black-box nature of deep neural networks blocks humans from understanding them, as well as trusting and using them in real-world applications. This talk will introduce interpretation techniques that bridge the gap between humans and models for developing trustworthy natural language processing
(NLP). I will first show how to explain black-box models and evaluate their explanations for understanding their prediction behavior. Then I will introduce how to improve the interpretability of neural language models by making their decision-making transparent and rationalized. Finally, I will discuss how to diagnose and improve models (e.g., robustness) through the lens of explanations. I will conclude with future research directions that are centered around model interpretability and committed to facilitating communications and interactions between intelligent machines, system developers, and end users for long-term trustworthy AI.

Biography

Hanjie Chen is a Ph.D. candidate in Computer Science at the University of Virginia, advised by Prof. Yangfeng Ji. Her research interests lie in Trustworthy AI, Natural Language Processing (NLP), and
Interpretable Machine Learning. She develops interpretation techniques to explain neural language models and make their prediction behavior transparent and reliable. She is a recipient of the Carlos and Esther Farrar Fellowship and the Best Poster Award at the ACM CAPWIC 2021. Her work has been published at top-tier NLP/AI conferences (e.g., ACL, AAAI, EMNLP, NAACL) and selected by the National Center for Women & Information Technology (NCWIT) Collegiate Award Finalist 2021. She (as the primary instructor) co-designed and taught the course, Interpretable Machine Learning, and was awarded the UVA CS Outstanding Graduate Teaching Award and University-wide Graduate Teaching Awards Nominee (top 5% of graduate instructors). More details can be found at
https://www.cs.virginia.edu/~hc9mx