Instructors: Instructors: Brian Dillon (University of Massachusetts, Amherst), Suhas Arehalli (Johns Hopkins University)
Deep learning models have become part of the standard toolkit in areas of Natural Language Processing, including language modeling. They have also become an important topic in linguistics, by helping researchers explore how these learning architectures do and don’t generalize linguistic patterns like humans do. But how can these tools be used to help understand human language processing? In this Minicourse, we will explore this question. We will provide a hands-on introduction to basic deep-learning techniques (LSTMs) using out of the box models. We will introduce some of the different ways that such models have been used in the study of linguistic and psycholinguistics. In particular, we will explore the relationship between the workings of deep learning models and key claims in contemporary psycholinguistic theories such as incremental probabilistic prediction and interference in working memory. We do this through two case studies from the psycholinguistic literature, agreement attraction and garden path effects, with the goal of better understanding how these models can illuminate the cognitive processes that underlie widely studied psycholinguistic phenomena. Students in this minicourse should have basic familiarity with Python, but no particular expertise in machine learning or deep learning is expected. Students will be exposed to hands-on modeling using LSTM models through the use of Google CoLab and Jupyter Notebooks, and will be given experience developing their own novel LSTM simulations on psycholinguistic phenomena of their choice.
Date: Thursday, January 6, 2022
Time: 10:00 AM - 3:00 PM, with a lunch break
Room: Columbia 6
Instructors: Instructors: Brian Dillon (University of Massachusetts, Amherst), Suhas Arehalli (Johns Hopkins University)
Deep learning models have become part of the standard toolkit in areas of Natural Language Processing, including language modeling. They have also become an important topic in linguistics, by helping researchers explore how these learning architectures do and don’t generalize linguistic patterns like humans do. But how can these tools be used to help understand human language processing? In this Minicourse, we will explore this question. We will provide a hands-on introduction to basic deep-learning techniques (LSTMs) using out of the box models. We will introduce some of the different ways that such models have been used in the study of linguistic and psycholinguistics. In particular, we will explore the relationship between the workings of deep learning models and key claims in contemporary psycholinguistic theories such as incremental probabilistic prediction and interference in working memory. We do this through two case studies from the psycholinguistic literature, agreement attraction and garden path effects, with the goal of better understanding how these models can illuminate the cognitive processes that underlie widely studied psycholinguistic phenomena. Students in this minicourse should have basic familiarity with Python, but no particular expertise in machine learning or deep learning is expected. Students will be exposed to hands-on modeling using LSTM models through the use of Google CoLab and Jupyter Notebooks, and will be given experience developing their own novel LSTM simulations on psycholinguistic phenomena of their choice.