School of Informatics

Jon Oberlander Memorial Lecture

Mari Ostendorf - Contextualised Language Processing with Explicit Representations of Context

Note: the lecture will be followed by a drinks reception from 5:00 to 6:00pm.

Abstract

Mari Ostendorf

Understanding human language is facilitated by considering the context in which it occurs, for both humans and computers. For computational systems, the significance of this observation is demonstrated by the impact that contextualized word embeddings have had on natural language processing tasks in the past few years, leading to dramatic improvements in performance. In most work, "context" is defined by the sequence of neighboring words, which is useful for tightening the representation of syntactic function and semantic meaning of a word. However, there are other types of context that are easily available for many forms of language: speech vs. text, author or speaker index, location, social context, etc. This talk describes three different architectures for contextualizing language that involve learning embedded representations of context as a separate factor in the model. The examples leverage three different types of information: metadata associated with online text, turn-taking structure in dialog, and alignment of audio and word transcripts in speech.  In experiments with a variety of contexts and language processing tasks, we show that explicit representation of context both improves performance and provides insights into characteristics of language associated with different contexts.

Biography

Mari Ostendorf joined the University of Washington in 1999. She is an Endowed Professor of System Design Methodologies in the Electrical & Computer Engineering Department, an Adjunct Professor in Linguistics and in Computer Science & Engineering, and Associate Vice Provost for Research. She is a Fellow of the IEEE, ISCA and ACL, a Scottish Informatics and Computer Science Alliance Distinguished Visiting Fellow, a former Australian-American Fulbright Scholar, and a member of the Washington State Academy of Sciences. In 2017, Prof. Ostendorf served as a faculty advisor for the student team winning the inaugural Alexa Prize competition to build a socialbot, and conversational AI is a focus of her current work. Her research explores dynamic models for understanding and generating speech and text, particularly in multi-party contexts, and it contributes to a variety of applications, from education to clinical and scientific information extraction.

School of Informatics, CDT in NLP, UKRI logos

This event is organised by the School of Informatics and hosted in conjunction with the UKRI CDT in NLP.

Dec 12 2019 -

Jon Oberlander Memorial Lecture

Mari Ostendorf will describe three different architectures for contextualising language that involve learning embedded representations of context as a separate factor in the model.

G07/G07A

Informatics Forum, University of Edinburgh
10 Crichton Street
Edinburgh
EH8 9AB