Inferring Social Life from Everyday Conversations
|Author:||Izhak Shafran Center for Spoken Language Understanding, Oregon Health & Science University|
|Date:||May 10, 2012|
|Host:||College of Arts and Sciences|
How many friends do you have? How often do you speak with your family? Or friends? How often do you engage in deep conversations? We don't have any means to answer these simple questions scientifically, even though they are at the core of the social nature of our human lives. In the last five years, we have been developing a framework for quantifying social life of older adults, which is motivated in part by our desire to understand the link between social disengagement and higher risk of cognitive decline with age. In this talk, I will set the context with a brief review of previous work in measuring social behavior and delve into our work on analyzing social lives of older adults, aged 79 or above. Since they are less mobile and rely heavily on telephones, we start by analyzing their telephone conversations. For this purpose, we collected all their telephone conversations over the duration of a year. We demonstrate how machine learning techniques can be applied to classify different types of social conversations: business vs residential, family vs others, familiar vs unfamiliar and family vs other residential, achieving accuracies as high as 88%. Our unique corpus of naturalistic everyday conversation affords us an opportunity to probe the nature of telephone conversations. Among other things, we revisit the decade old hypotheses of Schelgoff and Sacks to confirm that "openings" signal the type of conversations but do not find any support that "closings" do. Our results show that the social calling habits are remarkably stable over periods as short as 2-3 months. Our findings can enrich social network models with information about type of social relationships, which so far has been elusive.
This work is a result of continuing collaboration with Anthony Stark, Jeffrey Kaye, and Nicole Larimer.
Izhak Shafran is an Assistant Professor at the Center for Spoken Language Understanding in Oregon Health and Science University (OHSU) in Portland. His primary research is in large vocabulary speech recognition and acoustic modeling. Recently, he began investigating novel methods for assessing cognitive and social abilities of older adults in the context of neuro- degenerative diseases. Before joining OHSU, he was a research faculty at the Center for Speech and Language Processing in Johns Hopkins University and a research member at AT&T Research Labs in Florham Park. He received an NIH Career Development Award in 2010.