Computer Science Events
This calendar presented by
Computer Science
[PAST EVENT] Measuring the tone of news coverage: challenges for unsupervised sentiment analysis
November 13, 2015
3pm
---------------------------------------------------------------------------
Measuring the tone of news coverage: challenges for unsupervised sentiment analysis
--------------------------------------------------------------------------
Social scientists are often interested in the tone of news: Does news about the economy grow more positive as the economy grows? Is coverage of Donald Trump more positive than that of Ben Carson? Has coverage of Muslims become more negative since 9/11? Etc. These questions fall within the broader rubric of sentiment analysis, a subject that has been studied extensively by computer scientists and linguists. While that literature primarily engages commercial applications (categorizing movie reviews, product reviews on Amazon, etc.), social science applications present some particular challenges that have not been examined extensively.
This talk addresses three such challenges. First, we demonstrate how to calibrate sentiment measures so as to find the correct cut-off between positive and negative tone, and also to be able to compare across different domains. Second, we show how combining multiple general-purpose lexica improves the quality of our measures. Together, these allow us to identify patterns over time in news coverage at a more fine-grained level than the traditional classification into a few broad categories (positive, neutral, negative). Finally, we present a method to identify the target of negative sentiment. For example: is a negative article about Muslims negative about Muslims, or negative about the treatment of Muslims by the authorities? For obvious reasons, this is a crucial distinction to make; it also turns out to be a very difficult one to make successfully.
--------------------------------------------------------------------------
Maurits van der Veen is associate professor of government at William & Mary. His research focuses on the beliefs and ideas that shape foreign policy decisions, with a particular emphasis on automatically extracting those ideas by mining political texts. He is the director of STAIR lab (Systematic Text Analysis for International Relations, stair.wm.edu) at W&M. The lab offers opportunities for student-faculty research collaboration in developing new methods of automated text classification, drawing on innovations in machine learning and computational linguistics.
Measuring the tone of news coverage: challenges for unsupervised sentiment analysis
--------------------------------------------------------------------------
Social scientists are often interested in the tone of news: Does news about the economy grow more positive as the economy grows? Is coverage of Donald Trump more positive than that of Ben Carson? Has coverage of Muslims become more negative since 9/11? Etc. These questions fall within the broader rubric of sentiment analysis, a subject that has been studied extensively by computer scientists and linguists. While that literature primarily engages commercial applications (categorizing movie reviews, product reviews on Amazon, etc.), social science applications present some particular challenges that have not been examined extensively.
This talk addresses three such challenges. First, we demonstrate how to calibrate sentiment measures so as to find the correct cut-off between positive and negative tone, and also to be able to compare across different domains. Second, we show how combining multiple general-purpose lexica improves the quality of our measures. Together, these allow us to identify patterns over time in news coverage at a more fine-grained level than the traditional classification into a few broad categories (positive, neutral, negative). Finally, we present a method to identify the target of negative sentiment. For example: is a negative article about Muslims negative about Muslims, or negative about the treatment of Muslims by the authorities? For obvious reasons, this is a crucial distinction to make; it also turns out to be a very difficult one to make successfully.
--------------------------------------------------------------------------
Maurits van der Veen is associate professor of government at William & Mary. His research focuses on the beliefs and ideas that shape foreign policy decisions, with a particular emphasis on automatically extracting those ideas by mining political texts. He is the director of STAIR lab (Systematic Text Analysis for International Relations, stair.wm.edu) at W&M. The lab offers opportunities for student-faculty research collaboration in developing new methods of automated text classification, drawing on innovations in machine learning and computational linguistics.