Yes or No, Does Pitch Matter? Modeling the Conditions for Meaning in Polar Questions

Cristina Hain, Nanette Veilleux, and Byron Ahn, Department of Linguistics, Princeton University, 330 Alexander Street, Princeton NJ 08540

Machine Learning tools can be useful in investigating long standing questions about spoken communication beyond the word sequence. Specifically, intonation is a feature of human speech that can change the meaning of sentences, e.g., affecting whether a string is a question or an assertion. Within yes/no questions there is understudied (and potentially meaningful) intonational variation. This study uses a corpus of American English polar yes/no questions, naturalistically produced in a game context: speakers were recorded while playing a modified version of “Guess Who?”.  This corpus had been previously annotated for context, and for this work was intonationally annotated with respect to final pitch movements. Following previous work that successfully used linear mixed effects models to identify variables that conditioned final pitch movements, this work uses Random Forests to predict the type of meaning conveyed by the question. (Random Forests are a machine learning technique that minimizes the impact of heterogeneous data (different scales) and redundant features.) Two types of questions (about a property of the object vs. about the identity of the object) were predicted with 71.9% accuracy. In addition, the intonation features (f0 measured at the beginning and the end of the last syllable) were the most important for prediction.  

Additional Abstract Information

Presenter: Cristina Hain

Institution: Princeton

Type: Poster

Subject: Linguistics & World Languages

Status: Approved

Time and Location

Session: Poster 8
Date/Time: Tue 5:00pm-6:00pm
Session Number: 5545