The predictive power of Prosody in Spanish question classification using error-driven learning

Heuven van staereling, Daniel van (2023) The predictive power of Prosody in Spanish question classification using error-driven learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bsc_Thesis_AI__Spanish_Question_Prosody_and_Error_Driven_Learning_Revised_Final_s3585131.pdf Download (3MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (127kB)

Abstract

Speech contains more information than just the grammar structure and semantic meaning. The intonation of a speaker gives important information to the sentences meaning. This information is difficult to analyze for a machine. In Spanish, where questions and statements have the same word order, intonation helps distinguish between them. However, context is important in most cases. Therefore, intonation is not the only discerning factor. The lack of context and the noise created by individual differences in speakers make it a difficult problem. This research explores how to develop a process of abstracting pitch to a set of cues that is easier to interpret and what cues to use. Furthermore, this research makes a model to evaluate the predictive accuracy in classifying questions with only pitch information without contextual information to see how much is possible with only pitch. The machine learning model of choice is error-driven learning because of its interpretability. The model needs to train on the created set of cues to learn the important patterns in speech for classifying questions and statements. Analysis of the results shows that only a little information is necessary. It is only necessary to look at the first and last change in the pitch to classify questions. The cues used give predictions with a balanced accuracy of 64.9% and an averaged F1-score of 62.6%. The accuracy is not 100%. The reason is that people use context in regular conversation.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Jones, S.M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	28 Aug 2023 09:21
Last Modified:	28 Aug 2023 09:21
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31276

Actions (login required)

View Item