Accions del document
LARCA seminar, Feb 23rd, Julia Sidorova
Speaker: Julia A. Sidorova, Alex Computing
Time: Feb 23nd, 10:00
Place: S208, floor -2, Omega Building.
Title: Emotion recognition and the TGI+.x classifier
Abstract: My application is recognition of emotions from speech. My contribution is a novel classification method.
I. the TGI+ classifier:
I adapted a philosophy of combining decision trees and tree automata, which originally comes from optical character recognition. The application to a new domain of speech emotions instead of optical character recognition requested devising new algorithms for the tree grammar inference part. The resulting classification algorithm incorporates statistical and syntactic learning. The syntactic part implements a tree grammar inference algorithm. The statistical part implements the entropy decision tree classifier C4.5.
I have further extended the idea with an built-in feature selection procedure. I tested the classifier on a benchmark data set of ACTED EMOTIONS. The proposed classifier outperformed a state of the art classifier, the multilayer perceptron, with a statistically significant difference in accuracies of 4.68% and the baseline of the C4.5, by 26.58%.
The main property of the TGI+ is human-readability of the classification process, which is of a potential application for example in the clinical context as an assessment and training tool for patients with impaired capabilities to express speech emotions.
II. further evolution to face big and noisy data sets:
I propose the high-level features, which are the distances from a feature vector to a tree automaton accepting class i, for all i in the set of the class labels. I propose to early-fuse the set of the low-level features and the set of the high-level features, submit the resulting set to a feature selection procedure and then do the classification step with the RIPPER-k classifier, which is, like the C4.5, a decision tree classifier yet more robust on big and noisy data sets. The proposed classification scheme outperformed the state of the art top performer (the SVM) with a statistically significant difference in accuracies of 7% on AUTHENTIC EMOTIONS.
Acted and authentic speech emotions are both important for practical applications, yet have been shown to be very different pattern recognition tasks.
Time: Feb 23nd, 10:00
Place: S208, floor -2, Omega Building.
Title: Emotion recognition and the TGI+.x classifier
Abstract: My application is recognition of emotions from speech. My contribution is a novel classification method.
I. the TGI+ classifier:
I adapted a philosophy of combining decision trees and tree automata, which originally comes from optical character recognition. The application to a new domain of speech emotions instead of optical character recognition requested devising new algorithms for the tree grammar inference part. The resulting classification algorithm incorporates statistical and syntactic learning. The syntactic part implements a tree grammar inference algorithm. The statistical part implements the entropy decision tree classifier C4.5.
I have further extended the idea with an built-in feature selection procedure. I tested the classifier on a benchmark data set of ACTED EMOTIONS. The proposed classifier outperformed a state of the art classifier, the multilayer perceptron, with a statistically significant difference in accuracies of 4.68% and the baseline of the C4.5, by 26.58%.
The main property of the TGI+ is human-readability of the classification process, which is of a potential application for example in the clinical context as an assessment and training tool for patients with impaired capabilities to express speech emotions.
II. further evolution to face big and noisy data sets:
I propose the high-level features, which are the distances from a feature vector to a tree automaton accepting class i, for all i in the set of the class labels. I propose to early-fuse the set of the low-level features and the set of the high-level features, submit the resulting set to a feature selection procedure and then do the classification step with the RIPPER-k classifier, which is, like the C4.5, a decision tree classifier yet more robust on big and noisy data sets. The proposed classification scheme outperformed the state of the art top performer (the SVM) with a statistically significant difference in accuracies of 7% on AUTHENTIC EMOTIONS.
Acted and authentic speech emotions are both important for practical applications, yet have been shown to be very different pattern recognition tasks.
