Improved semi-supervised autoencoder for deception detection.

Area of Science:

• Computational Linguistics
• Machine Learning
• Speech Processing

Background:

• Current speech-based deception detection algorithms are hindered by a scarcity of labeled training data.
• Vast quantities of readily available unlabeled speech data remain underutilized in deception detection research.

Purpose of the Study:

• To propose a semi-supervised additive noise autoencoder model for enhancing deception detection.
• To leverage both labeled and unlabeled data for more robust and accurate lie detection from speech.

Main Methods:

• Developed a semi-supervised autoencoder with a two-layer encoder-decoder structure and a classifier.
• Modified the hidden layer activation function to suit deception speech characteristics.
• Incorporated dropout at each layer to mitigate overfitting and directly linked the supervised classification task to the encoder output for efficiency.

Main Results:

• The proposed semi-supervised model achieved superior performance compared to alternative methods on deception detection tasks.
• Experimental results demonstrated effectiveness even with a limited amount of labeled data.
• Performance was validated using the INTERSPEECH 2009 Emotion Challenge feature set on the CSC and a custom deception corpus.

Conclusions:

• The semi-supervised additive noise autoencoder model offers a promising solution for speech-based deception detection.
• This approach effectively addresses the challenge of limited labeled data by utilizing unlabeled data.
• The model provides a more concise and efficient method for deception detection with advanced performance.