ML for classifying deceptive speech from audio/text

• Original data is labeled transcribed sequences of Deceptive/Not Deceptive speech audio from a (few hundred?) people
• Data is broken into 3 types: Acoustic, Lexical, and Personal
• Acoustic and Lexical are at two granularities: IPU level, and Turn level
  • Acoustic: OpenSMILE (IS13) acoustic/prosodic feature extractor
  • Lexical: DAL lexical sentiment feature extractor
  • Lexical: LIWC lexical sentiment feature extractor
• Personal is collected per person.
  • These are NEO-FFI personality indicators: Neurotiscism, Extraversion, Openness, Agreeableness, and Conscientiousness
  • Also the persons gender
  • And whether their native language is Mandarin Chinese or American English (1/0?)

Given train / dev/ test IPU splits:

** Will combine training and dev and do cross validation to have roughly 60/40 split **

Naive baseline accuracy: 62.12% train, 59.04% test

• Logistic Regression w/
• L1 regularization
• L2 regularization
• Grouped LASSO -- we should discuss what groupings make sense
• Random Forest
• Neural Sequence model
• Maybe do some simple dimension reduction in preprocessing w/ Chi Square or PCA?