Type of Document Dissertation Author Calix, Ricardo A. Author's Email Address email@example.com URN etd-06292011-153733 Title Automated Semantic Understanding of Human Emotions in Writing and Speech Degree Doctor of Philosophy (Ph.D.) Department Engineering Science (Interdepartmental Program) Advisory Committee
Advisor Name Title Knapp, Gerald Committee Chair Constant, David Committee Member Gunturk, Bahadir Committee Member Harvey, Craig Committee Member Houston, Andrea Committee Member Yu, Boryung Dean's Representative Keywords
- human computer interaction
- machine learning
- natural language processing
- cognitive science
Date of Defense 2011-06-13 Availability restricted AbstractAffective Human Computer Interaction (A-HCI) will be critical for the success of new technologies that will prevalent in the 21st century. If cell phones and the internet are any indication, there will be continued rapid development of automated assistive systems that help humans to live better, more productive lives. These will not be just passive systems such as cell phones, but active assistive systems like robot aides in use in hospitals, homes, entertainment room, office, and other work environments. Such systems will need to be able to properly deduce human emotional state before they determine how to best interact with people.
This dissertation explores and extends the body of knowledge related to Affective HCI. New semantic methodologies are developed and studied for reliable and accurate detection of human emotional states and magnitudes in written and spoken speech; and for mapping emotional states and magnitudes to 3-D facial expression outputs. The automatic detection of affect in language is based on natural language processing and machine learning approaches. Two affect corpora were developed to perform this analysis.
Emotion classification is performed at the sentence level using a step-wise approach which incorporates sentiment flow and sentiment composition features. For emotion magnitude estimation, a regression model was developed to predict evolving emotional magnitude of actors. Emotional magnitudes at any point during a story or conversation are determined by 1) previous emotional state magnitude; 2) new text and speech inputs that might act upon that state; and 3) information about the context the actors are in. Acoustic features are also used to capture additional information from the speech signal. Evaluation of the automatic understanding of affect is performed by testing the model on a testing subset of the newly extended corpus. To visualize actor emotions as perceived by the system, a methodology was also developed to map predicted emotion class magnitudes to 3-D facial parameters using vertex-level mesh morphing.
The developed sentence level emotion state detection approach achieved classification accuracies as high as 71% for the neutral vs. emotion classification task in a test corpus of children’s stories. After class re-sampling, the results of the step-wise classification methodology on a test sub-set of a medical drama corpus achieved accuracies in the 56% to 84% range for each emotion class and polarity. For emotion magnitude prediction, the developed recurrent (prior-state feedback) regression model using both text-based and acoustic based features achieved correlation coefficients in the range of 0.69 to 0.80. This prediction function was modeled using a non-linear approach based on Support Vector Regression (SVR) and performed better than other approaches based on Linear Regression or Artificial Neural Networks.
Filename Size Approximate Download Time (Hours:Minutes:Seconds)
28.8 Modem 56K Modem ISDN (64 Kb) ISDN (128 Kb) Higher-speed Access DissertationRCalix.06.29.2011.pdf 14.43 Mb 01:06:48 00:34:21 00:30:03 00:15:01 00:01:16indicates that a file or directory is accessible from the LSU campus network only.
If you have questions or technical problems, please Contact LSU-ETD Support.