Personal tools
You are here: Home SIGs and LIGs Speech SIG Non-Prototypical-Emotion

Non-Prototypical-Emotion

Initialisation of the Virtual Panel Discussion on Recognition of Non-Prototypical Emotion

Many hurdles have already been taken when it comes to integration of emotion recognition engines and systems into real-life applications: test and training on natural and spontaneous data, independency of the subject, coping with noise and variety as non-limited text in the case of speech processing, to name just the most relevant. However, one crucial issue is usually still neglected: testing models and engines on non-prototypical cases: the ground truth with respect to annotation is generally formed by a majority vote of labellers. By that, cases with low agreement are simply discarded. Yet, an application cannot afford such luxury: it has to process everything that "comes in", whether it is call centre data analysis, user interaction, media retrieval, or – let alone – surveillance. Thus, we believe that many reported results tend to be over-estimates, and new paradigms might need to be investigated: detection over classification, including consideration of "out-of-emotion-vocabulary" events, and reliable, non-trivial confidence measurements. Thereby evaluation may change (or be added) from reporting accuracy or F-measures in the case of categorical description of emotion, and mean linear error and cross-correlation in dimensional modelling, to Receiver-Operator Curves, Equal-Error-Rates or the Area Under Curve and related measures. While this problem generally touches all modalities suited and used for analysis, we limit the topic to speech processing – to not lose focus and stay within our own main expertise.

As we believe that it is important to initiate a discussion on this topic – we ask you to contribute your oppinion and experience and set the floor for a virtual panel discussion. So please make sure to mail us your contribution:
  • Did you do experiments with non-prototypical emotion, yet?
  • Which measures do you consider best (ROC curves, EER, accuracy, ...)?
  • Do you consider the representation in the emotion space sufficient in this respect?
  • ...
Thank you!

Björn Schuller, Anton Batliner, and Stefan Steidl
Document Actions
Powered by Plone

Portal usage statistics