- From: Burkhardt, Felix <Felix.Burkhardt@t-systems.com>
- Date: Wed, 27 Sep 2006 14:00:15 +0200
- To: <public-xg-emotion@w3.org>
Dear all, here comes a usecase for the Emotion ML derived from our work in anger-detecting voice portals following Marc's format in [1]. Sorry it's kind of longish ... belonging to Use case 1: Annotation of emotional data Use case: Anger Annotation in Voice Portal Recordings # description: Felix has a set of Voiceportal recordings, a statistical classifier, a group of human labelers and a dialog designer. The aim is to define a training database for the classifier that gives the dialog designer a detector of a negative user state ("anger"?) in two stages so that he/she can implement dialog strategies to deal with the user's aggression. # issues: The following issues need to be handled from this scenario: a) The classifier needs each recording labeled with (the degree of?) the anger-related state. b) Because a sensible dialog strategy depends on the user state, the dialog designer and the human labelers (which decide which emotional state to tag to which recording) need to have a fine agreement on the emotion terms and their meaning. c) Defining emotional states is a subjective task and depends strongly on the labeler, therfore the group of labelers should (at least partly) annotate the same recordings, the labels have to by unified. d) Also there are many words connected with negative emotions and the labelers need a common bag of words (that can be boiled down for the classifier later) and a common understanding of their meaning. e) The training data should be as large as possible and human labour is expensive, so the labeling process should be fast and easy, near real-time. f) For the dialog design, beneath to know IF a user is angry, it's even more important to know WHY the user is angry: is the user displeased with the dialog itself, e.g. too many misrecognitions? Does he hate talking to machines as a rule? Is he dissatisfied with the company`s service? Is he simply of aggressive character? This should be labeled also. g) often voice portal recordings are -not human but DTMF tones or background noises (e.g. a lorry driving by) -not directed to the dialog but to another person standing beneath the user this should be labeled also. h) the classifier might use human annotated features for training, e.g. transcript of words, task in application, function in dialog, ... this should be annotated also i) the training data should be usable for several dialog applications # requirements: (probably missing some things ;-) scope of the emotion annotation: refer to one voice portal recording emotion description: emotion categories: set of labels, with intensity (from 0 to 1) several labels for each recording (originating from several labelers) other: make connection with various other annotations, including meta-data about context regards, Felix [1] http://www.w3.org/2005/Incubator/emotion/wiki/UseCases Dr.phil. Felix Burkhardt T-Systems Enterprise Services GmbH Business Unit Technologiezentrum - Standort Berlin Goslarer Ufer 35, 10589 Berlin Tel. +49 30 -3497 2218 Mobile: +49 151 16710189 Fax +49 521 -92100512 E-Mail: Felix.Burkhardt@t-systems.com Internet: <http://www.t-systems.com>
Received on Wednesday, 27 September 2006 12:00:30 UTC