Upload
primrose-bailey
View
221
Download
4
Tags:
Embed Size (px)
Citation preview
Emotion Detection in Email Customer Care
Narendra Gupta, Mazin Gilbert, and Giuseppe Di FabbrizioAT&T Labs - Research, Inc.
2010 ACL
Outline
[1.Introduction] [2.Emotion detection in emails]
2.1 Classifier
2.2 Feature extraction [3.Annotation] [4.Experiments and evaluation]
1.Introduction
Emotional emails often contain complaints about unfair treatment due to negligence, incompetence, rigid protocols, unfriendly systems, and unresponsive personnel.
Reduce the churni.e., retaining customers who otherwise would have taken their business elsewhere.
1.Introduction
A negative emotional component articulated by phrases like:you suck, disgusted.
Enumerating factual sentences such as:you overcharged, take my business elsewhere.
Outline
[1.Introduction] [2.Emotion detection in emails]
2.1 Classifier
2.2 Feature extraction [3.Annotation] [4.Experiments and evaluation]
2.Emotion detection in emails
We use supervised machine learning techniques to detect emotional emails.
2.Emotion detection in emails2.1 Classifier
We used Boostexter as text classification.
The output of the final classifier f is:i.e., the sum of confidence of all classifiers
The final classifier f can be mapped onto a confidence value between 0 and 1 by a logistic function;
2.Emotion detection in emails2.1 Classifier
Training model:
2.Emotion detection in emails2.2 Feature extraction
Salient features For our data we have identified the eight features listed below.
2.Emotion detection in emails2.2 Feature extraction
2.Emotion detection in emails2.2 Feature extraction
2.Emotion detection in emails2.2 Feature extraction
In the labeling guide for critical emails labelers were instructed to look forsalient features in the email and keep a list of encountered phrases.
We further enriched these lists by: a) using general knowledge of English, we added variations to existing
phrases. b) searching a large body of email text (different from testing) for different
phrases in which key words from known phrases participated.
For example from the known phrase lied to we used the word lied and found a phrase blatantly lied.
Outline
[1.Introduction] [2.Emotion detection in emails]
2.1 Classifier
2.2 Feature extraction [3.Annotation] [4.Experiments and evaluation]
3.Annotation
1) Two different labelers.
2) Kappa : 0.814.
3.Annotation
Outline
[1.Introduction] [2.Emotion detection in emails]
2.1 Classifier
2.2 Feature extraction [3.Annotation] [4.Experiments and evaluation]
4.Experiments and evaluation
We used cross validation (leave-one-out) technique on the test set.
4.Experiments and evaluation
4.Experiments and evaluation