AFFECTIVE COMPUTING

Extracting emotion from the language and discourse information

 

Another way to deduce an emotion out of the utterances is to look at the content of the speech, because people use different words when expressing different emotions.

 

 

Emotionally Salient Words

 

One way to evaluate the content of the utterances for emotions is to look at each word of the utterance for the “emotionally salient word” and analyse the salient of those words.

 

The “emotionally salient words” of each emotion are the worlds that appears more often in that certain emotions compare to the other. The underlying idea behind calculating the salience of the word is to calculate the self mutual information.

 

The formula for calculating it is given by:

 

 

 

 

 

 

 

where,   ek is the emotion

            vn is the word

W= { v1,v2 .. vn} = words in the utterances

E = {e1,e2 .. en} = emotion classes

 

By evaluating the database for the probability P(ek | vk) and P (ek), the self mutual information will be positive when the word makes the emotion more likely to happen. This is because the P(ek | vk), or in other word, the probability of emotion (ek) given the word (vk) will be greater than the P(ek) – probability of emotion (ek) given nothing. And so the log of more than 1 is positive and less than 1 is negative. This shows that if the self mutual information is positive it means that the word correlates to the emotion.

 

The emotional salience can then be calculated by:

 

 

 

 

 

 

this is the example of words and its emotional salience with only two emotions classes (negative and non-negative)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Given a sentence input, each word can be filtered by the dictionary of the emotionally salient words, and the emotional salience can be extracted and used as a feature for a classification of emotion of the sentence.