It is an emotion identification project that tries to identify emotions from text namely : anger, disgust, fear, sadness, joy, surprise.
Some notable works had been done on classifying sentences based on emotions. But this one is ambitious. It assumes that a sentence can be a mix of emotions.
Currently fine grained evaluations are being done i.e. each of the emotion classes are further subdivided into high or low, eg. high anger / low anger.
Initially each of the sentence is annotated with a continious value [0,100] for each emotion.
A sentence can be therefore 30% anger, 20% disgust,10% fear and so on.
Fined grained classification treats [0,50) as low and [50,100] as high.
Will be enhanced for coarse grained evaluation where it may be possible to tell the quantity of emotion.
For each emotion, false positives are currently high (will be improved).
Every sentence is passed through a POS-tagger.
Only adjectives are taken as features.
TF-IDF scores are applied for each feature and occurence matrix is created.
The occurence matrix is then trained using softmax regression of Tensorflow.
Sentences are classified with fine-grained evaluation.
Lots of false positives. Feature selection needs to be improved. Dimensionality needs to be reduced.
True positive is high.
Coarse grained classification is to be done.