[Week 7 — Emotion Detection]

Team members: Ali Baran Tasdemir, Akif Cavdar

We come to end for our project this week. If this project topic is interesting for you, you can read our story from the beginning.

So for wrap up all the things, we have done so far; Our aim for this project is detecting (classifying) emotions from sentences. We used SemEval 2007 Task 14 dataset at first but we decided to use ISEAR dataset after the 3rd week. Because SemEval dataset was too small. We used some natural language processing tools to prepare our data for some learning algorithms.

Our first learning algorithms were the basic ones. Firstly we tried Multinomial Naive Bayes. We get low accuracy rates with basic learning algorithms. Our best results were with SVM.

Multinomial Naive Bayes (left), SVM (right)

After this results, we decided to try some deep learning methods for our problem. And we implemented CNN and RNN. We added LSTM layer to CNN to try increasing success rate. We implemented 2 different neural networks.

CNN with one LSTM layer (left), RNN (right)

As we see the results above, we can say that some emotions are easier to separate from others. For example, joy and fear. But feeling like anger and guilt are sometimes very similar and very challenging to separate these feelings. For these emotions, our algorithms predict rate is lower. But for fear, disgust, joy, sadness predictions are quite successful.

The details of these networks will be on project presentation. And here is our project’s video presentation.