[Week 1 — Emotion Detection]
Team members: Ali Baran Tasdemir, Akif Cavdar
As a part of being human, we all experience some emotions like anger, joy, fear, love, etc.. And we act differently for every emotional aura. For example, eating tons of chocolate bars when sad or cheering when happy.
There are some websites or applications which proposes product for your mood. Spotify creates playlists for your mood. What if Spotify can detect your emotions and recommend songs for you?
What is Our Project?
The goal of our project is what we write in this situations. In Facebook, we share posts under the effect of some feelings. And with a successful analysis, we give some good data about ourselves for companies. But text analysis (sentimental analysis) can be hard for machines.
We can analyze a sentence is positive or negative with our brain. For computers this process much harder than that. There are some approaches to this problem.
For our project, we will work on newspaper headlines. And determine this headlines corresponding emotions.
Solving Approaches
We have two main stages for this project. First, we need to make a text analysis. Second is we need to build a prediction method for our texts. For the first stage, we can use “keyword method”. This method we pick some keywords for emotions. And we can search these keywords in headlines.
But as mentioned before sentimental analysis is hard. There can be some negative nouns before some keywords or people can use irony. But we will try to maximize our analysis in this part to get the most accurate result at next stage.
The second stage is learning — predicting stage. We will use learning algorithms for our tokenized headlines and we build a classifier model. We are planning to use multi-layered neural networks.
Dataset and Related Works
We will use Affective Text — SemEval 2007 Task #14 dataset. This dataset contains 1250 headlines (labeled). We work with 6 main emotion: anger, joy, surprise, fear, disgust, sadness. 2 other files contain every header’s class values and headers positivity or negativity value.
1250 headers may not be sufficient for training neural networks and testing. Because of that, we are looking for more data.
There is a text analysis-tone analysis API published by IBM Watson.
And there are some researches about our topic.
