As an assignment of AI course, we implemented this text classifier. The link for the data used during development and tests can be found below. There are distinct reviews about different movies, ranked as positive or negative, so we can actually understand whether our algorithm works well or not. We only used data from neg and pos folders for supervised learning. Also, we splited up the train folder randomly; 90% of the data was used during training and 10% for the development. Inside the code you will find m and n hyper-parameters, the meaning of which is: From the current dictionary skip the first n words and then keep the following m words. We tried different values for them and we also run the train and development process for different percentages of the total data.
This is the learning graph of our algorithm for m = 100 and n = 300:
Also here is the graph showing the best Recall, Precision and F1 values, reached by our algorithm for m = 200 and n = 100:
