Grid Search for twitter sentiment detection model

Project aims a basic hyperparameter search on RCNN model for text classification task. Models are taken from brightmax's text classification models.

Getting Started

Run the main file and add your embedding into a04_TextRCNN folder with name "wiki.tr". Format should be txt file. Following word2vec is used. https://github.com/facebookresearch/fastText/blob/master/pretrained-vectors.md

Prerequisites

Tenorflow 1.10.0 gensim 3.5.0 word2vec 0.9.4 For conda package manager

conda create --name tw_sentiment

conda install --name tw_sentiment tensorflow=1.10.0

conda install --name tw_sentiment gensim=3.5.0

conda install --name tw_sentiment word2vec=0.9.4

Any instance of skopt library import can be eliminated from code if gives error.

To download word2vec wget -P 'directory_of_twitter_sentiment' https://s3-us-west-1.amazonaws.com/fasttext-vectors/wiki.tr.vec

##Usage Activate your environment in anaconda shell by

activate tw_sentiment

navigate the each model's main.py file and write:

python main.py

Make sure that python is added to your path variable.

Tests

F1 scores for RCNN model is obtained as in the table. Unfortunately, no worthy progress is obtained for other models.

Table	Negative	Notr	Positive
experiment: learning_rate= 0.0005, batch_size= 20	0.69	0.32	0.35
experiment: learning_rate= 0.0005, batch_size= 40	0.61	0.28	0.30
experiment: learning_rate= 0.0005, batch_size= 80	0.59	0.25	0.24
experiment: learning_rate= 0.001, batch_size= 20	0.74	0.46	0.41
experiment: learning_rate= 0.001, batch_size= 40	0.71	0.45	0.37
experiment: learning_rate= 0.001, batch_size= 80	0.70	0.41	0.35
experiment: learning_rate= 0.01, batch_size= 20	0.42	0.33	0.35
experiment: learning_rate= 0.01, batch_size= 40	0.48	0.29	-
experiment: learning_rate= 0.01, batch_size= 80	-	-	-

Tests are done with dataset: 1509 negative sentence 1002 notr sentence 987 positive sentence

95% percent of the data allocated for training, remaining is for testing.

Pretrained word2vec is with shape(200.000x300)