Twitter Sentiment Classification

Messages on social networks include many information about people’s opinion and feelings. So, developing tools for automatic sentiment classification of these messages seems to be attractive. In this project, we focused on Twitter messages. The goal is to predict whether a tweet is expressing a positive or negative feeling; therefore, it is categorized as the binary classification problems. We implement neural networks and deep learning architectures (Using TensorFlow).

Used Libraries

- Python 3.6.5
- numpy 1.15.4
- pandas 0.23.4
- Tensorflow 1.4.1
- keras 2.0.8
- scikit-learn 0.20.1
- h5py 2.8.0

Regenerate our best final result

To regenerate our final result from pre generated embedding data and pre trained model, you only need to execute "python run.py" This file download our embedding data for 300 dimension and our best model checkpoints (for LSTM). Afterwards, predict labels for test tweets data. The final .csv result is saved in results folder.

Dataset

The dataset is available from "https://nlp.stanford.edu/projects/glove/" page. Download the Twitter datasets (glove.twitter.27B.zip) and Common Crawl (glove.840B.300d.zip)

• Extract zip folder and put glove files in "data/glove_embeddings" path
• Put twitter datasets including train pos, train neg, and test tweets in "data/twitter-dataset" path

Setup

To run for every data and model. First, locate datasets in data folder. You can use train_predict.py to extract your own results. Next, specify {full, glove_dimension, max_words, model_name} parameters and then execute "python train_predict.py". Final results will be saved in results directory.

Folder structure of project

├── data # This folder includes twitter datasets and glove dataset

├── data_preprocessing # Functions for reading data, preprocessing, generating data from glove embedding matrix, and loading data

├── generated_parameters # This folder includes generated checkpoints from training of models

├── models # This folder includes the implementation of neural network models + RLR and SVM models as baseline

├── paths # json file to specify relative paths

├── results # This folder consists of .csv result file for submitting to crowdAI

├── README.md # README file

├── run.py # Script for running the best model and creating result file

└── train_predict.py # This file generate and load data, train the desired model, and finally generate prediction for test data

Training models

For this classification project, we have implemented 3 neural network models:

• LSTM
• CNN
• GRU

Author

• Mohammad Vahdat [email protected]