📱 Twitter Sentiment Analysis 📱

Build a deep learning sentiment classifier to determine the tone (positive, negative, neutral) of a tweet.

💡 RESULTS

Entering the Twitter username of a person, we get his/her last tweet with a Sentiment Score prediction.

Here are some 'Mood Checks' I did to test the model :

📝 PROCESS

SA_Building_the_Models_(Word2Vec, SA_NeuralNet).ipynb

• Scrap Twitter with GetOldTweets3 library
• Train a Word2Vec model for tweet vectorization
• Predict the sentiment score of the tweet (using a Neural Network model)

🔎 MODELS

Word2Vec Model

Word2Vec is a neural network model that transforms words into vectors (words embeddings).

We used the Gensim Word2Vec model and trained it using a corpus of tweets and there corresponding sentiment scores (http://help.sentiment140.com/for-students).

Once trained, the Word2vec model can understand semantic similarity as shown on the picture below.

TF-IDF Matrix

The Word2Vec model converts each word from the corpus into a high dimensional vector.

To classify tweets, we have to turn them into vectors as well.

We "combine" these words-vectors together to get the tweet-vector by computing a weighted average (TF-IDF score - importance of the word in the corpus).

Keras Neural Network model

An other neural network model is trained to predict the sentiment score for a given tweet.

It is a Keras Sequential model using sigmoid activation function for the output layer (always return value between 0 and 1).

We used EarlyStopping to prevent from overfitting and callback function (ModelCheckpoint) to save the best model during training.

🚝 PIPELINE

• Scrap the last tweet of a given Twitter username (GetOldTweets3 library)
• Tokenize the tweet => ['i', 'love', 'you'] (nltk.tokenize.TweetTokenizer())
• Build tweet vectorization using Word2Vec model and TF-IDF matrix
• Feed the scaled tweet vector to the Keras Neural Net model to get the prediction