I wrote an article titled “Sentiment Analysis Using Keras” in the November 2018 issue of Visual Studio Magazine. See https://visualstudiomagazine.com/articles/2018/11/01/sentiment-analysis-using-keras.aspx.

In the article, I use the standard IMDB movie review dataset and create a prediction model that will accept the text a movie review, and emit a value between 0.0 and 1.0 where a value less than 0.5 means a negative review (“this film is terrible”), and a value greater than 0.5 means a positive review (“great experience”).

Sentiment analysis is a very challenging problem — much more difficult than you might guess. The Keras library has excellent support to create a sentiment analysis model, using an LSTM (“long, short-term memory”) deep network.

I’m quite proud of this particular article because explaining Keras and LSTMs is difficult, and especially difficult to explain succinctly.

One aspect of the problem that I didn’t go into is data preparation. The Keras library comes with the IMDB dataset built-in and ready to go. This is a huge time saver. I’m implemented a sentiment analysis system starting from the raw IMDB data — 50,000 individual files with one review in each. Getting the data ready was a big effort and took me a full week of pretty concentrated effort.

As I point out in my article, “As recently as about two years ago, trying to create a custom sentiment analysis model wouldn’t have been feasible unless you had a lot of developer resources, a lot of machine learning expertise and a lot of time. Instead, you’d likely have had to use a canned approach from a service such as Azure Cognitive Services text analytics. But now, even though sentiment analysis is a very challenging problem, the existence of neural network libraries like Keras with built-in LSTM functionality has made custom sentiment analysis feasible.”