"Rom-com or not rom-com" is an ML.NET project designed to run a machine learning model over a text document and determine if it exhibits the classic characteristics of a rom-com film script (or a horror, heist, comedy or science fiction film).
It was inspired by Mark Kermode's Secrets of Cinema excellent series on the BBC, which examines classic film genres, such as rom-coms, comedies, etc, and explores the structures, formats, film techniques and plot points that are common to each of these genres.
In the project we take a simpler, alternative AI approach to this same idea, and perform textual analysis on a load of movie scripts, by genre, and then use a machine learning algorithm to codify this, so that we can then take any arbitrary piece of text and rate it for the likelihood of it being in one of these genres.
If you wish to see the offical video of Phil and I presenting a session on this project, see An Introduction to Machine Learning with ML.NET
So, with a big 'Hello to Jason Isaacs', let's see how it works....
This project is pretty easy to get up and going, but does assume you have a good working knowedge of C# and .NET core, and it is a great way to get more familiar with ML.NET. At its core, this AI project boils down to a three stage process:
- Run the Script Builder to churn through all the film scripts and create a single TSV file with the script text and genre type for each film, one per line
- Use Model Builder to create and train an ML.NET machine learning model (and do a quick prediction check at the same time)
- Deploy the model behind .net core web api, PredictAPI, that allows a HTTP POST of a body of text for analysis using the saved machine learning model, and returns a probability rating for each genre type
The project comes with a simple jquery web page that can be used to call the api.
This project, as of Novemeber 2019 was been updated to .net core 3.x, so please make sure you have the relevant SDKS etc installed.
The Script Processor takes a pile of movie script file, in the default "Scripts" directory and tidies them up to create a single Tab-Separated Values (TSV) file. It is vital that each movie script in the scripts folder is named with the following convention
X_move name.txt
where 1 - rom com 2 - horror 3 - heist 4 - comedy 5 - science fiction
eg 1_500 days of summer.txt
We can't legally share or distribute film scripts, but there are many websites on-line that contain copies of popular film scripts, for research and the like.
The output TSV file has two columns
- Label (001 for rom com, 002 for horror, etc.)
- Key Phrases - the script as a single line with extraneous stuff removed
Originally we performed text analytics at this stage, but ML.NET has this capability built in, so so we now just use this process to create a single TSV file
There are some dummy text placeholders in this directory to start you off - simply delete them and replace them with the real film scripts.
At the command line, run this command
dotnet script .\ScriptProcessor.csx
Depending on the scripts you have sourced, you should see something like this:
Having created the data.tsv file, the next stage is to build a machine learning model. We use a multi-class classification model, which will give us a probability from the model for each genres. This is essentially like running a binary classifier for each of the genres in turn.
If you haven't worked with this type of model, the Sentiment analysis sample is a classic example of a binary classifier, which predicts if a piece of text has a positive or negative sentiment, based on supervised learning data set that has already been classified. The tutorial Categorize support issues using multiclass classification with takes you through the steps to create a multi-class classification model in more detail.
To use the model builder, at the command line run this command
dotnet script ModelBuilder.csx [train | load] 'some words to test'
The fist time you run this command, or when the data.tsv file has been updated, you need to run the command with the "train" option, so it builds and saves the model. As this can take several minutes depending on the szie of the data, subsequent runs can use the "load" option to load the previously saved model.
Train:
Load:
Notice how Model Builder also does a predictive test on a phrase you can pass in on the command line, as a quick test to validate the model with a probability rating for each genre. As you can see from the above screen shots, there is a very high probabilty that this is rom-com, with a small chance it might be a horror.
The final step is a .NET Core webapi application that can be used to take HTTP Posts and run the machine learning model on the body of ther request, allowing us test a more substantive szied chunk of text.
The applciation is a simple MVC application with a Predict controller which has a single POST method. It returns a JSON string with the probability data for each genre.
You can test the application with programs like Postman:
The final stage is to deploy the Web API with the model to Azure. There are a variety of DevOps ways to do this, but one of the simplest is to use an Azure AppService, and to use its Deployment Centre to do all the heavy lifting.
Once deployed, use the front end script project, which is a simple Javascrip web interface, to enter some text, and call the API, such as this from "Dawn of the Dead":
