This is a data science project which analyzes different movie scripts and extracts the evolution of the emotional content throughout each movie ("emotional plotline").

The aim is then to cluster the movies based on their emotional plotline. A visualization dashboard helps with the exploration of these clusters.

Example output

An example of an interactive dashboard allowing the user to explore the results of clustering (3 clusters) is shown here.

Page: https://annavm.github.io/Project_Plotline/example/

Screencast:

the page can take a little while to load, you can then use the tools to select a point, see its name and the corresponding emotional plotline.

Screenshot:

It is recommended to use the Anaconda distribution, to install a set of standard required packages. Once Anaconda is installed, please type:

conda install numpy pandas matplotlib numba jupyter

The additional required Python packages are listed in the file requirements.txt. In order to install them, please type:

pip install -r requirements.txt

In addition, you will need to download the nltk corpus (i.e. the data which is needed for the Natural Language Processing package nltk). To do so, type the following code in a terminal:

python
>>> import nltk
>>> nltk.download()

Then click on Download on the graphical window.

If you wish to reproduce the analysis, here are the different steps to carry out.

The scripts are obtained by scraping the website IMSDb. You can automatically download approximately 1000 scripts from this website by running the code in code/scraping_script.py:

cd code/
python scraping_script.py

The code creates a directory data/scraping, where it stores the movie scripts, along with some metainformation.

For each movie script, the text is divided into windows of 100 consecutive words, and a quantified emotional content is associated to each window.

This is done by looking up each word of a given window in the Word-Emotion Association Lexicon, which associates words with 8 emotions (anger, anticipation, disgust, fear, joy, sadness, surprise, trust) and 2 sentiments (negative, positive). A copy of the NRC lexicon is stored in data/emotions/NRC_emotions.txt.

The code that extract the emotional content of each movie is in code/emotions_script.py. It can be run by typing:

cd code
python emotions_script.py

The code creates a directory data/emotions/arrays, where it stores the datapoints (as .npy) needed to trace the graph for each movie.

Option 1: To visualize the graphs, type:

cd code
python load_plotline.py

and answer yes (y) to the prompt "Do you want to save plots as png (y/n)?". The graphs will be stored in a directory data/emotions/graphs.

Option 2: To explore the data dynamically, open the corresponding Jupyter Notebook by typing:

cd jupyter
jupyter notebook Visualize_Emotions.ipynb

The dashboard generated with iWidgets allows an interactive view of the different plots. Screencast:

Screenshot:

The aim is to compare the evolution of emotions in two movies. This relies on a building a comparison tool to contrast a set of 'emotion' plotlines defining a given movie to the set obtained for another. The approach retained here is based on Dynamic Time Wrapping, which calculates the pairwise distance between all the movies.

The code that returns a dictionary containing the pairwise distances in pickled form is in code/dtw_script.py. It can be run by typing:

cd code
python dtw_script.py

The dictionary of pairwise is stored in data/distances.pkl, the lookup structure is as follows: distances[filename1][filename2].

A Jupyter Notebook, jupyter/Explore_closest_movies.ipynb, is available to give an easy access the top 10 closest movies to a selected movie.

The motivation is to group the movies according to the evolution of emotions in their scripts. This is achieved thanks to the pairwise distances calculated previously and a modified Kmeans clustering algorithm called medoids (instead of taking the mean as the prototype of the cluster the median is retained). As with any unsupervised algorithm, assessing the performance of the clustering is not straightforward. Here, I develop 2 ways to investigate the results of clustering: first observing the cost associated with a given number of clusters (option 1), second, analyzing how reproducible the clustering is (option 3).

The code code/medoids.py has 3 main features:

option 1: picking the number of clusters

cd code
python medoids.py pick_k

option 2: running the k medoids algorithm (here for 3 clusters)

cd code
python medoids.py k=3

option 3: investigating the reproducibility of the clustering is (change 3 to number of clusters chosen)

cd code
python medoids.py 3_stability

The most meaningful clustering occurred for k = 3, as the vast majority of movies stay in the same clusters.

I build an interactive scatter graph with linked emotional plotline thanks to the mpld3 package (http://mpld3.github.io/). This package is a great way of combining Javascript, D3js and python. I used an example of custom plugin to write the appropriate D3js script to develop a tool where the plotlines and the movie name appear on hover.