Population Infinite: Runway ML
This week I trained a Machine Learning Model on RunwayML which imitates the way I talk by importing data from my Facebook. Creating a bot which imitates the way someone talk has been something I’ve always wanted to try. I chose to use myself as the reference point because I find it interesting to see a reflection of myself from an external perspective, especially from an algorithm which sometimes spits out incoherent sentences filled with words that I commonly use. Most of the sentences generated seem to be a satire of how I talk online due to the inclusion of words I commonly use inside sentences without coherent logic. I had a lot of fun seeing generating sentences with my model.
Searching for the dataset to use for training was the most difficult part of the process for me. I found several ways to extract messages from different social medias including facebook, instagram, weChat, and iPhone SMS. It was the most straightforward to extract text messages from the phone since there are several software which automatically extracts the text messages and export into PDF, but all of them require purchasing an expensive license to use the full features of the software, so I decided to try out other means of extracting my messages after going through several SMS extraction software. I also tried downloading my messages from instagram, which can be done directly from the website’s profile settings. However, Instagram only has the option to download all messages have in my account, and the resulting file was too big for me to work with in converting it into google sheets for organization/filtering, I just get a notification that tells me there are too many lines in the document when trying to import. Out of all the platforms I tried, Facebook provides the most customizations in downloading the account’s data, letting me pick the time range and format of the data I want to download. After downloading my messages from Facebook, I had to first organize and filter out on Google Sheets the unnecessary information from the file such as the name of people sending the messages and the timestamp for when the messages were sent. I ended up only importing conversation histories from some of my friends instead of all because the files are given in separate files and I had to manually import each one for filtering and exporting as text. I picked a few friends who I talked the most with for the data set.
Once I filtered out the unnecessary information and converted it into text file, I was able to feed it as a dataset into Runway ML. Below are some sentences generated using the model
