Text Style Transfer

Not bad aside from the repeated a

Inspiration

Image style transfer has been around since 2015/2016 and is a powerful tool for manipulating images. I wanted to create something similar for sentences and phrases.

What it does

Given a content sentence and style sentence, the script produces a final sentence.

How I built it

I took a pretrained sentence embedding model called InferSent along with precomputed word vectors fastText and iteratively updated a random sentence tensor to minimize the content loss, style loss, and regularization term. I use Flask to make a frontend for the script

Challenges I ran into

Nothing like this has been done before. In the literature, style transfer is done using different techniques centered around GANS. Text models are also wayyy behind their vision counterparts, making this a much harder problem. For example, they don't understand grammar so generated sentences are often unnatural. Also I have never used Flask before.

Accomplishments that I'm proud of

Stacked word vectors seem to be a decent representation of style, especially the flow of a sentence. In particular, the final sentence usually captures the flow of the style sentence well, especially if the style sentence is distinctive. Introducing a regularization term prevents the final sentence from repeating words. Also figuring out how to use Django/Flask (I initially wanted to use Django) last minute.