Table of contents:
If you know the clothing company Taylor Stitch, then some first thoughts about them might include distinct product photography, a connection to California, and some really unique fabrics -- at least that what's I tend to think of. After a while I also started to notice a very consistent set of product names and a pretty recognizable writing style used for their marketing. Taking all of this in, I wondered if I could train a NLP model to write product descriptions for real or made up Taylor Stitch product names.
Take this real product description as an example:
The Jack in Prussian Blue Oxford
The Oxford button-down is a stylish wardrobe staple with endless versatility, so it’s no surprise that our signature Oxford, The Jack, has become a mainstay of our roster. Its burly 100% organic cotton basket weave and double-needle felled construction make it sturdy enough to stand up to the roughest workday, while peerless tailoring and a clean, pleatless design means it’s snazzy enough for your next night on the town. We’ve given this classic Prussian Blue iteration a thorough garment dye and wash for a soft, lived-in feel on the very first wear.
A few things here: the product naming conventions are always [product name] in [colorway], and note how the description focuses on the durability of the product, its history within the company, and the wide variety of the product's use cases. I won't pretend that these sentiments are exclusive to just Taylor Stitch's marketing, but I've found that the tone and consistency of these product descriptions make for what could be a fun modeling problem. Put more broadly: with enough examples, is it possible to train a NLP model to generate fake marketing descriptions in same writing style?
In addition to the consistent product names and descriptions, Taylor Stitch's website made gathering product descriptions to be relatively easy. They have a very accessible archive of nearly 1,000 of their products, and otherwise a simple website to scrape. With just a few URLs, I was able to scrape all the info I needed to create nearly 1,200 training examples from old and in-stock products alike. I've made my current version of the training data file available in ref_files/product_info.csv; the original script I used to create this file is titled info_scrape.py and I created an interactive notebook which allows for regular updates of the training data, titled update_text_file.ipynb
For the actual model training, I decided to take advantage of many of the already sophisticated NLP models out there, and focused on fine-tuning a light version of GPT-2 provided by Hugging Face. In other words, this became a transfer learning problem. I'd also like to give credit to François St-Amant's great article and whose code served as a foundation which I then adapted to my needs. And while there are more powerful versions of GPT-2 available, the lightest version is nonetheless >500mb! As such, having a functional model which was also on the lighter side was important to me.
OK, so how good is the model? Why don't we see with some examples, all of which you can find in ref_files/sample_output.txt...
For a real product that was not used for training:
The Jack in Chestnut Oxford:
Originally conceived as a companion piece to our signature Heavy Bag Polo, our Oxford jacket has developed an incredibly soft, supportive softness that can be trusted to stay on for years to come. Built from rugged 100% organic cotton and beautifully textured with a subtle speckling at the collar, The Jack is perfect for weekends, camp, or any other time of the day. With a button down collar and button up sleeves, The Jack is just as cozy as ever—tough enough to dress down to your favorite sweater, jeans, or whatever you wear them for at night.
Not bad, I guess? Note that the model got quite a few things wrong. It was asked to write about a shirt (not a jacket), and I don't know about "dress[ing] down to your favorite sweater..". Still, I'm quite pleased to see how it hit several of the original beats in the real product description above: there was a mention of 100% organic cotton, some talk about durability, and even a mention of different use cases! All while having a tone consistent to the brand's actual writing. Think it's overfitting? Here's another take:
The Jack in Chestnut Oxford:
Inspired by early pulp and hardwood shirting, The Jack in Chestnut Oxford is our take on the classic British shirt. This ultra-soft, comfortable oxford iteration is made from 100% organic cotton and is exceptionally soft—so you won't feel the cold at the edges, for example. Our collaboration with the renowned camp collar expert Dan Taylor of Pantyhose to develop the chunky Oxford stitch means it's constructed from a very special blend of organic cotton and 100% recycled yarns that’ll hold up well for years to come.
All right, the mention of the cotton is back (and more than once) and the durability is mentioned again. It's even cool to see the attempt to chacterize the Oxford as a "British" shirt (Even if it that isn't exactly true.) However, this example is definitely not as clean, so it's safe to say that while the tone and some main points are there, I wouldn't exactly call this one a compelling descripton. Just ask "renowned camp collar expert Dan Taylor of Pantyhose".
Of course what makes this fun is that you can get the model to write about anything. For instance, how would James Dean's iconic jacket from Rebel Without A Cause be described? Here are two takes for "The James Dean Jacket in Racing Red":
1/2: While our namesake may have emerged from the shadows of the pulp pulp pulp industry, The James Dean has emerged as one of the most recognizable products out there. Built from 100% recycled materials and equipped with rugged construction, it’s built to be taken on outings, backcountry, and on hikes, so you can take a breather when the clouds start to recede. Now, that’s even more impressive when you consider that The James Dean Jacket is made in Mexico, not California.
2/2: With so many stylistic options out there, it's no surprise that the legendary silhouette of The James Dean is renowned for its refined, clean lines and rugged construction. With its brushed chest pocket, short hem and comfortable comfort, The James Dean Jacket is a true asset to your wardrobe.
Once again, it's a bit hit or miss. That first one? Not so great. But to my surprise, I'm quite pleased with how the second one turned out. Regardless, I think it's pretty cool how the descriptions still have a writing style consistent with the brand's marketing and include common elements such as nods to the history or the legacy of the product, the materials, product durability, and a use case or two.
Interested in generating some descriptions of your own? Clone this repo and give it a shot. One caveat is that you will need to create virtual environment specific to the model and I've provided everything in a .yml file for your use. Note the trained model is ~500mb.
Once everything is ready to go, product generation is simple:
From generator import Generator
sample_products = ['The California in Rust Geo',
'The Short Sleeve Hawthorne in Navy Pique',
'The Thor Hammer in Natural Uru']
writer = Generator()
writer.write(sample_products,num_attempts=2,when_ready=True)
Note that all products follow the [product name] in [colorway/material] naming pattern. This is important as the model was trained specifically using this template for each prompt. If you decide to try it out for yourself, please add any descriptions you particularly like to ref_files/sample_output.txt and initiate a pull request. I'd love to have that file as a collection of everyone's favorite entries.
Finally, I am planning on adding a few more features such as the opportunity to train the model a little further. Also as more products are introduced I can add them to the training file so the model can get more examples to learn from. Still, I don't think it will ever be perfect and that's the beauty of it. A model which creates near perfect examples every time would be boring anyway. Instead it's the unexpected nature of all these descriptions that makes it entertaining.
Any other questions? Create an issue and we'll go from there.