blog/garment-parsing-gpt3/

[utterances-bot]

Parsing garment descriptions with GPT-3 • Max Halford

The task You’ll have heard of GPT-3 if you haven’t been hiding under a rock. I’ve recently been impressed by Nat Friedman teaching GPT-3 to use a browser, and SeekWell generating SQL queries from free-text. I think the most exciting usecases are yet to come. But GPT-3 has a good chance of changing the way we approach mundane tasks at work. I wrote an article a couple of months ago about a boring task I have to do at work.

https://maxhalford.github.io/blog/garment-parsing-gpt3/

[alexmolas]

Super-interesting post! I had to do something similar in my past job (classify cloth pieces in 10 different categories using their description). Of course my problem was much more easier than yours, and in fact with just a simple SVM we got pretty good results.

I want to try generating new training data by taking the samples I’ve already parsed, and formatting them to different text formats while adding noise

This sound cool! I would give it a try for sure, it sounds like an easy way to increase your training dataset. I guess the problem is how to generate noise in the same way as humans generate noise.

[MaxHalford]

Super-interesting post! I had to do something similar in my past job (classify cloth pieces in 10 different categories using their description). Of course my problem was much more easier than yours, and in fact with just a simple SVM we got pretty good results.

Nice! I'm supposing the content of each description are semantically similar to the categories. Maybe my article on unsupervised text classification would help if you're lacking labels.

I guess the problem is how to generate noise in the same way as humans generate noise.

That's a good point. I think there's no point generating templates which already have been seen by the model. I think it's useful to generate typos though, it'll make the model more robust. There's probably loads of ways to generate typos in words. For instance, one could take the code from Peter Norvig's spelling corrector article.

With regards to what templates to pick: when I receive a new dataset to parse, I could look at the templates which exist in that dataset, take all the existing parsed data I have, and format the latter according to the several templates I've identified. I'm idealizing a bit, but in theory I think this is a good way to bootstrap oneself using existing data.