{"id":23,"date":"2022-02-05T17:38:00","date_gmt":"2022-02-05T17:38:00","guid":{"rendered":"https:\/\/ghothi.co.uk\/?p=23"},"modified":"2025-05-28T14:15:47","modified_gmt":"2025-05-28T14:15:47","slug":"categorising-data-with-zero-shot-classification","status":"publish","type":"post","link":"https:\/\/ghothi.co.uk\/?p=23","title":{"rendered":"Categorising Data with Zero-Shot Classification"},"content":{"rendered":"\n
Recently, I needed to group about 1000 terms into a number of categories.<\/p>\n\n\n\n
Doing it manually was far too painful, so I turned to machine learning as it can be so good for these kinds of tasks.<\/p>\n\n\n\n
Specifically, I used the torch<\/code> and transformers<\/code> library from Hugging Face. Their pipeline<\/code> function makes setting up a model and performing various tasks really simple. In this case, I used the zero-shot-classification<\/code> pipeline, which lets us classify text into categories without needing a pre-trained model specific to our labels. In other words, someone else has already done the heavy lifting of training a model perfect for this use case.<\/p>\n\n\n\n
Given how much of a time-saving this was, I thought I’d record the process in case it’s useful for anyone else. This will be a simplification, but it should be simple enough to read between the lines.<\/p>\n\n\n\n
Zero-Shot Classification Explained<\/h2>\n\n\n\n
Unlike traditional machine learning models that need to be trained on specific pre-defined categories, zero-shot classification can categorise text into labels it hasn’t seen before (i.e. data it wasn’t trained on). It’s similar to having a grad student organise your stuff without needing to give them examples first, they just understand how the info needs to be organised based on the categories you tell them. The below image probably speaks a thousand words:<\/p>\n\n\n\n<\/figure>\n\n\n\n
This is really useful when you need to sort data into custom categories without having a huge dataset (or the time) to train a new model.<\/p>\n\n\n\n
This is all possible because of the ‘Transformers’ architecture – these are just models that understand and generate human language by understanding the meaning, context and relationship between words. We’ll be using the Hugging Face library, which gives us easy access to these pre-trained models.<\/p>\n\n\n\n
The Goal<\/h2>\n\n\n\n
Categorise a list of about 1000 terms into predefined categories using zero-shot classification to save me having to do it manually.<\/p>\n\n\n\n
Setting Up the Code<\/h2>\n\n\n\n
First, import the necessary libraries and set up our environment:<\/p>\n\n\n\n
![\"\"](\"https:\/\/ghothi.co.uk\/wp-content\/uploads\/2022\/02\/few-shot-classification.png\")
<\/figure>\n\n\n\n