Hand-labelled training sets are expensive and time consuming to create usually.
Some datasets call for domain expertise (eg: medical/finance datasets etc).
Given these factors around costs and inflexibility of hand-labelling it would be nice
if there are tools which can help us get started quickly with minimal labelled dataset – enter weak supervision.
But what if you do not have any labelled data at all? is there a way to still label your data automatically in some way?
That’s where zeroshot_topics might be useful! to help you to be up and running quickly.
zeroshot_topics let’s you do exactly that! it leverages the power of zeroshot-classifiers, transformers & knowledge graphs to automatically suggest labels/topics from your text data. all you need to do is point it towards your data.
The algorithm contains, 4 stages:
- Keyword & Keyphrase extraction: This is done with the help of KeyBERT. but really any sort of keyword extractor can be used.
- Keyword/Keyphrase expansion via knowledge graphs/Taxanomy: Then we expand the important keywords we discovered by using some sort of taxanomy/knowledge graph like wordnet, conceptnet etc.
- Trace the Hypernyms for the keywords: Identify the Hypernyms(the root/parent word) and use this as the psuedo-label for the zeroshot classifier.
- Zeroshot classification: Use the Hypernyms and documents to label via zeroshot classifiers.
Note: Currently, this tends to work well on short-texts in general, in the future I intend to experiment and see how we can support long texts as well.
zeroshot_topics is distributed on PyPI as a universal
wheel and is available on Linux/macOS and Windows and supports
Python 3.7+ and PyPy.
$ pip install zeroshot_topics
from zeroshot_topics import ZeroShotTopicFinder zsmodel = ZeroShotTopicFinder() text = """can you tell me anything else okay great tell me everything you know about George_Washington. he was the first president he was well he I'm trying to well he fought in the Civil_War he was a general in the Civil_War and chopped down his father's cherry tree when he was a little boy he that's it.""" zsmodel.find_topic(text, n_topic=2) # Output - Topics: ['War', 'Head Of State']
Some things that i plan to add in the coming days, if there’s some interest in this work by the community.
- Support custom keyword extractors.
- Support Custom Knowledge-graphs & taxonomy.
- Support Custom Zeroshot-classifiers in the pipeline.
- Add Usecase examples & improve documentation.
- Optimise the overall library and make it a faster.
- Support Long Text documents.
zeroshot_topics is distributed under the terms of