Show HN: Active learning and model explainability for document classification
github.com
github.com
I want to take in massive sums of articles from a news API which will be placed in their corresponding file based on where my classifier says I should put them. I have to generate my own labeled data for this. That is a problem. Most people don't realize that the sample efficiency in models which utilize transfer learning is so great that AI-assisted data labeling is extremely useful and can significantly shorten what is ordinarily a painful data labeling process.
We need a way to quickly create word embedding powered document classifiers which learn with a human in the loop. For some classes, an extremely limited number of examples may be all that is necessary to get results that a user would consider to be successful for their task.
I want to know what my model is learning - so I integrate the word embeddings available with Flair, combine with Classifiers in Sklearn and Tensorflow/Keras/PyTorch, and finish it off with a nice squeeze of the LIME algorithm for model interpretability (implemented within the ELI5 Library)