NLPatl (NLP Active Learning)

This python library helps you to perform Active Learning in NLP. NLPatl built on top of transformers, scikit-learn and other machine learning package. It can be applied into both cold start scenario (no any labeled data) and limited labeled data scenario.

The goal of NLPatl is to make use of the state-of-the-art (SOTA) NLP models to estimate the most valueable data and making use of subject matter experts (SMEs) by having them to label limited amount data.

At the beginning, you have unlabeled (and limited labeled data) only. NLPatl apply transfer learning to convert your texts into vectors (or embeddings). After that, vectors go through unsupervised learning or supervised learning to estimate the most uncertainty (or valuable) data. SMEs perform label on it and feedback to models until accumulated enough high quailty data.

Installation

pip install nlpatl

or

pip install git+https://github.com/makcedward/nlpatl.git

Examples

• Quick tour for text input
• Quick tour for image input
• Custom Embeddings, Classification, Clustering and Learning function

Release

0.0.2, Dec 17, 2021

• [Completed] Transformers supports Tensorflow
• [Completed] Performance tuning during clustering
• [Completed] Support multi-label
• [Completed] Custom Embeddings, Classification, Clustering, Scoring(Learning) function
• [Completed] Support TorchVision for image embeddings
• [Completed] Support SentenceTransformers
• [Completed] Add Least Confidence Sampling and Most Confidence Sampling
• [Completed] Add Semi-supervised learning
• [Completed] Add Farthest (Clustering) Sampling, Mismatch (Uncertainity) Sampling
• [Completed] Add Mismatch-farthest Learning

Citation

@misc{ma2021nlpatl,
  title={Active Learning for NLP},
  author={Edward Ma},
  howpublished={https://github.com/makcedward/nlpatl},
  year={2021}
}