Labelling Tool with a little bit of Active Learning

Web based labelling tool with active learning. Model is defined in LSTMNet.py

The trained model is trying to predict labels for your examples and ask for human labelling for these which he is most uncertain. Loading batch size is set to 20.

Data for labelling are not included because of licence. The tool is still in development but can be used to speed up labelling data.

Design 1.0

Usage

If you want to run in on your machine, download this repository and run:

export FLASK_APP=application.py; export FLASK_DEBUG=1; python -m flask run

Then you can access website at your localhost address. For example http://127.0.0.1:5000/

For own model see abstract_model.py USAGE

• Pytorch implementation: lstm_net_pytorch
  • Legacy code, slow
• Keras implementation: lstm_net_keras
  • Improved loading and creating model, more faster

Input format of data

what	0	0
are	0	0
you	1	B-pronoun
talking	0	0
about	1	B-generic_entity
.	0	0

let's	0	0
talk	0	0
about	0	0
politics	1	B-topic
.	0	0

Between each example is newline and each word is in this format:

what\t0\t0\n

Workflow

After labelling some data you can retrain your model (it does not save a model). Then you can test the model and choose to save new one on load the previous one. After saving there is no backup of old model (so by careful).

Todos

• Loading configuration from JSON
• Modular architecture
• Research for better models