NER with Deep Learning

This project uses Combination of Deep Neural Networks for Named-Entity Recognition Task. The Project implements the SOTA method proposed by (Ma and Hovy et al., 2016) in this paper. The implementation uses Keras 2.0 library with tesnsorflow backend.

Some differences from the original proposed approach:

The model architecture is a bit different from the original implementation in the following ways:

• A Dense Layer (100 units) has been added to the model for imporving performance.

• Hyperparameter optimization has done for imporved results and rich features learning.

Dataset:

• The data used for training is CoNLL 2002 dataset for NER and POS Tagging.

Training & Inference

• In order to train model, first update following paths in ner_tagger.py file:

TRAINING_DATASET_PATH = "./ner_dataset.csv"
NER_MODEL_OUTPUT_PATH = "./saved_model/My_Custom_Model3.h5"

• Run the following command in project root to train model and do inference:

python ner_tagger.py

• To avoid training model every time you run the code, comment the following line in ner_tagger.py file: NERTagger().train(TRAINING_DATASET_PATH)

• To run model inference on different text, change the value of variable texty defined on top of ner_tagger.py