ner-lstm

Why we included this project

This is a small, readable implementation of a two-layer bidirectional LSTM for named entity recognition, built on TensorFlow with gensim for word embeddings. The pipeline is split into separate scripts for embedding generation, input preparation, training, and evaluation, so it is easy to trace how a sequence tagging model goes together, which makes it a good starting point for learning NER or reproducing classic results. The authors report F1 scores on the CoNLL 2003 English benchmark and the ICON 2013 Hindi dataset, and they compare word2vec, GloVe, and RNN-based embeddings, so you can see how the choice of embedding changes the outcome. It is a research codebase from 2016 rather than a maintained service, so treat it as material to study and adapt, not something to deploy as-is.