Named_Entity_Recognition_ML_experiments

Description

An exploration of machine learning approaches to Named Entity Recognition (NER) including classical classifiers as well as Neural Network (NN) and embedding techniques

This was a portfolio project for a course on Machine Learning for NLP.

Neural net consists of a feed forward model with: 1 Dense layer- relu, 1 Dense output - softmax, 1 hidden layer Dense - relu batch size = 50 epochs = 6 optimizer = 'adam' loss = categorical_crossentropy

Run

Run from commandline:

python .\main.py

3 arguments can be specified (but need not be): argv[1] argv[2] argv[3]

use arguments to pass in path to:

• training data;
• test data;
• embeddings (word2vec GoogleNews-vectors-negative300)

defaults assume everything is stored in local folder. E.g.:

• train_file = 'reuters-train-tab.en'
• test_file = 'gold_stripped.conll'
• embed_file = './GoogleNews-vectors-negative300.bin'

These can also be set inside the petercaine_program.py file in the first section.

Alternative paths

3 possibilites are available:

• use embeddings (y/n)
• use neural net (y/n)
• select algorithm (initials given at prompt) press Enter to indicate selection is finished

3 options are available 2 with embeddings; 1 without.

(1) tokens and features as data - using machine learning algorithms

• recommended features are (copy/paste into input): token pos chunk stem caps prev_caps short_shape prev_short_shape next_short_shape shape prev_shape

(2) embeddings and features as data - using svm

• recommended features are (copy/paste into input): pos chunk caps prev_caps short_shape prev_short_shape next_short_shape shape prev_shape

(3) embeddings and features as a neural net.

• Features are (no user choice involved): pos chunk caps prev_caps short_shape prev_short_shape next_short_shape shape prev_shape