README.md

NLP Project

Implement CRF by hand based on 2 POS datasets

Setup

1. install virtualenv
2. run pip install -r requirements.txt

Instructions to run

• P1: run python P1.py <train file>
• P2: run python P2.py <train file> <dev in file> <dev out file>
• P3: run python P3.py <train in file>
• P4: run python P3.py <train file> <dev in file> <dev out file>
• P5 Bert: run python bert.py <train file> <dev in file> <dev out file> <lang>, lang is "EN" or "ES"
• P5 Bi-LSTM-CRF: run python bilstm-crf.py <train file> <dev in file> <dev out file> <lang>, lang is "EN" or "ES"
• P5 attention model: run python transformers.py <train file> <dev in file> <dev out file> <lang>, lang is "EN" or "ES"

Results:

P4:

Metrics	EN	ES
# Entity in gold data	210	235
# Entity in prediction	35	21
# Correct Entity	22	6
Entity precision	0.6286	0.2857
Entity recall	0.1048	0.0255
Entity F	0.1796	0.0469
# Correct sentiment	14	6
Sentiment precision	0.4000	0.2857
Sentiment recall	0.0667	0.0255
Sentiment F	0.1143	0.0469

Using BiLSTM-CRF:

Metrics	EN	ES
# Entity in gold data	210	235
# Entity in prediction	158	156
# Correct Entity	118	125
Entity precision	0.7468	0.8013
Entity recall	0.5619	0.5319
Entity F	0.6413	0.6394
# Correct sentiment	84	104
Sentiment precision	0.5316	0.6667
Sentiment recall	0.4000	0.4426
Sentiment F	0.4565	0.5320