Joint Event-Relation Extraction using Encoder Decoder Architecture

Natural Language Processing Course Project of Group 19 at IIT Kharagpur (Autumn'21). For a detailed description of the methodology and results, please view the Project Report.

Group Members:

• Angana Mondal (19IE10039)
• Abhikhya Tripathy (19EC10085)
• Debaditya Mukhopadhyay (19IE10036)
• Aditya Basu (19IE10002)

Introduction

Joint-event-extraction is a significant emerging application of NLP techniques which involves extracting structural information (i.e., event triggers, arguments of the event) from unstructured real-world corpora. Existing methods for this task rely on complicated pipelines prone to error propagation.

Model Architecture

An encoder-decoder based architecture for joint entity-relation extraction was proposed by Tapas Nayak et al., and we further develop the architecture to deploy it to predicting trigger, argument and relation tuple (including the classes of the trigger and argument). We also utilise pretrained BERT embeddings to preprocess our data.

Datasets

The data is available at: https://drive.google.com/drive/u/1/folders/1fYP9PUQYRV0JWBa-N3CwuGkOCeBielT9
To obtain the Word2Vec embeddings and BERT embeddings, download 'w2v.txt' and 'BERT_embeddings.txt' from the aforementioned link.

Requirements

• Python 3.5 +
• Pytorch 1.1.0
• CUDA 8.0

Running the Code

• Python3: python3 Joint_Event_Extraction.py gpu_id random_seed source_data_dir target_data_dir train/test w2v/bert
• IPython: Run individual cells of NLP_Proj6_Grp_19.ipynb

Command Line Arguments

• Source_data_dir: Path to source data directory
• Target_data_dir: Path to target data directory
• train/test: Job mode (Choose only one of the two modes at once)
• w2v/bert: Embedding type

Default Command Line Arguments for Google Colaboratory

• os.environ[‘CUDA_VISIBLE_DEVICES’] (= gpu_id) = ‘0’
• random_seed = 42