Domain Specific Tweet Generation

This project is part of the course "Information Retrieval and Extraction" from IIIT-H.

In this modern world, a vast amount of news articles are being published online across domains every day. These articles carry vital information embedded in large texts, ready to be consumed by the readers. It might be easier for the end-user to comprehend short rich text snippet to absorb the key information. This can be done in multiple ways like by posting on micro-blogging sites, social-media or any other public platforms. Twitter can be used for sharing information to wider audiences, but generating the specific tweets for each article will be a tedious and time consuming task.Automated generation of tweets for these articles can contribute for faster write-up and domain related tweets which can be deployed as an independent system.

Requirements

Install required packages using pip:

pip install -r requirement.txt

Performance Analysis

Three different experimental approaches are carried out and following are the preformance analysis for each of them

Approach 1 (Pointer Generator network)

Domains	BLEU (1-gram)	ROUGE (1-gram)
Entertainment	0.13	0.19
Sports	0.08	0.18
Politics	0.16	0.33
Space-Science	0.23	0.38
Business	0.20	0.31

Approach 2 (SVO based approach)

Domains	BLEU (1-gram)	ROUGE (1-gram)
Entertainment	0.09	0.08
Sports	0.17	0.15
Politics	0.17	0.18
Space-Science	0.06	0.07
Business	0.13	0.13
Combined	0.12	0.12

Approach 3 (Topic sentence)

Domains	Input	Train Model	BLEU (1-gram)	ROUGE (1-gram)
Entertainment	first sentence	first sentence	0.49	0.50
Sports	first sentence	first sentence	0.42	0.49
Politics	first sentence	first sentence	0.53	0.60
Space-Science	first sentence	first sentence	0.56	0.62
Business	1 sent. summary	1 sent. summary	0.38	0.43
Combined	first sentence	first sentence	0.20	0.24

Website

project site