Incongruity Detection between Bangla News Headline and Body Content through Graph Neural Network

TL;DR

This paper introduces a graph neural network model specifically designed for detecting incongruity between Bangla news headlines and content, addressing the language's unique syntactic challenges and resource limitations.

Contribution

It presents the first graph-based hierarchical dual encoder model tailored for Bangla, achieving high accuracy in incongruity detection tasks.

Findings

Achieved above 90% accuracy on Bangla news datasets.

Addresses low-resource language challenges with a novel graph neural network approach.

Demonstrates effectiveness of graph-based models for complex syntactic languages.

Abstract

Incongruity between news headlines and the body content is a common method of deception used to attract readers. Profitable headlines pique readers' interest and encourage them to visit a specific website. This is usually done by adding an element of dishonesty, using enticements that do not precisely reflect the content being delivered. As a result, automatic detection of incongruent news between headline and body content using language analysis has gained the research community's attention. However, various solutions are primarily being developed for English to address this problem, leaving low-resource languages out of the picture. Bangla is ranked 7th among the top 100 most widely spoken languages, which motivates us to pay special attention to the Bangla language. Furthermore, Bangla has a more complex syntactic structure and fewer natural language processing resources, so it becomes challenging to perform NLP tasks like incongruity detection and stance detection. To tackle this problem, for the Bangla language, we offer a graph-based hierarchical dual encoder (BGHDE) model that learns the content similarity and contradiction between Bangla news headlines and content paragraphs effectively. The experimental results show that the proposed Bangla graph-based neural network model achieves above 90% accuracy on various Bangla news datasets.