Automatic Incongruent News Detection (From the Perspective of Body and Headline Centric Representation)
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The prevalence of deceptive and incongruent news headlines has demonstrated their significant role in propagating fake news, which worsens the dissemination of both misinformation and disinformation. In the literature, incongruent news article detection has been studies from two aspects- body-centric and headlinecentric encoding. However, earlier headline-centric and body-centric approaches in the literature fail in the following scenarios.(i) The hierarchical encoding in the earlier studies is limited to paragraph level only, and headline-guided attention highlights paragraphs that are contextually similar to headlines. However, considering
the underlying incongruent news detection task, highlighting the paragraphs or sentences that are not contextually similar to the headline is essential. (ii) It fails to detect partially incongruent news articles. To address the first limitation of studies in literature, this thesis proposes a Gated Recursive And Sequential Deep Hierarchical Encoding GraSHE method for detecting incongruent news articles by extending the hierarchy structure of news body from body to word level and incorporating incongruent weights. The proposed
model, (GraSHE) captures the long-term dependencies and syntactic structure by incorporating sequential information at the paragraph and body level (using BiLSTM), and syntactic structure at the sentence level using child-sum Tree LSTM. Further, unlike headline guided attention models, (GraSHE) also incorporates
incongruity weight to capture non-dominant textual segments which are not congruent with other part of news body. To address the second limitation of studies in literature, this thesis proposed dual summarization and graph context matching based methods. This thesis proposes dual summarization-based methods Multi-head Attention Dual Summarization MADS and dual-summarization based approach, namely DuSum, which divides the news article body into two sets, positive and negative set. Sentences congruent to the headline are
Description
Supervisor: Singh, Sanasam Ranbir