Copy the page URI to the clipboard
Bharati, Nitu
(2023).
DOI: https://doi.org/10.21954/ou.ro.0001590a
Abstract
The overarching aim behind this research is to automatically detect the stance of the body of a news article relative to the article’s headline. The news headline may not always reflect what is in the news body. The stance of a news body to its headline can be agree, disagree, discuss or unrelated (Pomerleau & Rao 2017). Central to this work is the use of a specific discourse relation, the attribution relation (AR), for detecting the stance of a news article body relative to its headline. An attribution relation is a span of text which links a source to content through a cue. For example, consider The boy said it was a spider. Here, the boy is the source, said is the cue and it was a spider is the content. This thesis also examines how the expertise of sources affects stance detection. The main research question of this work is “Can attribution relations and source expertise be useful in detecting the stance of a news article’s body towards its headline?”.
To address this research question, I developed a new attribution detection model that can tag components of attribution relations in news texts. I developed a new stance detection model which uses these tags as input, rather than working on the whole article as a single piece of text, with performance comparable to state-of-the-art. Furthermore, once we add the source expertise information to our stance detection model, this has a positive effect on the F-score for stance detection (increase by 14%).
The work is novel in a number of further specific ways. Firstly, it is the first time a single-step deep learning approach has been applied to AR detection and been released as open source code. Second, this is the first time that attribution relations from a news article body have been used as input for a stance detection model instead of the full text of the news article body. As part of this research, I created an extension to the Fake news challenge corpus (Pomerleau & Rao 2017) with addition of source expertise data. Finally, I separately confirmed, through an empirical study, that source expertise is positively correlated with the credibility that readers assign to claims from a source.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)
