Abstract:
Ideology is defined as the set of attitudes or beliefs shared by members of a social
group. While news agencies cover events of death and violence, the way these events
get framed has a direct impact on how societies respond to such dramatic events.
It could promote military intervention, call for escalation of violence, or advocate
for peace, making them candidates for special attention. Since the way stories
get framed signals hidden ideologies, text discourse, and specifically opinionated
discourse, provides the ultimate gateway for the identification of such ideologies.
This thesis presents a systematic framework for detecting hidden ideologies. Using
two large Arabic newspapers with different political views, An-Nahar and As-Safir,
and by focusing on the Arab-Israeli conflict, we show that political stereotypes are
manifested as biases in word embeddings, and that shifts in stereotypes along the
temporal axis act as a response to events happening locally and in the region. We
show that the application of contrastive viewpoint summarization of certain war-
related entities signal hidden ideologies. Finally, we transform the problem into a
supervised learning problem tapping at the sentence level, were Van Dijk’s discourse
context model is used as the annotation scheme, operating in a challenging set-up
where not a lot of annotated data is available, and for that, we leverage the usage of
meta-learning approaches that thrive in under-resourced settings. Finally, we answer
the hypothesis of whether opinionated discourse is the host of hidden ideologies and
reflect on its importance