10-22, 10:00–10:25 (Europe/Stockholm), Data
Live stream: https://www.youtube.com/watch?v=Ew4tKVem6F8
In this talk, we aim to find if polarization is induced in a neural
network by feeding it newspaper articles with manufactured sentiments according to the
Allsides Media Bias chart for the level of faith people on various aisles of the political
spectrum. This project consists of a set of experiments on similar data-sets from news
agencies across the various subsets in the ”media-bias” chart. News Media perceived bias
is common across consumers that belong to various political affiliations. While anecdotal
evidence of this exists and there exist annotated datasets that aim to annotate the ”spin”
a news agency puts on certain events and entities, whether this is a widespread problem
and whether it can be detected by the neural network topically or temporally is a problem that needs to be explored. The news media bias analysis is modelled as a Natural
Language Processing sentiment analysis task and a fake news binary classification task to
deduce the level of polarization in a neural network by feeding it headlines embedded using
pre-trained sentiment models from news publications across the political spectrum. When
it came to fake news vulnerability, news from all kinds of perceived politically affiliated
news media holds up well against a fake news dataset with a very good accuracy. None of
the accuracies dropped below 95%. This is a significant result that sort of debunks the AllSlides
Media categorization - if taken as simplistically as it is presented. These experiments can be extended to include entity based topical
studies in the future and to also educate the populace about their perceived biases.
As social media sites across the world grapple with hate speech, calls for genocides and sexual
harassment on their platforms, as policy scientists look up the various biases in our justice system’s usage of language and as most of the people in the world struggle with what is globally
called ”media bias”, I believe that as mathematics and statistics became commonplace measures, so will Machine Learning models. This work is an example of an intersection of a non
scientific field with computer science and mathematics, trying to quantify, measure and identify
non mathematical phenomena in the language of mathematics. It is important because it could
be the basis of the scientific approaches that the next generation policy makers, voters, non
profit social organizations and governments could use to make life changing decisions for their
citizens.
2
The questions that this study tries to answer is whether a neural network can learn biases from
the news media based on perceived bias scores obtained from independent agencies. It also
seeks to answer whether any of these political leanings of the news media affect the vulnerability of their consumer when it comes to fake news. The results of this experiment aim to show
Conclusions
1. SVMs perform better clustering with respect to the categories than neural networks, however the maximum does not cross 67%
2. The most significant conclusion from this work is that though there is a perceived bias
when it comes to news agencies, when looked at from a neural networks standpoint, it
is negligible. Mainstream news agencies are not able to polarize a neural network with
inherent biases in their headlines.
3. There may be topical biases that need to be examined by using an Entity linking and bias
calculation approach
4. Most mainstream news agencies do not make the consumer vulnerable to believing fake
news. This study needs to be conducted with data from popular social media ”news”
groups or popular TV shows that masquerade as news but may technically not even be
news channels.
5. It is safe to conclude that the perceived bias that stems from social media polarization is
being extended to news media when their contribution to the polarization may be negligible.
Significance of Work
1. The significance of this work is to be able to transform a social problem into a technical one and using neural networks and Machine learning techniques to try to gain some
insights.
2. Hopefully using these techniques to find deeper trends will become mainstream and help
policy makers and the general citizens approach news media bias in a better light.
Future Work
Some further studies to take up are as follows:
1. Effect of news media on Perceptron Networks
2. Better Annotated Datasets to perform bias analysis
3. Effect on memory models of media bias
4. Experiment on some of the most polarizing news epochs in time
5. Studies on country level news bias
Aroma Rodrigues is a master's student at UMass Amherst. She believes that Automation is the path to Inclusion. In 2016, a teammate of her "Shoes for the Visually Impaired" project presented it at the FOSSASIA. She reads, writes and enjoys walking to explore places. She presently works in a financial services firm and believes that solving problems that she has would solve problems for a large chunk of the world. An ML enthusiast she has about 20+ Coursera Certifications with the respective project work to support her learning in that field. She presented a talk on “De-mystifying Terms and Conditions using NLP” at PyCon 2018 and a talk called “Propaganda Detection in Fake News using Natural Language Processing” at PyCon ZA 2019 in Johannesburg. She spoke on detecting gender roles based biases in school textbooks at PyOhio 2020.