The Problem: Media Bias
Link to Project: https://bit.ly/2KbunTR
I would say I spend a fair amount of time per week reading up on global affairs. It’s an interest of mine, I think it’s important to stay up to date, and it’s a topic of conversation I have a lot with my folks. I read mainly from US centric news outlets, but I do try to diversify and read outlets based outside of the United States. Personally, I’m a big fan of The Economist and The New York Times.
I’m well aware that different news outlets have their own agendas and point of views, some news outlets hold a high level of integrity and provide great journalism, however some news outlets may be less so. I can’t say. However, I know that there are news outlets that produce content that I completely disagree with.
An obvious example of different media biases are in political news, there are liberal voices, conservative voices, libertarian, corporate, etc. Thus this project was born, and I wanted to figure out a way to gain some insights to the possibility of different media biases to different topics.
What did I do?
First of all, let’s talk a little bit about the data I used. I didn’t use any datasets that were on kaggle or elsewhere. I created my little data set from scratch, from web scraping articles online. To which I wrote a tutorial on. Please check it out, it’s a shameless plug.
After accumulating data comprising of various articles from various news outlets, I went to work. However, the data I got online was not labeled with the topics it was talking about. Sure it was possible to get the category of the article, but I wanted the specific detailed topic for each article I scraped. For example, if article A discussed about the COVID-19 response in Canada, that would be a category in itself and hopefully the various news outlets had articles that covered the same topic.
I had the bold idea to perform an unsupervised machine learning method to find clusters of different topics in my corpus, to which I would then perform a sentiment analysis and subjectivity analysis to see if that would provide me with any interesting insights.
Essentially I needed to do Latent Dirichlet allocation Topic Modeling, which is an unsupervised clustering algorithm specifically used in natural language processing.
Topic Modeling: Its job is to identify topics from analyzing a collection of documents, to spot word patterns within them, and automatically cluster these patterns that best describes a set of documents.
After topic modeling, I calculated a sentiment score and subjectivity score for each document. The sentiment score ranges from -1 to 1. With 1 being a positive sentiment, meaning the article is mainly positive and -1 being a negative sentiment. The subjectivity score which ranges from 0 to 1, with 0 being very objective in the writing style and 1 being very subjective in the writing style.
From here I suspected that with the different topics identified and grouped together from our LDA topic model. I made the following assumptions:
- High subjectivity score meant the news outlet may have a certain agenda for that topic or could be very sensationalist in the writing style.
- Opposite Sentiment for each topics meant they have different agendas.
- Low Subjectivity & Low Sentiment meant that this news outlet generally stayed neutral and focused on the facts.
The Outcome
I’ll highlight some of the interesting insights from this project below.
| Insight | Note |
|---|---|
| Topic - Sports | All news outlets had positive sentiment score. |
| Topic - COVID-19 - Trump | Most news outlets had low Sentiment, with Washington Post having negative sentiment & highest subjectivity |
| Slate.com | Had the highest subjectivity score across all topics, which makes sense since the news outlet is often criticize for adopting contrarian views. |
| Channel News Asia | A Singaporean news outlet, had very low subjectivity scores and consistently low sentiment scores. |
Project Shortcomings & Potential Updates:
The project does have its fair share of shortcomings. Most importantly, my dataset that I scraped was quite small. I had roughly 4000 articles. After cleaning, it went down to around 3000. Which is not enough if I wanted more unique insights. However, I do plan on scraping a bunch more articles and run it through my code in the future and see If I could get a better analysis and do a better job at answering my question.
Another improvement I could make is by studying my texts a little closer, and do a better job at cleaning each document during the nlp preprocessing phase.
Preview:
However, this is a general overview of my project, if you would like to get to know the nitty gritty details of my project. Please do check out my project’s GitHub repository here. It’s much more well documented to what I did, and I’ve taken the courtesy to write up some deeper analysis in the markdown cells of my Jupyter Notebooks.
If you like to check out my online dashboard of the various topics sentiment & subjectivity score. You can find my Google Data Studio Dashboard here.
Thank you for reading! 🙂
Extra Resources & References:
Great Video on LDA Topic Modeling by Andrius Knispelis.
Feature graphic on LDA topic modeling belongs to Christine Doig. Great presentation by Christine on the Introduction to Topic Modeling in Python to be found here.