Expose

• 

  Logo

• 

  Website homepage

• 

  Search Result Page

• 

  Evaluation page demo *displaying purpose only

Informational Inequality

Today, we are witnessing an unprecedented level of political bipolarization. Divided into ideological battlefields, Americans have drifted apart. Perhaps worst of all, even aware of the division, many Americans willfully turn a blind eye on their political opponent’s point of view and seek comfort in biased news feeds. The informational industry (e.g. media, big technology companies), instead of moderating the national tension through presenting each side of the story fairly, takes advantage of it and fuels it.

There is an immense informational inequality between the media and the public. Our project, then, attempts to mitigate that gap and reduce the current informational inequality, in order to restore fundamental American values such as civil discourse and progress.

What are we exposing?

Our mission: To ensure equitable access to reliable information for all, leveraging the power of technology​

The ranking of the articles that show up on a search engine, for example, is systematically determined by your previous preferences, advertisement endorsements, and so on, rather than only relevancy. In other words, search engine not only reinforces your biases but also allows the interests of particular wealthy individuals and organizations to influence your informational inquiries.

Especially since certain technological mechanisms remain unseen by the public, we are not even constantly aware of the fact that our own biases are being confirmed and strengthened while we search up information. In other words, we are subject to the news media’s control that is rooted in the informational inequality without knowing.

Our own formulations of ideas are highly influenced by the presentation (rhetoric, pathos) of information and information itself. Equipped with a brief analysis of the presentational and informational characteristics of an article, we can better defend our own thinking and stay unbiased.

What Expose does

Search: Search for online articles under the topics of the user's interest through default news searching engines (e.g. Google news). Notice that we are not changing the algorithm of the search engine. If we give a different ranking of the results, we are somehow still trying to make people conform/obey the standard we set - even if this is relatively more objective, this is never not the right way to address the inequality.

Evaluate: Use the pre-trained natural language processing model to evaluate the subjectiveness of the news articles, generate a percentage of subjectiveness for the article, 0% the lowest and 100% the highest. Also evaluates the subjectiveness the certain new press, generate the distribution of its articles based on subjectiveness. All of the ratings we provide are given purely by natural language processing. Quantified, accurate, objective, high-correlation.

Display: Provide the user with the evaluation result of the articles and the news press through percentage and graphs/bars, so they would be informed with how subjective the article is before actually reading it​.

How we built it

Front-end: We designed the front-end website and justified its usage through a video demo.

Back-end: We used a pre-trained natural language processing model to give subjectiveness evaluation to the news articles. We also worked on implementing a web crawler that obtains information from online news webpages to provide data to process. We used a breadth-first-search based web crawler to obtain news articles radiating from a webpage (for example, New York Times, shown in the NLP section below). We focused on analyzing and displaying the distribution of articles published by the same press to our customers.

(See presentation slide for detailed NLP model description, page 7-9.)

Justification

Why can we use a model trained on movie reviews on news?

Initial thoughts: Both movies and news models are subjectiveness evaluations based on sentences. The language unit is small, and the theme difference will not serve as a determining factor.

Model analysis: As we look into the histogram plots of both the movie review model and news article model, we can see that the distributions are similar.

(See presentation slides for details) If we leverage the effect of the huge accumulation around 0.0 in movie reviews histogram, both distributions share approximately the same structure, where scores concentrated more on the left side of the [0, 1] range.

Challenges we ran into

We faced the challenge of implementing the pre-trained natural language processing to run at back-end while the user is searching at the front-end. We worked on transferring the back-end to run through Google Cloud AppEngine, but are still working on realizing it. We successfully implement the web crawler on back-end locally, but for real-time generation, we still need more time. Also, we had trouble building an actual search engine at the front-end, so we only have a demo website for now.

What's next for Expose

Finalize and optimize the rating model by taking more relevant factors into account: ​the subjectiveness score of the title and in-article images​, ratings given by the readers at the end of the article​, the authority of the publications, etc

Interactive web application with rating model running at the back-end in real time​

Chromium-based extension (Google Chrome compatible)​