ActEarn

Inspiration

Team members have long been dissatisfied with news:

when users open 5 articles looking to better understand something, 80% of the content are exactly the same
also it's hard to know whether what the user read is the full story - is it biased? too subjective?
sometimes users felt strongly about certain topics, and wanted to voice their own opinion, to join the debate, to shape the narrative
but users, even citizen journalist don't always get the distribution and compensation they deserve

Therefore, we want to build a platform where we aggregate every news articles, every opinions on a topic into one "co-created" news "article". This way there is no duplication, there will be more voices from different perspectives, and users get to "shape the narratives" together.

But user comments can be noisy, and sometimes not necessarily constructive. So at this Hackathon, we also want to explore the potential of using LLMs to moderate "debates".

What it does

ActEarn is a platform for balanced news content that's co-created by citizen journalist and everyone, structured by NLP. ActEarn first uses NLP to structure news content based on opinions; and then lets citizen journalists/users “join the debate” by adding opinions and curate content to back/refute opinions.

At this Hackathon, we are focused on 1. using LLM to structure content based on their opinions, 2. using LLM to rate the quality and validity of arguments, 3. generate a final narrative based on "high-quality" arguments

How we built it

Built a crawler to automatically crawl news from reputable sources
Use a ML pipeline to extract key opinions of these new articles, and structure these articles based on their positions
Let users debate the topic
Rank and rate users' response by Claude and GPT
Generate a "narrative" based on the inputs of different articles and user debates that is the most comprehensive, in-depth, and interesting

Challenges we ran into

Setting up an automatic pipeline to extract key opinions, and structure them based on opinion is challenging because we initially decided to build custom models along the way. We used a out of box NER model and the Robert-small-MNLI model, but ran into some backend challenges. So we end up not shipping the pipeline with custom models. We instead used GPT API and it works well!

Accomplishments that we're proud of

We proved that LLMs can be very effective at moderating user debates! We shown that LLMs can be used to structure content based on semantic meanings! We iterated on product targeting to focus on news junkies and people who actually care enough about news topics to want to voice their thoughts.

What we learned

Although most people wouldn't necessarily want to see the other perspectives, we have found users who are passionate about "joining the debate". They are people who comment a lot on reddit, but they are also people who have been silent thus far - they were not motivated to "join the debate" because the current designs of the "commenting section" is not optimized to incentive them. They felt their messages don't matter under the current commenting sections.

By giving them an opportunity to "shape the narrative", users are more motivated to voice their opinions.

By using LLM to moderate the debate, we can incentive high quality discussions.