WorkThoseAsserts

The web app in action!
Prompt Engineering in Cohere Playground
Data collection and manual classification

Inspiration

Women sometimes struggle to communicate confidently in the workplace. Fueled by the versatility of Cohere's LLMs, I set out to explore how sentiment analysis can transform negative thoughts and expressions into a more uplifting embedding space. Inspired by the practice of cognitive behavioral therapy, I sought a way systematically challenge disempowering thoughts.

What it does

There are 2 components to WorkThoseAsserts: the Assertiveness Booster, and the Cognitive Fallacy Crusher.

The Assertiveness Booster empowers you to tap into your inner she-hulk with tricky workplace situations. Given a workplace situation, it transforms a tentative draft of what you're trying to say into a powerful statement that will be sure to leave an impression on your boss. Whether you choose to use them is your prerogative; regardless, reading the uplifting and self-confident statements will leave you feeling empowered.

The Cognitive Fallacy Crusher takes a thought that's been troubling you, learns the cognitive fallacies lurking beneath the surface, and generates an uplifting response to combat it.

How we built it

Tuned with prompt generation on the [Cohere Playground (https://os.cohere.ai/playground/large/generate)
Crafted prompts structured according to 10-15 recognized subcategories of cognitive fallacies.
Powered by React, Node, and Express

Challenges we ran into

Prompt generation was unreliability
Assertiveness Booster sometimes got too assertive... imagine telling your manager "I think my performance would qualify me for a raise, I just need you to review it with me."
Labeled datasets on cognitive distortions are nowhere to be found :'(

Accomplishments that we're proud of

Creating my first React app myself!
Coming up with great transformations! Check them out: gSheet
WorkThoseAsserts encouraging my friend after her breakup

What we learned

Classification of cognitive distortions into subcategories seems unreliable in mental health datasets
In classification, there’s a marked improvement in crowdsourced data vs the assumed high-quality data from counseling, which could be attributed to larger dataset and leading questions all the data seems self-annotated by researchers or employed psychologists, no public datasets available

What's next for WorkUpAssert

Experiments into whether LLMs acquire sufficient abstract reasoning to directly address different logical fallacies, by constructing more uniformly propositional input data
Employ Cohere's Classification endpoint to explicitly extract one of the 15 types of cognitive fallacies in the Cognitive Fallacy Crusher, to help people learn to identify their own cognitive fallacies. Break those negative cycles of thought!