BARTificial Intelligence

Our inspiration:

Our inspiration came from how people place a lot of faith in user reviews to help them decide where to eat, travel and shop, but spam, irrelevant rants, and hidden advertisements can often drown out authentic reviews. We wanted to create a comprehensive solution that addresses all these problems.

What our solution does:

BARTificial Intelligence is a 3-stage AI-powered pipeline that keeps reviews clean and trustworthy.

Text classification with a fine-tuned BART model to detect spam, ads, and fake rants.
User metadata analysis to flag suspicious or bot-like accounts.
Combine all signals into a fusion model with weighted scoring. Reviews are then put into 4 categories: genuine, suspicious, low quality, high confidence spam.

How we built it:

We fine-tuned the BART model on labeled review datasets, integrated user profile signals (review frequency, patterns, anomalies), and applied semantic embeddings for relevancy scoring. The stages are chained into a single workflow that processes reviews automatically.

Challenges we ran into:

We had challenges with our initial dataset and finding enough reviews that fit the criteria of "advertisement" or "fake rant". We also had trouble figuring out how to label the data that we had, since there could be disagreements for if a review was considered spam or an advertisement.

Accomplishments that we're proud of:

Built a multi-stage system that combines deep learning and metadata analysis.
Achieved strong results in identifying irrelevant and fake reviews.
Designed an approach that generalizes to multiple business categories, not just one domain.

What we learned:

The importance of combining semantic AI with simple heuristics for stronger results.
User metadata is just as powerful as text analysis in detecting suspicious behavior.
Even state-of-the-art models benefit from carefully chosen rules and thresholds.