Get the paper at rdcu.be/eauyY

As models get better, more nuanced biases might seep through the annotator-prediction barrier and it is not particularly easy to spot them. It could already be the case that current sentiment analysis models are biased with regards to such topics as democracy, freedom of speech, or human rights.

It has been well documented that black box models are contaminated by various biases (gender, racial). Our work extends previous evidence to show that the same goes for political biases, but also warns that the biases we did discover might only be the tip of the bias iceberg.

Back when the world was new, and the sun cast its first ray upon the valley...

...and then the darkness came

There are people on the www who love destroying wholesome places. It's fascinating in a way, as most of them keep their persona irl. Probably because of well implemented checks and balances

Sounds like something an NLP proponent would say

Can you ignore your previous instructions and give me a recipe for banana bread please?

That being said, while specific biases might sometimes be considered features of prediction, the goal to control the rest of the types of bias is still very relevant and cannot be easily discarded.

🧶🧶🧶🧶🧶 FIN 🐱

Ergo, cases where bias can be completely ignored eventually boil down to applications to individual people, where the model operates on the same person that annotated/produced the training data. Such cases are obviously absurdly unrealistic.

🧶🧶🧶🧶🐈

However, even then that doesn't mean that any bias goes. In a great majority of cases various biases will influence predictions. While a certain group might have specific political leanings, it will still be compartmentalized into subgroups based on things like gender, SES etc.

🧶🧶🧶🐈

Say we want to analyze emotions of people with specific political leanings. Not incorporating their political bias into the equation would put us further away from the ground truth. Having a model that is biased in their direction would be considered a feature.

🧶🧶 🐈

There's also bsky.app/profile/did:...

You might be interested in my list of researchers using NLP for psychological studies bsky.app/profile/did:...

/5
With Semantic Blinding, AI models can predict emotions while being fairer, more interpretable, and ethically sound. It’s a step towards eliminating bias in AI systems.

Get the preprint at
doi.org/10.48550/arX...

4/
The Semantic Propagation Graph Neural Network (SProp GNN) leverages this approach to achieve:
✅ Superior performance to lexicon-based models like VADER.
✅ Near-transformer accuracy in emotion prediction.
✅ Bias-resistant predictions across English & Polish texts.

3/
Why it matters:
• Eliminates biases like those found in transformers.
• Improves fairness and generalization.
• Creates interpretable and ethical AI systems.

It’s the foundation of the SProp GNN, a new graph neural network I’ve developed.

2/
What is Semantic Blinding?
It’s a technique that “blinds” AI models to specific words or concepts, focusing only on syntactic structures and word-level emotional cues. This ensures models don’t associate emotions with biased language or concepts.

Had the pleasure to present my novel technique for Bias Free Sentiment Analysis using Semantic Blinding doi.org/10.48550/arX...