SBERT-Based Semantic Similarity Engine

[student-smritipandey]

Sentence-BERT (SBERT) – A Siamese & Triplet Network modification of the BERT architecture designed for sentence-level embedding and semantic similarity.

This model is designed to detect semantic similarity between an incoming SMS and a set of known spam messages, even when the message is paraphrased or reworded.

It enables Spamlyser Pro to:

Identify variant spam attacks (e.g., same meaning, different wording).

Improve recall by catching previously unseen but similar messages.

Reduce false negatives from traditional classifiers.

Describe alternatives you've considered

No response

Additional context

No response

[github-actions]

Thank you for raising this issue!

We'll review it as soon as possible. We truly appreciate your contributions! ✨

Meanwhile make sure you've visited the README.md, CONTRIBUTING.md, and CODE_OF_CONDUCT.md before creating a PR for this. Also, please do NOT create a PR until this issue has been assigned to you. 😊

[student-smritipandey]

@eccentriccoder01 please assign me here in my issue,I am gssoc'25 contributor

[eccentriccoder01]

This is an exciting idea @student-smritipandey!

Quick curiosity though, how do you see yourself implementing this? Right now all fine-tuned models are trained in Kaggle and fetched from my HuggingFace repos. Since SBERT here means introducing another model stream, do you plan to train & host it separately, or integrate it alongside the existing pipeline?

Would love to hear your take I’m especially interested in how you’ll connect this with the current Hugging Face setup without breaking the flow...

[student-smritipandey]

I want to integrate in existing pipeline but don't know whole process so need guidance for that

[eccentriccoder01]

Got it @student-smritipandey, sorry for the delayed response as I was caught up with exams...No worries at all, I’ll be happy to guide you through integrating SBERT into our current HuggingFace pipeline so it blends smoothly with the existing models. That’s the fun of building together❕

Since Spamlyser’s models are trained in Kaggle and then fetched from my HuggingFace repos, the same flow applies here too. You’ll need to:

1. Start with a pre-trained SBERT model from HuggingFace (e.g. sentence-transformers/paraphrase-MiniLM-L6-v2).
2. Fine-tune it on the sms_spam dataset (you can follow a process similar to the existing backbones).
3. Push your tuned SBERT model to your HuggingFace repo.
4. Update the Spamlyser code (in your fork) to fetch this new model alongside the existing ones.

Please tell me wherever you feel hurdles or need guidance, I'll be happy to assist you :)

[student-smritipandey]

@eccentriccoder01 okay

[eccentriccoder01]

@student-smritipandey Hope you're alive, LOL❕