Word-level embeddings

[Mystic-Slice]

The system currently considers each sentence to represent a single value. This is a good tradeoff between granularity/accuracy and the speed of recommendation.

World-level embeddings could provide more granular and accurate classification. This is a discussion to explore that idea.

[santanavagner]

Hi @Mystic-Slice,

Thank you for starting this thread.

We need to gather evidence that by using word-level embedding we would also be able to identify more accurately the input prompts.

How do you see this applied to the datasets and API endpoints we have today?

[Mystic-Slice]

Hi @santanavagner
I opened this thread based on a suggestion from a community member, to know more about his idea.

@ArionDas You are welcome to share your thoughts here. Thanks!

[ArionDas]

Thank you @Mystic-Slice.

Hello, @santanavagner, so here's some context on how I came to know about this project.
I attended a call where @Mystic-Slice presented the work through Cohere Labs, where he showed the embedding value space. Each value in that space represents an entire sentence.

Now, when a user puts in some prompt, I vividly remember (from the presentation), the application currently recommends whole sentences in place of faulty (negative sentiment, foul words, etc) sentences (given by the user). This is because we are using sentence-level embeddings currently.

Two problems could arise here (correct me if I'm wrong):

• we miss out the context of some positive sentiment / useful information in some user's sentence to remove (let's say) a few faulty words somewhere in that sentence.
• the sentence we recommend maybe close to what a user would've meant, but maybe it could miss out on certain essential "jargons" specific to their domain of interest (we don't have such words in our static dataset of sentences).

I had proposed, why not use word-level embeddings instead of entire sentences, so the application can be more dynamic, and recommend swapping faulty useful words instead of removing entire sentences.

Happy to discuss more on the same, and now that I have a healthy knowledge of the repository, I'll be happy to help to implement it out as well, if it seems feasible.

Thank you.

[santanavagner]

Hi @ArionDas ,

Got it. I agree with you and results from our latest paper are aligned with this, i.e., users don't like when the system removes the whole sentence in case one or more words are identified as harmful.

Actually, this idea on swapping terms is being implemented by @luanssouza in a feature for recommending a reprhasing of the sentenced identified as harmful (#17) instead of removing it.

@luanssouza ,

Please consider sharing the working code you have so @ArionDas can collaborate with you on issue #17.

Thank you all!

Cheers

[luanssouza]

Hello, @santanavagner and @ArionDas!

Interesting idea @ArionDas. The way I am doing the negative sentence rephrasing goes in a different direction: I am using prompts to ask LLMs to rephrase negative sentences considering our social values.

The work I have been doing so far can be found in the brach rephrase of our repository.

Best regards,
Luan

[ArionDas]

Hey @luanssouza,
I went through your code where you used an LLM to generate more dynamic, "word-level" recommendations.
Very nicely designed task with all the tags, etc.

Is there a place I could test it out maybe? (on the UI I mean - because I can't host the LLM on my personal machine.)

Maybe, then we can connect and discuss about doing something about recommending "domain-specific jargons"?
Or you can let me know if you have something else on your mind, and if you want me to help you out.

cc: @santanavagner @Mystic-Slice

Thank you!