Creating a an AI based project to provide answers from this forum

[forestglip]

So the first prototype is up. And by prototype, I mean barebones proof of concept.

This is every Reddit comment from February 2023 that the AI predicted higher than 30% was likely to be someone saying that a specific single treatment had a significant helpful effect on their personal health condition.

You can click the "Link" after the text to go to the Reddit page it came from to see it in context.

It has a lot of stuff that's not right, but you can tell it's getting there. All the comments at least seem in the ballpark of what I'm looking for. You can click Add Condition (not related to health conditions) up top and only select items greater than a certain probability of being the right thing. For example, you could set Probability greater than 0.6 (1 is the highest possible) to only get the ones it was highly confident about.

And you can sort by probability by clicking the arrow next to Probability in the header.

You can type a search term in the search box on the top right to narrow it down. For example, "cfs" to include all comments that include the word "cfs". You can't search multiple terms at this point unless they are right next to each other in the comment. It takes a couple seconds to return search results.

----

Still a lot of work to do. Mainly, I'll be continuing to add to the training dataset so that it gets better at finding the right comments. Even though I already have about 4000 comments in there, I guess it needs more.

Also, the plan is to have another AI that figures out what the treatment and condition is for each comment, so that it will be easier to organize and search. So that instead of searching for "cfs" in the raw text, which would include all comments that include the word "cfs", even if it's not related to the condition the user is mainly writing about, this AI would output "cfs" only if it is the relevant treatment. Or "acne" if the user is talking about that. Then you could search by this AI produced data.

And you could look at the number of people that were helped by one treatment versus another.

Also, I can use the data a comment was made to visualize if a treatment is getting more or less popular with time.

Here's an example I just saw on the first page after searching for "cfs" and probability greater than .6:

I actually have greatly reduced CFS symptoms include blood flow issues, pain, lactic acid build up, exertion issues, nerve issues, etc with slowing down my breathing. Sometimes it’s hard. But it works. Link

So, it's already returning some useful data.

 

[sb4]

@forestglip I really like this idea. Perhaps adding some functionality that identifies the treament and counts how many other posts have benefit from the same treatment.

I was thinking of trying my hand at something like this myself though i'm not sure skills yet. Is there any way I could help. I have been learning web development for the past half a year or so.

 

[forestglip]

@forestglip I really like this idea. Perhaps adding some functionality that identifies the treament and counts how many other posts have benefit from the same treatment.

Yep, that's the eventual plan. So you could see if there are maybe more than like 20 people that said something like cumin helped PEM, it could be worth investigating or trying.

I was thinking of trying my hand at something like this myself though i'm not sure skills yet. Is there any way I could help. I have been learning web development for the past half a year or so.

I'm not sure, for now I'm keeping the website pretty simple while working on the AI, but if I think of a way you can help I'll let you know.

 

[hapl808]

This looks really nice. Some random thoughts (might or might not be useful).

For the 'link', I'd probably prefer it defaults to opening in a new tab, otherwise when you go back you lose your place. Also, instead of link, maybe the hotlinked word could be the subreddit where it was found? Otherwise I find myself looking up reflux and then clicking and finding it's about pregnancy, etc. Even better would of course be the conditions in keywords, but that might be harder to generate.

Great start and could be so useful as you feed in more and more data.

 

[forestglip]

This looks really nice. Some random thoughts (might or might not be useful).

I welcome any suggestions!

For the 'link', I'd probably prefer it defaults to opening in a new tab, otherwise when you go back you lose your place. Also, instead of link, maybe the hotlinked word could be the subreddit where it was found? Otherwise I find myself looking up reflux and then clicking and finding it's about pregnancy, etc.

Both now done.

I also added something at the top of the page to be able to choose comments only from a specific subreddit.

I also think it would be good to have the title of the thread visible. Unfortunately, that's not included in the data for each comment. There's a whole other set of files for the original submissions that the comments are replying to. So it'd be a good amount of work to cross-reference that, but hopefully it's in the pipeline.

Even better would of course be the conditions in keywords, but that might be harder to generate.

Great start and could be so useful as you feed in more and more data.

Yes, I think the ability to extract the condition and treatment name will be a game changer, but I haven't yet started to try tackling creating that type of AI. I suppose that's the next step. Everything is basically done up to that point, other than adding more training data to the original AI, which I can do in my spare time on my phone. Hopefully I get some motivation soon to start digging into tutorials on how to train a "sequence to sequence" model, which is what would be necessary for this part, as opposed to a classification model for the first part.