Did you open the paper and read it? The hypothesis are very simple.

They need to be set up with two parts, the first a predicate, then the second part is a couple options..

So for example a hypothesis can be set up in two parts as follows:

Part A:

"The author of this comment { } about a border wall"

Part B:

["thinks negatively" | "thinks positively" "is neutral"]

The options are intended to fill in the gap in the curly braces.

The model will give a probabilistic ranking of the three options, so you need to think carefully about how you set up your hypothesis.

Like I said drop them here or dm me and I can run them once I've scrapped UMs comments.

[Edit: I've got UM's comments, and I've saved them to disk. Let me know if you've got your questions ready, or if you still need help understanding how to set up a hypothesis]

[Addendum] @SatansMaggotyCumFart@lemmy.world

I'm going to give you a worked example.

This is on UM's most recent comment:

"She was on my ballot, so she is a candidate. I don’t know how to explain this any better."

So I set up the predicate:

'The author of this post {} Joe Biden.'

with the options:

['supports', 'opposes', 'is not talking about']

and we get the result:

{'sequence': 'She was on my ballot, so she is a candidate. I don’t know how to explain this any better.',

'labels': ['is not talking about', 'supports', 'opposes'],

'scores': [0.9906510710716248, 0.008063388988375664, 0.0012855551904067397]}

So this comment we would score as "not talking about Joe Biden". Anything you can think of that can fit within that framework. I dont know UM, but you seem to, so you probably know what would be interesting to ask.

I'm literally setting up for that right now, and for the third time you are accusing of not doing exactly what I'm trying to do for you. It takes a while to download all the comments. I'll let you know when I have them.

The fuck are you talking about. I literally said I was doing it in my response.

Bruh this is why you come up as a troll in so much of your comments.

I'm putting dinner together but I'd be happy to run UM if you would like me to. In exchange would you read the paper so you can understand how the sentiment analysis works? Its important for hypothesis testing. You need to set up good hypotheses for this to be effective.I'm going to down load their comment database now. You work on coming up with some hypotheses.

UniversalMonk

@satansmaggotycumfart@SatansMaggotyCumFart@lemmy.world

I'm downloading the data and will save it to disk, but am about to sit down to dinner. I'll save the output and either drop the hypotheses you want to test here or dm me, but I may not respond tonight.

Sorry I should have been more clear. That was for the "high confidence that the conversation is around Biden" cohort of comments. So within a subset of about 5% of your overall number of comments., so maybe 2.5 - 5% of comments in total you are making one of these kinds of accusations, or about 1:20 or 1:40. I ran a frequency analysis, and at several points you just spam the same comment over and over again, so that might be skewing things. I'm not sure that should be filtered out, because it is trolling.

And yes, I think more testing is required, but most importantly, I think I need to get more of a context window around comments. I want to do this using the whole comment chain or thread. That gets more complicated because now you have 'identities' (speaker A, speaker B, C.. etc), which is where the graphical approach is going to show its benefits. Again, work for another time. At least at a first pass, a few minutes of work adjacent to some other work I'm doing level of effort, its more than sufficient to make my point.

By what basis do you consider a comment abusive or trolling?

Its part of how the hypothesis is set up. You can read the paper I cited here: https://arxiv.org/abs/2409.02078

So specifically for that question the hypothesis “The author of this reply {} Biden.”, with h0 being “is being abusive, or trolling.” and h1 being “is being honest and genuine.”

And on your second point, since I've still got the data up, we can address that specifically. We'll address the following hypotheses. 'The author of this reply {}': 'is accusing someone of being a russian asset.', 'is accusing some one of trolling.', 'is accusing someone of being a bot.', 'is accusing someone of engaging in bad faith', 'is having a normal conversation'.

you make a bunch of charts that don’t really prove anything instead.

Only if you lack reading comprehension.

@SatansMaggotyCumFart@lemmy.world

Ok, results are in. Firstly, just some preliminary stuff on your posting behavior. I only pulled your last 6k comments, which goes back to February. You caught me at a good time, because I had been working on parts of this for a while for some network analyses I'm working on looking at the relationship between moderation bias and community sentiment, so I had some of these tools just laying around.

It looks like you really got posting in around April, and hit your stride over summer. You've slowed down a bit since. Also, you tend post most frequently at about 19:00 GMT or 3PM EST/ 12 PST, and then again around midnight GMT, or about 8PM EST/ 5PM PST.

For this work, I'll be using some models from this paper: https://huggingface.co/papers/2409.02078, "Political DEBATE: Efficient Zero-shot and Few-shot Classifiers for Political Text". This tool allows me to set up hypotheses like the following.

samples = list(test['premise']) template = 'The author of this reply {} Biden.'

multilabel entailment labels

labels = ['is talking about', 'is not talking about']

The multilabel option determines if more than one hypothesis can be true for the document. If false, the most likely label is returned. If true, a dictionary of labels and their estimated probability is returned.

res = pipe(samples, labels, hypothesis_template = template, multi_label = False)

Below is the result of the hypothesis 'The author of this reply {} Biden.', with the two options: {h0: 'is talking about', h1: 'is not talking about'], where we accept h0 at >0.5

It appears that your posts mention Biden at a relatively uniform rate. Please note that we're in percent of posts, not count (as with the previous two figures), since your post frequency has changed over time. It seems like for any given week, 5-15% of your posts typically mention Biden.

So for the below analysis, I tested the hypothesis "The author of this reply {} Biden.", with h0 being "is supportive of" and h1 being "is in opposition to". I only performed this analysis that had a very high probability of being about Biden.

So, "generally" supportive, but not crazy. You started less supportive of Biden than you are now, but like I said, I only grabbed the previous 6k comments of yours. Generally you seem to be about 50/50 on Biden. Which is against my previous assumption, I thought you were more supportive of Biden (closer to 80-90%).

The next experiment I ran was the test (on all of your comments, not just the ones mentioning Biden) was the hypothesis "The author of this reply {} Biden.", with h0 being "is being abusive, or trolling." and h1 being "is being honest and genuine.". I ran this test on all comments.

Honestly, @SatansMaggotyCumFart@SatansMaggotyCumFart@lemmy.world , I think you can up your game. You've got ample headroom to live up to your legacy.

HOWEVER.. If we look at the same results for posts which are explicitly about Biden... we can see that you are trolling and abusive at a rate much higher than your background rate.

So there is your answer. Not as bad as I thought, but not great. Definitely an abusive troll when it comes to political discussions.

Some limitations about this approach. I want to expand it to include the context that a given comments sits in. Its fine for a cursory analysis like this to just use single comments, but context is key. I think we'll get much clearer signal/ noise with more context. Also, these conversations happen in a threaded manner. I need to develop a way of accounting for that. I'll probably pull some methods that I've used for network analysis for that component. But I got the major issues out of the way, and I can run these kinds of analysis for anyone on the fediverse. So for a preliminary step, its at least on its way to being sufficient to identify bad faith/ troll accounts.

[I had to switch to a larger compute instance]

You can just tell people now what you already know. Like its all out here in the open. You can't really lie in the fediverse.

What will I find if I look?

Your trolling says more about you than it does me. I advocated to swap the candidate for literally months because Biden was an absolute TRAINWRECK of a candidate. There was about a 0% chance of Biden winning this election, and if I recall correctly, you defended keeping him as the candidate, in-spite of the incredibly obvious fact that he was losing the election in a dramatic fashion. You defended him to the point of calling any one asking to remove him a trolls, bots, and Russian assets.

So you weren't interested in improving the Democrats chances of winning. You were just behaving like an abusive troll. And I called that Biden wouldn't be the candidate litterally 8 months out. I was right, and literally almost everyone in this forum was wrong.

And for a few weeks, we had a good thing going with Kamala. But she's shifted to campaigning with neocon's and being "pro-buisness", and accordingly, her polling is dropping off a cliff.

She needs to shift back to how she campaigned/ policies she had before and up until the convention. IF she doesn't, she will lose this election.

But you aren't here for nuance. You just want to pretend that you are on the right side of anything.

And just as before, you aren't interested in figuring out what it would take Kamala to win. You are just here to abuse.

This feels like one of those games where the QB crushed it in the first half, and just completely fell apart during the second half. Now its up to the star running back to make it happen while the QB is still blowing it.

Click on all busses:

There is a short story in here about someone who can't pass a captcha, loses their identity, and has to move on to becoming a fisherman in Norway.

Dude I can't even pass a catchpa these day. I still don't know if a e-bike is a scooter, a bike, or a moped.

Your dismissive tone suggests you don't think the thing exists or happened. Which would mean you are either oblivious or intentionally ignorant to basic news cycles. But even the most cursory or superficial level of knowledge and understanding of the candidate and their campaign can be used to highlight her shift to the right.

It started during the convention, with the final night highlighting Republicans while refusing to platform a Gazan: https://www.cnn.com/politics/live-news/dnc-democratic-national-convention-08-22-24/index.html

Intercept on her embracing of NeoCon's: https://theintercept.com/2024/09/14/dick-cheney-kamala-harris-neocons/

LA times article on it: https://www.latimes.com/politics/story/2024-10-10/2024-election-kamala-harris-policy-shifts

However, we don't need only rely on reported news, we also have data we can use to address questions like this. (Codeshare link for reproduction: https://codeshare.io/dejgKZ, data sources are Nate Silvers)

Regression analysis puts the shift manifesting in polling at about ~September 16th.

And its not unique to one swing state (you can run the analysis yourself):

So I don't make the assumption that your comment is in good faith, but I also don't want you to have any wiggle room here.

spoiler

My exception is that your rhetoric will go from "She didn't shift to the right" to "It was a good thing that she shifted to the right"

You have to ask who these things are convincing to. Which cohort has their minds moved because of this interview (or that gaffe, or this policy, revelation whatever).

Harris hasn't moved any where but down in polling since her rightwing shift and she's just about out of time to change that. So it should be obvious that this approach isn't working on any one.

Seriously. We had months, years of this shit. He's a convicted felon over it. meanwhile, one side is trying on Hugo Boss for the frog march and the other side is funding a ally who prefers burning children in hospitals.

We've bigger shit to be discussing.

yeah thats what I should have put.

This is the same thing Hillary did in 2016. "Trump is bad. Therefore, you have to vote for me." I mean, the identical rhetoric has already shown its ineffective head in this thread.

Its always the voters that are wrong. The party or candidate can never fail; it can only be failed.