it isnt about truth or fiction, it is negative/hate tweets. I run a politics forum and the one rule is no insulting other people, groups, or positions.
This makes for a very civil, but bland conversation. People engage with negative/hate tweets.
I personally like talking politics without insults, but most people are incapable of it.
I would fully expect the next thing for you to say is "I can program my own Twitter in a week"
Trying to figure out language intent is just the kind of thing an engineer/moderator says is easy and then is in deep water a month later after a phrase that means "you're great" in one language means "you're a donkey's anus" in another.
When you're moderating a small group it can be somewhat easy, everyone tends to speak the same language, and quite often it just falls into a groupthink that excludes situations like this. But when the situation scales you don't just have users that actively want to use the service, you have adversarial users that want to abuse your service and make it hell... and those users can be exceptionally clever.
>>I would fully expect the next thing for you to say is "I can program my own Twitter in a week"
as a perfect example you just called achenatx a moron by implying that they would insult twitter employees by implying twitter is trivial.
it's an insult by way of a hypothetically ascribed insult and there's no chance in hell that either of them would trigger sentiment detection because they are so context dependent, even worse it's cultural context not textual context
I don't know about English, but in other languages you need to know the context to distinguish "hate" speech and insults. If I call someone "You, motherf*er!", without context you don't know if I'm insulting that person or just acknowledging my friend who just made a great joke.
That's a truly amazing viewpoint, I honestly can't imagine how one could express the solution that clearly.
In case you can't guess, I'm not serious. However if you download a sentiment analysis model and feed it my first paragraph it'll claim it was positive.
Sentiment analysis is a really really really hard problem, especially for short texts.
Everyone is assuming OP is meaning ML and that they're in the just-enough-knowledge-to-be-dangerous phase.
The problem that needs solving isn't "catch anything that could, upon deciphering, hurt someone's feelings a bit". It's "catch enough despicable or aggressive comments before they cause problems for others".
The later is easily doable because part of the signal is the interactions and you only need to damp bad interactions down until they aren't self-sustaining wars across the feeds of the uninterested, not sanitize every post so that they're all toddler-safe.
Once you stop trying to prevent bad thoughts and switch to trying to create a good forum it becomes tractable at any scale.
This makes for a very civil, but bland conversation. People engage with negative/hate tweets.
I personally like talking politics without insults, but most people are incapable of it.
It is easy to detect insults.