@jserv, one thing to do is link to your question here in chat, so people know what you're referring to. Are you asking about this: Calculating a 95% confidence interval without the original dataset?

@gung-ReinstateMonica Yes, that's the question

@gung-ReinstateMonica This is not a homework question, and as far as I know, not something that can be easily looked up in a statistics textbook or as part of a statistics course. I have already explained the extent of what I understand in the original post, and further attempts in my later edit. This is a question specifically about whether it is possible to calculate a confidence interval figure from the paper I provided.

@Glen_b I still see a count for you. For example, stats.stackexchange.com/review/low-quality-posts/stats .

@jserv, I was responding to, "The question reads..." That sounds like a homework-type question. Note that our policy doesn't require that it is literally homework. What is the context for this question? Who asked you to compute the CI? Where was this question written?

@gung-ReinstateMonica The question came from an employer e-mail, but as far as I'm aware SE doesn't usually require you to say where the question came from. It sounds like I'd have to be extremely careful about my phrasing in the future if someone else has addressed a question to me and I bring it here

@jserv, I don't think you have to be extremely careful, & SE doesn't require you to say where questions come from. I'm just wondering why you have a question that sounds so much like a homework assignment. You write, "I have been asked to calculate a 95% confidence interval ... The question reads:", followed by what sounds like a homework question. That's why I ask.

Most of the people here answering questions teach stats. I know students sometimes go online to get someone to give them answers. I don't like it in my students when I catch them, & I don't want to do people's homework for them here. I'm not the only one that feels that way.

It was suggested to me that chat would be a better place to get answers to this type of question: stats.stackexchange.com/questions/561605 I have my doubts but whatever, I'll give it a try: I'm a professional programmer with basically no prior ML experience. I have a problem where ML seems like a likely tool for solving it and I'm trying to get something better than a random guess as to how much work it would be to lean enough ML to try and solve the problem.

See the link for the long version, but the short version of the problem is: help a single writer to write dialog that sounds like different people, and not just accents but things like idioms, word choice and complexity of speech.

If it's not homewok, you have nothing to worry about.

@BCS, I think the amount of work to do that would be prohibitive.

@gung-ReinstateMonica more details? Prohibitive at what level?

I know of examples of prior work for most of the sub components, author identification being the main one. Also note that I expect a human to be in the loop for other reason. For the editing, full automation is an anti-goal. Generating annotations like "this phrase doesn't seem right, the indicated speaker would likely use words like X, Y or Z in this context" would count as a valid solution.

Come to think of it, a plagiarism checker that works by spotting author switches would be about 70% of what I'm looking for.

@BCS building your own language model from zero is probably prohibitive and not necessary. you should look into adapting a large scale language model like GPT and figuring out a way to get it to make the kind of text you're interested in

@whuber Thanks. I wonder whether there was a temporary issue or whether I just got confused about what number I was looking at before. I'll assume the latter for now.

@Sycorax I have a suspicion that the hard part of that approach would be getting the language model to retain the properties of the input that the author wants it too. Figuring out what that constraint even means is likely hard. What I'm thinking of directing the human author to chunks of text that need to be re-written with some vague hints as to what direction to tweak things.

Having the human iterate that while manually retaining the needed properties seem much simpler than trying to figure out how to even encode the needed properties.