Welcome to Academia.SE! I do think of myself as theory-oriented, hence why I'd need external support to run experiments on a supercomputer rather than my small-scale verifications. With respect to single-author papers, this one is quite recent yet renowned in NLP, cited as Kudo (2018). To answer your question: no, I have not contacted the professor under whom I did the thesis. She is acknowledged at the end for assessing the thesis from which the paper came, but she was uninvolved in any of the research (since I only saw her 3 times for the whole year).

Because the university mandated periodic updates. With a mentor in the loop, it is more difficult to slack off for those students who are inclined to do so. In any case, as I mentioned in my post, I used these one-hour meetings mostly to discuss the meta (writing style, depth, markup...). The reason why discussions about content fell flat was the same reason why the final product ended up being 177 pages long: there was a lot to be aware of -- too much to communicate orally in one meeting, and the written notes I sent to my mentor in advance were never read.

I didn't mean differently. I meant it exactly as I wrote it, hence orally. Nowhere did I say that I failed to communicate them in writing. I'm not a mathematician, but I have an inkling that they too would rather have a piece of paper or a blackboard at hand to explain their ideas than to handwave, resulting in their peers failing to understand their in principle correct train of thought.

@ThomasBakx I know whose idea the proposal was: it originated as an idea for a hobby project of the PhD student. I figured out after a month of reading literature that the proposal would go nowhere, and hence the overlap between the proposal and the resulting thesis was approximately 0%. I did stick to the same field, but came up with my own topic (which is allowed and not uncommon).

(1) It's not a single paper, it's multiple papers. (2) It is a luxury to be generous. It's easier said than done when 100% of your net worth is in the balance. (3) I'm not asking to make a career from a single paper. I'm asking to capitalise the most possible off of it in order to maximise the opportunities to make a career afterwards. (4) As stated in my post and re-iterated in the edit, I already explicitly verified the idea.

For your TL;DR to be accurate, it would have to read: "Remove all contributions of your co-author from your paper. What is left of your paper? An idea with backing in literature, a correct mathematical formalisation of your idea, an algorithmic implementation, results from small-scale testing of the algorithm, but no large-scale high-resource tests." This is almost the polar opposite of how you wrote it.

I don't understand where you are getting the impression that I didn't formalise my idea. Communicating mathematics orally is extremely difficult because mathematics revolves around manipulating written symbols. My mentor did not formalise my ideas, I did all of that work myself; like I said, his only contribution in a paper describing, formalising and algorithmising the idea is providing access to indirect empirical verification. I already provided the direct empirical verification in my thesis because that was possible on a small scale, and it checks out.

@gidds I know my own language pretty well, but in your honour, I consulted the whole-word count statistics I have for the corpus that was used to train the model I am investigating. incest has 17 116 occurrences, princess has 2 379. pussy has 36 249, Debussy has 4 946. For comparison, the translation of princess (which doesn't contain -inces-) has 63 606 uncapitalised and 44 648 capitalised.

But I'm not the one using the corpus. Imagine you are using a language tool like ChatGPT, about which you are told that it was trained on general German text. Does it speak English? Not inherently, but you notice that any time you ask it to produce an English answer, it strings together a bunch of pornographic words each time. You hypothesise that this is due to an overrepresentation in the training data, but all you have access to are the internals of the model (and even if you have access to the training data, you want to know the fraction of the model focused on these data). That's me.

@gidds I'm going to assume this is a late April Fools, but for those still confused: wouldn't you think it to be peculiar seeing a bunch of English morphemes appear as highly frequent patterns in a corpus whose texts were automatically filtered to only keep non-English texts? And would it not be especially peculiar that most of the full-length English words are unambiguously pornographic, with the partial words also overlapping with them? I'm curious why, out of all webpages on the internet, innocuous English news articles would be the ones overrepresented in the set of misclassifications.

Good suggestions. I hadn't thought of saving these particular sources locally (but that folder will be a conversation-starter for sure!). The result could indeed fit in an ACL shortpaper, but considering this thread has been viewed >2500 times already, the odds of somebody plagiarising the idea are now non-zero.