I read the study in question, That's What She Said: Double Entendre Identification by Chloé Kiddon and Yuriy Brun, and found the experience quite startling. I usually enjoy a certain ease with scientific literatures, stretching myself out against the research with the languid, unwarranted comfort of a loutish lower class hooligan lounging on an aristocrat's sofa, although I don't quite bring the same lustful hostility to the encounter. But the article under discussion was from a very different literature, and I found myself suddenly and deeply sympathetic to those who are suspicious of and discomfited by a literature with which they are unfamiliar, rather like an aristocrat who isn't sure about the intentions of the sweating, bulging young lout on her sofa, smacking his lips and speaking a language that very much sounds like English but includes tantalising words that make her ears tingle and her cheeks blush.
First, I have to admit that it is entirely possible that the whole thing is a hoax. While much of it makes sense, certain passages are ostensibly incomprehensible, and I am not sure if they just reflect lacunae in my knowledge or if they're just nonsense? Of course, I wouldn't dare reveal a potential lacuna in my knowledge, so I rub my chin and say, "Well that's very, very interesting", fooling everybody but myself, hence, the uneasy feeling that I might be getting hoaxed. It would be as if I read in a blog that an Urdu phrase means "He who is in my backyard is a stranger to me" and then use that in my own blog, without knowing that it actually means "Why yes, I have goats trained to do exactly that." Hoaxes are a form of elaborate prank where vanity about intelligence (a refusal to admit to being confused or misled or not understanding, or, even more fundamentally, an inability to discern when something is incomprehensible) is exposed and mocked.
Second, I find myself pouncing on certain comments where I think I disagree with what the authors are saying and I want to hammer them for their credulity, but I could just be falling into a trap - one that I often accuse others of tumbling into. It is quite possible that where I see a "disagreement", they see no such thing: they may respond, "No, of course we understand what you're saying; but here, we're talking about this in such-and-such a way." Let me quote from a previous post:
skilled readers of the scientific literature often spot weaknesses in the research; foolish readers of the scientific literature also often spot weaknesses in the research, but they don't understand how that research is being used and read and interpreted - they tend to think that everything is being treated as pure gospel truth, instead of being considered, contemplated, and contextualised: a musing editorial can refer to various findings without constantly tugging at the reader's elbow to explain how weak those findings are, because the editorialist trusts the readers to understand that this is musing, about weak findings.I am very familiar with certain types of scientific literature. Sometimes when I read papers, the science may be over my head. Sometimes, when I peer review papers, the science is over my head. Hell, I've been co-author of papers where the science is over my head. But I know the tones, the rhythms, the way language and jargon and citations are being used. Coming up against another literature, I'm thrust into the position of being unsure and of risking naive criticism. For example, when the authors of this paper talk about the "problem" they are addressing, I'm drumming my fingers, frustrated that they've missed the "problem" of innuendo altogether: but we're talking at crossed purposes, just as someone may say that a statistical "problem" I've found interesting in a paper has nothing to do with the "problem" they're worried about. The authors of this study expect their readers to know that what they are calling the "problem" is an interesting one - and, crucially, may be interesting for reasons I cannot fathom; their paper is not intended to resolve all the epistemological, socio-cultural, and linguistic "problems" posed by sexual innuendo.
So, why is any of this relevant? Because, of course, we could also be talking about the relationship between artificial intelligence and comedy. Jokes, even basic ones like puns, depend upon lacunae, the space of the unspoken; although there are positive cognitive shifts, such as when a dual meaning is recognised, the realm of the unspoken includes the background noise of interpersonal registers and the haze of cognitive, semantic, social, cultural, and political relationships. Language is not just a model that can be taken apart, and comedy is not just a certain use of language. That "rod" can mean a hard, slab-like piece of metal or a long pole used for fishing but also a penis is (probably) not in itself funny; the pun comes to life in a context involving semantic rules but also social rules, words but also characters, and so forth.
Mass murderer to cell-mate: I couldn't figure out a way to unpeel my banana last night.
Cell-mate: That's what she said!
Nope, doesn't really work.
Your boss, putting down the phone and, with a look of despair: You're only allowed to come in the back door.
You: That's what she said!
Much better! And it smuggles in a reference to the best boyband ever, Dujour!
The issue for computers understanding comedy is obviously not one of intelligence, if by this we mean funds of knowledge, speed of calculation, capacity to solve problems, and many other features of cognitive processing. Cognitively, a 1970s record player is more sophisticated than I am; cognitively, computers today are more "intelligent" than all but a very few of my readers. You know who you are. But getting a joke, even one as unnuanced as "That's what she said!", requires a nuanced appreciation of multiple facets of the context (in fact, what is funny about "That's what she said!" is rarely the retrospective overlay of sexual content into what was previously thought not to be sexual, but the naively, almost innocently, boorish way it is done, and the faintly strange little drama it creates when one imagines a woman, the floating "she", saying whatever it is that was said). A lack of familiarity with the particularities of language as it is filtered through these other variables is the largest stumbling block for artificial intelligence.
At the outset, Kiddon and Brun acknowledge all this; science is about creating models, through rational approximations that need to be justified and transparent, and through which they can then hone in on the problem they want to solve. Unfortunately for me, I'm not entirely clear on what the "problem" is; I think the problem is one of recognition: can they create a model that correctly identifies sentences suitable for adding "That's what she said!" based on nouns and structure? They constructed DEviaNT (Double Entrendre via Noun Transfer - a much better choice than PERverT, Pun and Entendre Recognition, version Two), which they found to be a "promising" approach, high on precision (which means that of the examples it nabbed from the ether, it got a lot of sentences appropriate for adding "That's what she said!"), if low on recall (it didn't nab that many sentences), compared to other approaches.
Stand-ups don't need to worry just yet about comedian cylons, but they're coming. After all, the authors got some startling results:
However, 2 of the 8 false positives are in fact TWSSs (despite coming from the negative testing data): “Yes give me all the cream and he’s gone.” and “Yeah but his hole really smells sometimes.”
It's just a matter of time before the robots take over and I'm only hoping they're intending to make us laugh.