\n<\/p>\n
Listen carefully to a spoken conversation and you\u2019ll notice that the speakers use a lot of little quasi-words \u2014 mm-hmm<\/i>, um<\/i>, huh?<\/i> and the like \u2014 that don\u2019t convey any information about the topic of the conversation itself. For many decades, linguists regarded such utterances as largely irrelevant noise, the flotsam and jetsam that accumulate on the margins of language when speakers aren\u2019t as articulate as they\u2019d like to be.<\/p>\n
But these little words may be much more important than that. A few linguists now think that far from being detritus, they may be crucial traffic signals to regulate the flow of conversation<\/a> as well as tools to negotiate mutual understanding. That puts them at the heart of language itself \u2014 and they may be the hardest part of language for artificial intelligence to master.<\/p>\n \u201cHere is this phenomenon that lives right under our nose, that we barely noticed,\u201d says Mark Dingemanse<\/a>, a linguist at Radboud University in the Netherlands, \u201cthat turns out to upend our ideas of what makes complex language even possible in the first place.\u201d<\/p>\n If you’re enjoying this article, consider supporting our award-winning journalism by subscribing<\/a>. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.<\/p>\n For most of the history of linguistics, scholars have tended to focus on written language, in large part because that\u2019s what they had records of. But once recordings of conversation became available, they could begin to analyze spoken language<\/a> the same way as writing.<\/p>\n \u201cIf you want to have streamlined conversations, these are the tools you need.\u201d \u2014Mark Dingemanse, linguist<\/i><\/p>\n<\/blockquote>\n When they did, they observed that interjections \u2014 that is, short utterances of just a word or two that are not part of a larger sentence \u2014 were ubiquitous in everyday speech. \u201cOne in every seven utterances are one of these things,\u201d says Dingemanse, who explores the use of interjections <\/a>in the 2024 Annual Review of Linguistics<\/i>. \u201cYou\u2019re going to find one of those little guys flying by every 12 seconds. Apparently, we need them.\u201d<\/p>\n Many of these interjections serve to regulate the flow of conversation. \u201cThink of it as a tool kit for conducting interactions,\u201d says Dingemanse. \u201cIf you want to have streamlined conversations, these are the tools you need.\u201d An um<\/i> or uh<\/i> from the speaker, for example, signals that they\u2019re about to pause, but aren\u2019t finished speaking. A quick huh?<\/i> or what?<\/i> from the listener, on the other hand, can signal a failure of communication that the speaker needs to repair.<\/p>\n That need seems to be universal: In a survey of 31 languages around the world, Dingemanse and his colleagues found that all of them used a short, neutral syllable similar to huh?<\/i> as a repair signal<\/a>, probably because it\u2019s quick to produce. \u201cIn that moment of difficulty, you\u2019re going to need the simplest possible question word, and that\u2019s what huh?<\/i> is,\u201d says Dingemanse. \u201cWe think all societies will stumble on this, for the same reason.\u201d<\/p>\n Other interjections serve as what some linguists call \u201ccontinuers,\u201d such as mm-hmm<\/i> \u2014 signals from the listener that they\u2019re paying attention and the speaker should keep going. Once again, the form of the word is well suited to its function: Because mm-hmm<\/i> is made with a closed mouth, it\u2019s clear that the signaler does not intend to speak.<\/p>\n Sign languages often handle continuers differently, but then again, two people signing at the same time can be less disruptive than two people speaking, says Carl B\u00f6rstell<\/a>, a linguist at the University of Bergen in Norway. In Swedish Sign Language, for example, listeners often sign yes<\/i> as a continuer for long stretches, but to keep this continuer unobtrusive, the sender tends to hold their hands lower than usual<\/a>.<\/p>\n Different interjections can send slightly different signals. Consider, for example, one person describing to another how to build a piece of IKEA furniture, says Allison Nguyen<\/a>, a psycholinguist at Illinois State University. In such a conversation, mm-hmm<\/i> might indicate that the speaker should continue explaining the current step, while yeah<\/i> or OK<\/i> would imply that the listener is done with that step and it\u2019s time to move on to the next.<\/p>\n Continuers aren\u2019t merely for politeness \u2014 they really matter to a conversation, says Dingemanse. In one classic experiment from more than two decades ago, 34 undergraduate students listened as another volunteer told them a story. Some of the listeners gave the usual \u201cI\u2019m listening\u201d signals, while others \u2014 who had been instructed to count the number of words beginning with the letter t<\/i> \u2014 were too distracted to do so. The lack of normal signals from the listeners led to stories that were less well crafted<\/a>, the researchers found. \u201cThat shows that these little words are quite consequential,\u201d says Dingemanse.<\/p>\n Nguyen agrees that such words are far from meaningless. \u201cThey really do a lot for mutual understanding and mutual conversation,\u201d she says. She\u2019s now working to see if emojis serve similar functions in text conversations.<\/p>\n The role of interjections goes even deeper than regulating the flow of conversation. Interjections also help in negotiating the ground rules of a conversation. Every time two people converse, they need to establish an understanding of where each is coming from: what each participant knows to begin with, what they think the other person knows and how much detail they want to hear. Much of this work \u2014 what linguists call \u201cgrounding\u201d \u2014 is carried out by interjections.<\/p>\n \u201cWe are building common ground, and we are taking turns.\u201d \u2014Martina Wiltschko, theoretical linguist<\/i><\/p>\n<\/blockquote>\n \u201cIf I\u2019m telling you a story and you say something like \u2018Wow!\u2019 I might find that encouraging and add more detail,\u201d says Nguyen. \u201cBut if you do something like, \u2018Uh-huh,\u2019 I\u2019m going to assume you aren\u2019t interested in more detail.\u201d<\/p>\n A key part of grounding is working out what each participant thinks about the other\u2019s knowledge, says Martina Wiltschko<\/a>, a theoretical linguist at the Catalan Institution for Research and Advanced Studies in Barcelona, Spain. Some languages, like Mandarin, explicitly differentiate between \u201cI\u2019m telling you something you didn\u2019t know\u201d and \u201cI\u2019m telling you something that I think you knew already.\u201d In English, that task falls largely on interjections.<\/p>\n One of Wiltschko\u2019s favorite examples is the Canadian eh?<\/i><\/a> \u201cIf I tell you you have a new dog, I\u2019m usually not telling you stuff you don\u2019t know, so it\u2019s weird for me to tell you,\u201d she says. But \u2018You have a new dog, eh?\u2019 eliminates the weirdness by flagging the statement as news to the speaker, not the listener.<\/p>\n Other interjections can indicate that the speaker knows they\u2019re not giving the other participant what they sought. \u201cIf you ask me what\u2019s the weather like in Barcelona, I can say \u2018Well, I haven\u2019t been outside yet,\u2019\u201d says Wiltschko. The well<\/i> is an acknowledgement that she\u2019s not quite answering the question.<\/p>\n Wiltschko and her students have now examined more than 20 languages, and every one of them uses little words for negotiations like these. \u201cI haven\u2019t found a language that doesn\u2019t do these three general things: what I know, what I think you know and turn-taking,\u201d she says. They are key to regulating conversations, she adds: \u201cWe are building common ground, and we are taking turns.\u201d<\/p>\n Details like these aren\u2019t just arcana for linguists to obsess over. Using interjections properly is a key part of sounding fluent in speaking a second language<\/a>, notes Wiltschko, but language teachers often ignore them. \u201cWhen it comes to language teaching, you get points deducted for using um<\/i>s and uh<\/i>s, because you\u2019re \u2018not fluent,\u2019\u201d she says. \u201cBut native speakers use them, because it helps! They should be taught.\u201d Artificial intelligence, too, can struggle to use interjections well, she notes, making them the best way to distinguish between a computer and a real human (see below).<\/p>\n And interjections also provide a window into interpersonal relationships. \u201cThese little markers say so much about what you think,\u201d she says \u2014 and they\u2019re harder to control than the actual content. Maybe couples therapists, for example, would find that interjections afford useful insights into how their clients regard one another and how they negotiate power in a conversation. The interjection oh<\/i> often signals confrontation, she says, as in the difference between \u201cDo you want to go out for dinner?\u201d and \u201cOh, so now you want to go out for dinner?\u201d<\/p>\n Indeed, these little words go right to the heart of language and what it is for. \u201cLanguage exists because we need to interact with one another,\u201d says B\u00f6rstell. \u201cFor me, that\u2019s the main reason for language being so successful.\u201d<\/p>\n Dingemanse goes one step further. Interjections, he says, don\u2019t just facilitate our conversations. In negotiating points of view and grounding, they\u2019re also how language talks about talking.<\/p>\n \u201cWith huh?<\/i> you say not just \u2018I didn\u2019t understand,\u2019\u201d says Dingemanse. \u201cIt\u2019s \u2018I understand you\u2019re trying to tell me something, but I didn\u2019t get it.\u2019\u201d That reflexivity enables more sophisticated speech and thought. Indeed, he says, \u201cI don\u2019t think we would have complex language if it were not for these simple words.\u201d<\/p>\n To make artificial intelligence sound more natural, developers are building interjections into its responses. Google\u2019s NotebookLM<\/a>, for example, offers the option of summarizing information \u2014 say, one or more scientific papers \u2014 in the form of a podcast hosted by two AI-generated hosts.<\/p>\n On first hearing, the program does a pretty good job: The hosts joke, laugh and insert Mm-hmm<\/i> and Wow!<\/i> at superficially appropriate times. But to the ears of a trained linguist, there\u2019s something amiss. (Listen to an example<\/a>.)<\/p>\n \u201cThey almost work, but not quite,\u201d says theoretical linguist Martina Wiltschko of the Catalan Institution for Research and Advanced Studies. \u201cThey kind of remind me of AI art, where there\u2019s too many fingers. You don\u2019t see it at first, but if you look carefully, you see that something\u2019s wrong.\u201d<\/p>\n One tell is that when the listener mm-hmm<\/i>s or laughs, the speaker pauses while they do so, lending a slightly creepy note to the simulated conversation. \u201cTo me, it\u2019s almost like the uncanny valley,\u201d says Wiltschko. \u201cIt\u2019s close, but it\u2019s not quite close enough.\u201d<\/p>\n The biggest deficiency, though, is a robust sense of who knows what in the conversation. The AI hosts seem to flip back and forth on which one knows which pieces of information.<\/p>\n \u201cIt\u2019s not just what they\u2019re saying, it\u2019s who\u2019s talking in what context, and who knows what,\u201d says Wiltschko. \u201cI would be really surprised if the AI could ever handle that \u2014 and human beings handle that with ease.\u201d<\/p>\n This article originally appeared in Knowable Magazine<\/i><\/a>, an independent journalistic endeavor from Annual Reviews. Sign up for the newsletter<\/a>.<\/p>\n<\/div>\nOn supporting science journalism<\/h2>\n
\n
Wow! There\u2019s more<\/h2>\n
\n
\nCan AI learn to use interjections?<\/h2>\n
\n