## Disfluencies and smiles

A couple of days ago, in a café in Paris,  someone noticed a young woman intently watching the Clinton/Trump debate, and commented "Isn't watching the debate so much better than working?" But the debate watcher was Ye Tian, a postdoc at the Laboratoire de linguistique formelle, Université Paris Diderot (Paris 7), part of a project whose acronym is DUEL — "Disfluencies, Exclamations and Laughter in Dialogue". And so her interest in the video was a professional one, with preliminary results that she published as a blog post here. Ye Tian's analysis is reproduced below, with her permission, as a guest post.

The Monday night presidential debate between Donald Trump and Hillary Clinton was apparently the most watched debate in American history. When I woke up today, my facebook was “刷屏” (screen-painted) by everyone’s opinion on this, so I sat down, opened youtube and started watching. OK I didn’t just watch. I thought, why don’t I check out their disfluency patterns, and whether there were any smiles and laughter in this presumably hostile interaction? This took me a whole day. Someone in the cafe saw me watching the video and said, “isn’t watching the debate so much better than working?”. I thought, “this is working!”

So here are some of my initial observations. People have the impression that Donald Trump is way more disfluent than Hilary Clinton, that he has a lot of incomplete sentences and that he repeats himself a lot. This is partly true. Trump and Clinton have very different types of disfluency. However, in terms of occurrences of disfluencies, they are not so different.

First, a few words of types of disfluencies. First there are silent and filled pauses. Pauses can be filled by things like “um” and “uh”, but also by “discourse markers” such as “you know”, “I mean”, “like” etc. Second, there are repetitions. We sometimes repeat parts of what we said – can be as small as a syllable, or as large as a clause. Third, there are repairs. We may say something, stop, and trace back to change what we said. We may repair “small things” like a morpheme, such as a plural marker. Here is an example from Trump: “(the thing + the things) that business as in people like the most is the fact that I’m cutting regulation”. Note that I have annotated the repair in the form of (to-be-repaired + repair). Here Trump repaired “the thing” by adding a plural ending. We may also repair big chunks. Here is another example from Trump: “We have endorsements from, I think, almost every police group, very — I mean, a large percentage of them in the United States”. Here Trump first said “almost every police group”, and then changed it into “a large percentage of them”. Lastly, there are abandoned (incomplete) utterances. For example, Trump said “we have made so many bad deals during the last — so she’s got experience, that I agree.”. The first sentence was incomplete (“the last…”).

So I counted all filled pauses, repetitions, repairs and abandoned utterances. Note that I didn’t analyse interruptions, or disfluencies during cross talk (obviously there were a lot of repetitions and abandoned utterances from both of them during cross talk). Trump had a fair number of repetitions, repairs and abandoned utterances, and relatively few filled pauses. He often stops mid-sentence to insert extra information, called asides or parentheticals. Often they are anecdotes or comments about himself (“he called me the other day” or “I’m not going to get credit for it”), and he may or may not come back to his original sentence afterwards. Clinton, on the other hand, had almost no repairs or abandoned utterances, a few repetitions, and many more filled pauses. Overall, Trump had 67 disfluencies while Clinton had 53. Mind you, I haven’t counted the total number of words (and I suspect Trump said more). So their overall rates of disfluencies may be the same.

So here is your impression confirmed. Trump tends to repeat himself, he often stops mid-sentence to add something, and may or may not come back to his original partial sentence. He also often changes his mind about what he said (repairs). This is why he doesn’t come across as a well-prepared and eloquent speaker. Below are some examples:

Repetition:

TRUMP: New York — New York has done an excellent job. And I give credit — I give credit across the board going back two mayors.

Repetition with inserted asides:

TRUMP: And Sean Hannity said — and he called me the other day — and I spoke to him about it — he said you were totally against the war, because he was for the war.

Repair:

TRUMP: They (left + fired ) 1,400 people.

TRUMP: I could name + { I mean} there are thousands of them.

Abandoned utterance:

TRUMP: The African-American community — because — look, the community within the inner cities has been so badly treated.

TRUMP: whether it’s — I mean, I can just keep naming them all day long — we need law and order in our country.

Clinton, on the other hand, speaks more slowly, and she uses more filled pauses than Trump. The filled pauses, however, are not evenly distributed. There are long stretches of speech without a single filled pause, but there are pockets of utterances where filled pauses are frequent. One example was her discussion about cyber crime. Here is one paragraph from her, fillers are annotated like { F uh}.

CLINTON: Well, I think we need to do much more {F uh} with our tech companies to {F uh} prevent ISIS and their operatives {F uh} from being able to use the Internet to radicalize, even direct {F uh} people in our country and Europe and elsewhere. But we also have to intensify our air strikes against ISIS {F uh} and eventually support our Arab and Kurdish {F uh} partners to be able to actually take out ISIS {F uh} in Raqqa. {F uh} And we’re hoping that {F uh} within the year we’ll be able to push ISIS out of Iraq and then, you know, really squeeze them in Syria.

Looking at disfluency patterns over time, it shows that Trump’s disfluency increases steadily over time, while Clinton’s disfluency fluctuates, peaking at 60 to 75 minutes window, when they were discussing cyber crime and fighting ISIS.

What does this disfluency difference between Trump and Clinton mean? Do repetitions /repairs /abandoned utterances suggest a less clear mind? a freer mind?, a mind that gets distracted by its own thoughts? Do fluctuations in rates of filled pauses indicate fluctuations in confidence? I don’t know. The viewers were clearly annoyed by Trump’s disfluency much more than by Clinton’s. Trump style disfluency – repetitions, inserted asides, repairs and abandoned utterances – affects discourse coherence. It doesn’t necessarily mean his mind is incoherent, but it IS a style that is more egocentric and less considerate. It indicates less initial planning and preparation. Clinton’s disfluency, namely filled pauses, indicates the opposite: planning. She pauses in order to make the upcoming utterance clear. So Trump was just externalizing his inner (rather free and unique) trail of thoughts, but Clinton was aiming for getting the exact ideas across to us: she cared about how OUR trail of thoughts changes as a result of her speech.

Now, what about smile and laughter? Were there any? Of course. One widely held impression was that Clinton had to smile a lot while waiting for Trump’s nonsense. And again, the impression was confirmed! She did smile a lot, and very often for looonnnggg stretches of time!

I found a total of 42 smiles and laughs of Trump and Clinton (there were also 8 audience laughs). And yes, most of them came from Clinton (74%)! The total duration of Clinton’s smile/laugh was 124 seconds, compared to Trump’s total of 14 seconds (nearly 10:1). Compare to friendly conversations, while laughter happens between 10 – 50 times per 10 min, this debate was a smile/ laughter desert (at 0.5 times per 10min including smiles). In friendly conversations, there a lot of “dyadic” laugh, meaning when one person laughs, the other often joins in. In this debate there was only one occasion where both joined the smile/laugh. Trump had just said a lot of bad things about Clinton’s temperament. Clinton responded “Whew, OK”, followed by a laugh (and some shoulder wiggling). At that moment, the audience laughed and Mr Trump smiled at her. Here is a screen shot.

Very sweet huh? This was the only not-so-hostile laughter sharing moment between the two. All the other smiles and laughs communicates something hostile, or at least, non-cooperative. And of course when one does it, the other wouldn’t join. The most frequent “meaning” of smiles and laughs in this debate can be paragraphed as “ridiculous”. It is often when one person had said something about the other (often Trump was the speaker), and the other smiles to say “RIDICULOUS”.

Both Trump and Clinton used smiles and laughs in this way, but they look very different. Trump never showed his teeth. Very often he just lifted the corners of his mouth, and there was nothing around his eyes, which makes his smiles look “disingenuous”. Sometimes his smiles were accompanied with eye rolling or head shaking. Compared to Clinton, Trump’s smiles were much short (on average 1.5 seconds). Here are some of his signature smiles:

1. Trump smiling to Clinton’s “he owes about $650 million to Wall Street and foreign banks”: 1. Trump smiles to Clinton’s “I was so shocked that Donald publicly invited Putin to hack into Americans”: 1. Clinton said “I have put forth a plan to defeat uh ISIS”, and Trump reacted… In comparison, Clinton smiled a lot, and each time for a long time. Her average smile/laugh duration was 4.5 seconds (compared to 1.5 seconds of Trump), and the longest smile was 15 seconds long, 15 seconds long!!! That’s much longer than a usual natural smile! Also, though the function of her smiles and laughs were the same as Trumps – to dismiss what Trump just said, to communicate “that’s ridiculous” – her smiles look much more friendly. If I didn’t give you a context, you may well think she was hosting a party, or was talking to a friendly neighbour. Look: 1. Trump said “And you’re going to stop them [ISIS]? I don’t think so.”, and Clinton smiled, for 10 seconds: 2. Trump said “All of the things that she’s talking about could have been taken care of during the last 10 years, let’s say, while she had great power. But they weren’t taken care of. And if she ever wins this race, they won’t be taken care of”. Clinton smiled (at least 3 seconds, as the camera moved to the host). 6. Trump said “$200 million is spent [by Clinton], and I’m either winning or tied, and I’ve spent practically nothing”. Clinton smiled for 3 seconds.

So these were my initial observations. Trump and Clinton were disfluent in different ways. Trump tended to repeat, repair and abandon his sentences mid-air. Clinton used more filled pauses, but these filled pauses clustered in some stretches and were absent in others.

Trump didn’t show a great smile: no teeth, no eyes, and not so many. Clinton, on the other hand, smiled often and she kept them long. Her smiles were so friendly-looking, one might forget that linguistically they served the same functions as Trump’s bitter smirks.

I understand Trump’s bitter smiles. Clinton’s remarks were not friendly, why should Trump’s smiles be sweet? But why did Clinton have so many sweet and long smiles, after so many harsh attacks from Trump? Ahhh maybe her smiles were for the audience. She was strategically saying to us “look at how stupid Trump is, we (inclusive) are much better than him”. So Trump’s smiles stayed within their interaction, but Clinton’s smiles reached out. Did it work? What do you think😉

Above is a guest post by Ye Tian.

1. ### Mark Liberman said,

September 30, 2016 @ 8:52 am

Since I counted the words in the debate in an earlier post ("Debate words", 9/27/2016), I can fill in Ye's suggestion about overall rates. Trump used 8866 total words, so his 67 disfluencies occurred at a rate of 7.6 per thousand words. Clinton used 6580 total words, so her 53 disfluencies occurred at a rate of 8.1 per thousand words.

So the rates are indeed fairly similar — but the most striking thing about the overall rates is that both are relatively low. Thus in 13 FreshAir interviews that I transcribed carefully, the (very fluent and experienced) host Terry Gross used filled pauses at an overall rate of 9.6 per thousand words, and "fluent self-corrections" (that is, breaking off in the middle of a word and then restarting that word or switching to another one) at an overall rate of 26.9 per thousand words. Her guests used filled pauses at an overall rate of 17.5 per thousand words, and fluent self-corrections at an overall rate of 29.4 per thousand words.

Some details, from various stages of the analysis process, can be found in

"UM/UH accommodation", 11/24/2015
"Like thanks", 11/26/2015

2. ### Mike said,

September 30, 2016 @ 9:05 am

And those Fresh Air interviews are likely edited to remove some disfluencies, so host and guests' actual rates may be even higher.

3. ### Laura Morland said,

September 30, 2016 @ 10:01 am

Fascinating post!

It could, however, be improved by qualifying Clinton's smile. The writer states: "Trump didn’t show a great smile: no teeth, no eyes, and not so many. Clinton, on the other hand, smiled often and she kept them long. Her smiles were so friendly-looking…."

The nature of Trump's smile is well-detailed, but Clinton's only as "friendly-looking." Further, did Trump NEVER display a "teeth-visible smile" during the entire 90 minutes? That would be information worth noting.

Finally, I'd like to examine the following comment: "But why did Clinton have so many sweet and long smiles, after so many harsh attacks from Trump? Ahhh maybe her smiles were for the audience."

Again, Clinton's smile is described here as "sweet" without any evidence for what looks to be a subjective claim.

I would support Ye Tian's supposition that Clinton's smiles were directed at the audience (both the audience in the room and the audience at home), but I would respectfully suggest that her facial expressions were likely a result of her "intense debate preparation". And if it is true that her smiles were forced, it is all the most impressive that they come across as "sweet" and "friendly-looking."

It would be useful to compare Clinton's smiles and laughs (perhaps excepting the now-famous "shimmy") during the debate with her spontaneous facial expressions in this video (taken and posted yesterday): https://www.youtube.com/watch?v=vjK8yeMTo2Y (For those not in the know, at 0:06 she is mocking the inability of the Libertarian candidate, Gary Johnson, to answer the same question the night before.)

4. ### Ye Tian said,

September 30, 2016 @ 10:37 am

@Laura Morland Thanks for the comment! And good points! I don't think there was ever a smile showing teeth from Trump in the 90min. According to my annotation. He had one voiced laugh after Clinton said "so he has a long record of engaging in racist behavior.", and the rest were just low-arousal smiles. The only smile involved a bit of his eyes was the one after Clinton said "Whew, OK". You are right, calling Hilary's smiles sweet is completely subjective. I meant that while many of Trump's smiles won't be perceived as friendly ones out of context, hers certainly would.
And I agree with you, I think Clinton probably prepared every aspects- from content to form to gestures and facial expressions, in order to deliver a composed, intelligent and calm picture of hers, all with a lot of effort.

5. ### Riikka said,

September 30, 2016 @ 11:24 am

Apparently she has earlier been repeatedly told to smile. Also, considered that everyone already knows that she's qualified as far it comes to the substance matter, one of her most important goals were to look personal, relatable and trustworthy. And I suppose that in USA it means smiling like a tooth paste commercial, especially if you are a woman..

6. ### AntC said,

September 30, 2016 @ 6:53 pm

@Laura M I would respectfully suggest that her [Clinton's] facial expressions were likely a result of her "intense debate preparation".

So I don't have a vote in the Presidential race, but I am sure going to suffer/benefit from the effects of whoever gets elected. I want a President "of the free world" to be well-prepared, in fact intensely well prepared for every debate. And culturally sensitive/communicative to their audience — whether it be a studio/TV audience or the U.N. or the world's tyrants, despots and megalomaniacs and their henchmen [**] who are a threat to "the free world".

I do not want a President who bumbles into every encounter and blusters with 'negotiating positions' that can be overturned at whim; who has a loose relationship with truth; and thinks issues of security and democracy and financial support are merely deal-making.

[**] henchmen: apologies for the non-gender-neutral term, but nearly all people describable as hench- are men.

7. ### Michael Watts said,

September 30, 2016 @ 10:55 pm

I'm curious about the decision on which disfluencies to track. We've enumerated 5 types of disfluency: silent pauses, filled pauses, repetitions, repairs, and abandonings. My personal experience suggests that silent pauses are actually the most serious of those types of disfluency, from the perspective of how the audience perceives them. Giving presentations in school, my audiences tended to take silent pauses as a cue that I was entirely finished speaking, which was not my intent.

But silent pauses have been left out of the counts. There is a pretty good argument for this, in that silent pauses, unlike repairs and abandonings, might often be intentional rather than disfluent — but I believe this is also true of repetition.

For example, the example given earlier, "New York — New York has done an excellent job", looks like a disfluency when punctuated that way. If I change the punctuation to "New York! New York has done an excellent job", it looks like an intentional rhetorical device.

Of course, listening to the speech, it might be much easier to distinguish disfluent pauses/repetitions from intentional ones.

My instinct is that Clinton's high rate of filled pauses and very low rate of repairs and abandonings suggests she is trying to deliver memorized, targeted speech with defined wording. In other words, not only has she come into the debate with a list of points to make, she's come in with a bunch of professionally-written blurbs that she wants to use to make those points. (Alternatively, given that they're both displaying low disfluency rates overall, it suggests she's much better at sticking to the goal of delivering memorized blurbs than Trump is.)

8. ### AntC said,

October 1, 2016 @ 12:09 am

@Michael W audiences tended to take silent pauses as a cue that I was entirely finished speaking, …

Yes that's certainly a danger for politicians — especially in a Westminster-style bear-pit like Prime Minister's Question Time.

Notably Margaret Thatcher had a technique for that. She would put her pauses mid-sentence or mid-phrase, without a falling intonation — like she was going to say more (which she was) — then picked up at the same intonation level; completed the phrase and went straight on to the next. Those who would interrupt her were flummoxed.

(I guess that she already had the completion of the phrase in mind; but held it back to make time to prepare the next.)

To bring my rant (above post) to a linguistic point: if you're negotiating for peace/disarmament/human rights with (insert tyrant of choice), you don't want an ill-phrased off-the-cuff remark to tip some unstable regime into drastic action. It's all very well to claim retrospectively you were being sarcastic. Retracting the missiles not so much.

October 2, 2016 @ 9:58 pm

Like Michael Watts I am curious about some of the counting. Specifically, I wonder how you identify "pauses filled by… 'discourse markers'". Many discourse markers (e.g. 'like', 'you know', 'so') occur both in filled pauses and in planned speech. Are filled pauses identified independently of the word or sound that fills them? Are certain lexical items re-examined as likely indicators of filled pauses? Have you published (or posted) work that details this aspect of your methodology?

Thanks for a very interesting analysis.

10. ### Ye Tian said,

October 3, 2016 @ 3:23 am

@Michael Watts, good point about silent pauses. I will take them into account next time. I agree, silent pauses definitely have an effect. Very often there are silent pauses surrounding filled pauses (more for Clinton than Trump), and I did note those down. In terms of stand-alone silent pauses, from what I remembered, there were fewer of them than filled ones for both of them, though I definitely remember Clinton having a few silent pauses. One of them was "He even said, well, you know, if there were nuclear war in theee [silent pause] East Asia, well, you know, that's fine". I remember this one because the silent pause was very obvious.

In terms of other types of disfluencies, I only count those that sounds like disfluency, rather than, say, repetition used for emphasis. The annotation was done via video/audio rather than transcript.

And I agree with you, I also think Clinton has prepared pre-written speech, lots of it. And Trump really hadn't.

11. ### Ye Tian said,

October 3, 2016 @ 3:28 am

@Chad Nilep good point about what counts as fillled pauses. Theory-wise, I believe there whether something is a filled pause is a continuous rather than binary feature. It should be based on distribution of a term rather than semantic meaning. So "like" could potentially be just as much as a filled pause as "uh". However, for this analysis, I only counted "uh" and "um" (way more "uh" than "um). But I should probably have counted both "uh" and "um", and "you know" and "I mean", "well" etc. Something for the second and last debate!!

12. ### James Wimberley said,

October 4, 2016 @ 9:40 am

Ms. Ye's university fairly recently renamed itself from a Roman numeral (Paris-something) to "Denis Diderot". In this case the conjunction is felicitous. The author of Jacques le Fataliste – for my money a more engaging philosophical satire than the more famous Candide – would have enjoyed her analysis. Definitely a man you would have wanted to split a bottle with.