More about UM/UH on the Autism Spectrum

« previous post | next post »

At a workshop in June, a group of us will be presenting a report that includes this graph:

The x axis is the relative frequency of "filled pauses" UM and UH, from 0% to 8%, and the y axis is the proportion of filled pauses that are UM, from 0% to 100%. The individual plotting characters represent values from transcripts of 100 children's contributions to Q&A segments of a standard diagnostic interview, where the blue Ts are "typically developing" children, the green Ms are male children with an autism spectrum diagnosis, and the red Fs are female children with an autism spectrum diagnosis.

You can find the details in Julia Parish-Morris, Mark Liberman, Neville Ryant, Christopher Cieri, Leila Bateman, Emily Ferguson, and Robert T. Schultz, "Exploring Autism Spectrum Disorders Using HLT", Computational Linguistics and Clinical Psychology 2016.

This all started more than a decade ago because I was looking for real-world measures of sex differences in verbal ability, as part of a general evaluation of stereotypes about language and gender. The relative frequency of "filled pauses" like UM and UH is a possible index of fluency — but word-finding gets slower with age, and the dataset I used was not controlled for the interaction of sex and age, so I did a Breakfast Experiment™ that looked at age as well as sex. And what I found, reported in "Young men talk like old women", 11/6/2005, was that UH usage increases with age, and at every age men use UH more than women — but UM usage decreases with age, and at every age women use UM more than men.

In the summer of 2014, as we started to transcribe and analyze some Autism Diagnosic Observation Schedule (ADOS) interviews,  we found that the relative frequency of UM vs. UH was significantly lower in children with an ASD diagnosis than in typically-developing children. And a few weeks later, following a coffee-break conversation at a dialectology conference, an ad hoc group of researchers from around the world found age and sex effects in the relative frequency of the corresponding filled-pause variants in conversational datasets from British and Scottish English, Dutch, German, Norwegian, Danish, and Faroese.

This pattern —  age grading in a linguistic variable, and a sex difference where female speakers are more like younger speakers — generally indicates a language change in progress. And in those datasets where the apparent-time change (that is, younger speakers are different from older speakers) can be checked against historical recordings, the prediction of historical change in UM/UH frequency has been supported.

The results are surveyed in a blog post ("UM/UH update", 12/13/2014) and more formally and at greater length in a forthcoming journal article (Martijn Wieling, Jack Grieve, Gosse Bouma, Josef Fruehwald, John Coleman, and Mark Liberman, "Variation and change in the use of hesitation markers in Germanic languages", Language Dynamics and Change, forthcoming).

So it's well established that there's a pan-Germanic linguistic change, with UM gaining ground at the expense of UH. But in my opinion, it's still unclear why and how this is happening. Is there a single external forcing function, say the influence of English in American movies and television? Is there some general Euro-American cultural trend, say towards greater politeness or a more female-associated speaking style? Is there some internally-generated dynamics based only on the existence of two common hesitation markers, or somehow driven by the phonetic difference between them? And are there analogous processes in other language families?

It's even less clear to me why there should be the interaction between sex and ASD diagnosis seen in the graph above.

But in my opinion, this uncertainty is a Good Thing, though it bothered a few journal reviewers.

We've established some striking empirical generalizations, and the future investigation of their causes will engage scientific theories (about neurodiversity, gender,  and the dynamics of language and culture) and may produce practical results (for example in ASD diagnosis). Such descriptive explorations are an important part of rational investigation, and in linguistics, their results are generally more durable than new theories are.

I should add that our colleagues at the Center for Autism Research at the Children's Hospital of Philadelphia are joining us in plans to publish the audio, aligned transcripts and annotations described in our report; and we hope that other sites will join us in creating a much larger shareable dataset of this type. As the abstract for our CLPsych paper notes

The phenotypic complexity of Autism Spectrum Disorder motivates the application of modern computational methods to large collections of observational data, both for improved clinical diagnosis and for better scientific understanding. We have begun to create a corpus of annotated language samples relevant to this research, and we plan to join with other researchers in pooling and publishing such resources on a large scale.

The same idea is presented and discussed in another forthcoming conference presentation: Julia Parish-Morris et al.,  "Building Language Resources for Exploring Autism Spectrum Disorders", LREC 2016.

To my non-specialist eyes, "autism" looks more like (regions of) a multi-dimensional space than a one-dimensional spectrum — and it's a space that all of us inhabit, not just the people who end up with a clinical diagnosis. I hope that the creation of a very large collection of comparable conversations, with associated demographic as well as clinical metadata, will help future researchers to learn more about the dimensions of this space that are relevant to speech, language and communication.

The distribution of filled-pause variants is one small but interesting piece of evidence. Thus there's a highly significant overall difference between ASD-diagnosed children and typically-developing children in this feature — but there's little or no difference in the case of female children, and the distribution for male ASD children appears to be bimodal, with one subgroup having essentially the same UM/(UM+UH) ratio as the TD group. The true nature of this pattern will be much clearer when we can examine 1,000 or 10,000 interviews rather than 100, and correlate the results with other features. And there are hundreds of thousands of ADOS interviews recorded every year.


  1. Tim Leonard said,

    April 17, 2016 @ 8:14 am

    I think you mean that the relative frequency of UM vs. UH was significantly *lower* (61% vs. 82%) in children with an ASD diagnosis.

    [(myl) Oops — you're right, of course — fixed now.]

  2. Ethan Bradford said,

    April 17, 2016 @ 10:32 am

    For my speech, I worry that "uh" sounds stupid, so I modulate it to "um" when I think of it. If that is somewhat universal, it could explain several of your reported divergences.

  3. Brian Roark said,

    April 17, 2016 @ 12:40 pm

    FYI, the Gorman et al. IMFAR presentation that you mention in your original post from 2014 is now a paper in Autism Research, available online here:

  4. Rubrick said,

    April 18, 2016 @ 3:00 pm

    But in my opinion, this uncertainty is a Good Thing, though it bothered a few journal reviewers.

    Their being bothered bothers me, and I think your rejoinder — "their results are generally more durable than new theories are" — is spot on. What such results don't do as well is generate buzzy, misleading PR copy. "Old People and Women Say 'Uh' More Because of Texting" would play much better for the masses.

RSS feed for comments on this post