Linguistic synchrony
« previous post | next post »
Bruce Doré and Robert Morris, "Linguistic Synchrony Predicts the Immediate and Lasting Impact of Text-Based Emotional Support", Psychological Science 2018:
Emotional support is critical to well-being, but the factors that determine whether support attempts succeed or fail are incompletely understood. Using data from more than 1 million support interactions enacted within an online environment, we showed that emotional-support attempts are more effective when there is synchrony in the behavior of support providers and recipients reflective of shared psychological understanding. Benefits of synchrony in language used and semantic content conveyed were apparent in immediate measures of support impact (recipient ratings of support effectiveness and expressions of gratitude), as well as delayed measures of lasting change in the emotional impact of stressful life situations (recipient ratings of emotional recovery made at a 1-hr delay). These findings identify linguistic synchrony as a process underlying successful emotional support and provide direction for future work investigating support processes enacted via linguistic behaviors.
Where the data was from [link revised to work outside the Psychological Science site]:
Participants in this study were users of an online application, called Koko (previously Panoply), that facilitates text-based emotional-support interactions within an anonymous social network (Morris et al., 2015). We analyzed data from every user who interacted with this application between June 1, 2016, and June 20, 2017; a total of 169,376 unique users posted about 361,139 unique stressful situations and received 1,161,360 messages of support in response.
How they defined "linguistic synchrony":
Our selection of predictor variables was informed by a model positing that synchrony in textual content, style words, emotion words, and latent semantic content reflects a support provider who is able to use language from the recipient’s post, reference relevant actors or objects of the post using the appropriate function words, reference or re-express emotional states expressed by the recipient, and in a broader sense, speak to the semantic content conveyed by the recipient (i.e., semantic content that is not reflected in function words or emotion words).
We defined surface-level textual similarity as the opposite of the Levenshtein distance, which is the minimum number of single-character deletions, insertions, or substitutions required to change one text into another. The greater the Levenshtein distance, the more different two documents are in terms of the text they use.
We defined synchrony in linguistic style (typically called language style matching) as similar use of the function words from the Linguistic Inquiry and Word Count dictionary: negations, quantifiers, conjunctions, adverbs, auxiliary verbs, prepositions, articles, personal pronouns, and impersonal pronouns[…]
We also considered synchrony in the texts’ overall emotional character (i.e., the specific kinds of emotions expressed). We defined the texts’ emotional character using a lexicon-based algorithm that estimates expression of eight categories of emotion—joy, trust, fear, surprise, sadness, disgust, anger, and anticipation—as well as overall positive and negative valence, across 14,182 English words […]
Finally, we considered similarity in latent semantic content—that is, similarity in the kinds of underlying topics that are addressed in the stressor posts and support responses, despite potential differences in the words and phrases used. To do this, we used latent semantic analysis, which provides an algorithmic estimate of document similarity […]
The results ("Functions relating textual, linguistic-style, emotional-content, and latent semantic synchrony, as well as overall support-response valence, to immediate ratings of emotional-support effectiveness, whether a thank-you note was sent, the valence of thank-you-note language, and lasting emotional recovery"):
For some relevant background, see Molly Ireland et al., "Language Style Matching Predicts Relationship Initiation and Stability", Psychological Science 12/13/2010, discussed on LLOG in "Little Words", 6/7/2011.
Rick Rubenstein said,
August 8, 2018 @ 6:13 am
I'm a bit troubled that their method of analysis, taken to its logical conclusion, would seem to imply that the apex of emotional support would be a bot that simply echoed back the users' text verbatim.
[(myl) An excellent point. Though if you look at the detailed results, the effects of Levenshtein distance are equivocal, and it seems that in terms of "Lasting Emotional Recovery", all of the other metrics except Latent Semantic Synchrony are U-shaped, so that more is not necessarily better. Thus the predictions for the verbatim-bot are unclear.
Superficial metrics that work well in typical cases can often be "gamed" to produce strikingly good results for inputs that are actually very bad. Thus in tests of foreign-language fluency, speaking rate in syllables per second correlates very well with evaluators' ratings. But the speaking-rate metric would diverge strongly from human evaluation for someone who just gabbles nonsense very quickly.]
Cervantes said,
August 8, 2018 @ 8:16 am
Motivational Interviewing counselors are trained to use what are called "reflections," of which there are several types. These include nearly verbatim echoing, paraphrasing, and more complex forms but in general they would show a moderate to high degree of synchrony as defined here. Their use is generally perceived to be empathic.
KeithB said,
August 8, 2018 @ 8:37 am
"I'm a bit troubled that their method of analysis, taken to its logical conclusion, would seem to imply that the apex of emotional support would be a bot that simply echoed back the users' text verbatim."
Seems to work for ELIZA
http://psych.fullerton.edu/mbirnbaum/psych101/Eliza.htm
Haamu said,
August 8, 2018 @ 9:25 am
"Synchrony" seems an odd choice for a concept that appears to have no time component (particularly when applied to a potentially asynchronous medium like online/text-based communication). Is it a term of art?
Is the more obvious "synchrony" hypothesis — emotional support is more effective when the asynchronous response is more prompt — too obvious to test?
Cervantes said,
August 8, 2018 @ 12:35 pm
Yes, I was a bit puzzled at first. Synchrony is an existing technical term in linguistics and it's completely unrelated to this, which doesn't seem to be about time anyway. Similarity seems like a better choice.
RfP said,
August 8, 2018 @ 1:39 pm
As a technical writer who is always trying to improve how I connect with my audience, I wonder how much this kind of result maps into situations that are not thought of as "emotional." We use conversational language and contractions like "don't" because there are studies that (purport to) show that people remember conversations better than other types of discourse. This seems to reinforce that kind of thinking.
On a more personal level, I'll never forget the time I was locked in a stairwell I had taken to get to another floor at work. For whatever reason, I exclaimed "Doggone it!" This is something my grandparents might have said, and isn't part of my daily vocabulary. One of our security guards—for whom "Doggone it" is definitely a phrase of choice—happened to be in the hallway on the other side of the door and let me back in. He was nice to me for a long time after that!
Michele said,
August 9, 2018 @ 10:11 am
@myl: "Thus in tests of foreign-language fluency, speaking rate in syllables per second correlates very well with evaluators' ratings. But the speaking-rate metric would diverge strongly from human evaluation for someone who just gabbles nonsense very quickly."
Makes me wonder what they'd think of those fast-talking auctioneers!
Trogluddite said,
August 10, 2018 @ 2:10 pm
The analysis of the subjects' language doesn't worry me so much as the language used by the researchers in their paper. Terms like "support effectiveness" and "lasting emotional recovery" seem too far removed from what is really being measured by the linguistic analysis and posting behaviour, and strike me as hyperbolic and potentially misleading.
I use such online community-based support sites often; I am autistic, and the researchers might say that I often find greater "synchrony" with on-line autistic peers than with "IRL" non-autistic peers, and I would agree. Such places *have* helped me immensely; however, I also have seen many members for whom "quick fixes" of sympathy/validation have apparently become the end rather than the means; maybe even an addiction. Such members often upvote only those responses which confirm their existing biases while never showing any improvement in their handling of problems regularly restated in new threads, and sometimes for years.
It is also a truism that it easier to tell someone what they *want* to hear than what they *need* to hear. So, we also can't be sure that responders aren't using this principle in order to obtain an emotional uplift of their own from their upvotes. This is also something that I see often on social support websites; some members will post a "me too" response to problems for which they have no useful advice, which may be an *agreeable* response but not necessarily a *helpful* one.
There has been much debate about the use of crowd-sourcing and AI for mental health care, but there has been very little clinical testing of these services. As with any kind of health care, not everyone benefits from the same treatments, and some people may be vulnerable to negative side-effects. More research in this area is certainly required, and it good to see some being done; but it must be, and must be seen to be, disinterested and objective if we are to determine what the benefits might be and exactly who might benefit.
I think that the research makes a good case for the hypothesis that "synchronic" text communication helps to build rapport and that such rapport is emotionally uplifting; but most people would read far more than this into terms such as "successful emotional support" and "lasting emotional recovery". Language Log has many times exposed the ways in which decontextualised references to science papers can be used to support misleading stories; I do hope that the easily misinterpreted language in this paper does not lead to any new examples from those with a vested interest in uncritically promoting crowd-sourcing and AI for mental healthcare.