There is a lot I like about this book. It's well-written, and includes a lot on the scientific study of groups (i.e. it's not a business book containing some CEO's gut feelings based on a small number of anecdotes). But I want to point out how it gets language wrong. 1 /

This is mostly based on the first chapter, where the author talks about MIT's Human Dynamics Lab, and work on Social Signal Processing (SSP) by Sandy Pentland and others. 2 /

SSP, if you're not familiar with the field, focuses on nonverbal aspects of social interaction, e.g. gaze, gesture, prosody. Pentland is a major figure in the field (and essentially the founder, in that he coined the phrase Social Signal Processing). 3 /

SSP is an exciting field containing a lot of great, worthwhile research. But, the chapter goes too far by concluding that language is not important for understanding groups. "Normally, we think that words matter...But that assumption is wrong. Words are noise." 4 /

The context for this claim is work (at MIT and elsewhere) on automatic prediction of group performance on a task. And, indeed, nonverbal cues can be extremely useful for this type of prediction task. 5 /

I doubt many SSP researchers would explicitly say that language is not important for understanding social interaction (at least in their writings), but it may be a widespread attitude in the field. 6 /

However, there is *very* little research in SSP comparing verbal and nonverbal approaches for a particular task. 7 /

And I'm not criticizing SSP researchers for focusing only on nonverbal aspects of interaction, just like I wouldn't criticize a computational linguist for not studying gaze. 8 /

But we can't take SSP's focus on nonverbal features as a reason to conclude that language is not important for modelling / understanding / predicting group behaviour. 9 /

More fundamentally, there is no single task of predicting group performance. Performance has many aspects. Most research in this domain uses scenarios like a winter survival task, or an assembly task, or navigating a map or virtual environment. 10 /

So is language important for predicting group performance? We can't say yes or no, as these tasks and performance metrics are very different. In some cases (e.g. winter survival) we have evidence that combining verbal + nonverbal is best. 11 /

And the same goes for related tasks like detecting group affect. 12 /

If the goal is understanding / modelling / predicting group behaviour, this is hugely multi-faceted, and sometimes language will be important. Not just in the sense of ling. feas that improve performance, but also for providing context and interpretability for predictions. 13 /

And sometimes nonverbal alone is enough. After all, the fact that we are less conscious of many of these nonverbal aspects of interaction makes them very revealing. 14 /

I'll boil this thread down to two points:

1. SSP research doesn't tell us language is unimportant for understanding groups.
2. We need more #nlproc / computational linguistics work on group interaction.

15 /

I think the reason there is not more #nlproc work on group interaction is because of data scarcity. That is slowly changing, with recent corpora like ELEA, MULTISIMO, GAP, etc. 16 /

Back to the book: that strong claim in chapter 1 about words not being important is undermined by later chapters on how analyzing language in emails can predict employee affect and turnover, and the relationship between pilot communication and performance in emergencies, etc. 17/

Who knew language could be so illuminating? 18 /