Read on Twitter
What does it mean to make causal inferences about conversational behaviours? Why does that matter, why's it difficult, what would it take to get there? My #CSCW2020 paper w/ @m_sendhil & @Cristian_DNM takes a step towards addressing these questions https://bit.ly/2FbwwOv [1/]
Here's a boat I've often found myself in: Say you analyze lots of conversations, finding that in conversations that go better, people tend to use more positive language. If you wanted to foster better conversations, you might be tempted to encourage more positive language use.
There's a crucial and *causal* question in the gap between observation and data-driven policy: does being upbeat actually cause better conversations? Our paper formalizes this type of question, and argues why addressing it is important and hard.
The question is important because in such high-stakes settings as crisis counseling, correlation!=causation isn't just an adage: getting it right impacts the people in the conversations. (see my #ACL2020 paper w/ @CrisisTextLine https://bit.ly/2XcIKMU )
The question is hard because your behaviours in a conversation, and the outcome of the conversation, are tied up with the circumstances in which the conversation occurs, and the actions of the other person in the interaction.
To be concrete: maybe a counselor being more positive actually helps individuals in crisis; maybe it's simply easier to be positive to someone who's easier to help. Would being positive even make sense-let alone be beneficial-in more difficult situations?
We highlight & mathematically articulate two key difficulties stemming from this entanglement. In short: 1) making causal inferences from observational data is hard in general, but 2) especially when you're dealing with conversations: complex, back-and-forth joint activities.
Being precise about these challenges is generative: it points us to certain methodological steps that, for certain conditions and assumptions, actually enable us to provide causal interpretations for certain types of causal claims.
We make our case in a particular type of setting and policy, grounded in one potential application: assigning crisis counselors to future conversations, given observations of their past behaviours.
In the paper, we give empirical hints about what we might gain from making more careful causal inferences. At the same time, applying our framework to real data brings to the fore other questions that complement the causal one e.g., what are the behaviours & outcomes that matter?
There's many more causal questions about conversations, that map to impactful real-world applications, that inherit the basic challenges we described in our paper, and that lead us further into the thicket of what makes conversations challenging and interesting. [/n]
ps: personally, it's been fun to try and draw causal graphs of how conversations work. It's made me stare more closely at all the assumptions I'd had about conversations, and has pointed me to questions I'd like to keep asking.
pps: sociologically, the diversity of venues this paper has been rejected from (data science, NLP, HCI) is kind of funny. it probably says something snippy about doing interdisciplinary work, and the importance of audience design :)
