With Austin’s reputation as a world class technology hub, I recall feeling a bit of imposter syndrome upon arrival for my 8 week visit to The Computational Linguistics Lab at The University of Austin at Texas! Would shots be fired at this cowboy psycholinguist riding in with his relatively basic knowledge of computational modelling? Soon enough I reminded myself that the core of this trip was to bring along a holster of data that we have collected about how "humans" understand pronouns, and second to see what the computational linguist experts advise me about how modelling can inform our theory. Part 1 of my blog will outline what this human data is; Part 2 of my blog will cover the Texas visit itself and the more computational-y side of things!
What we’re studying.
A pronoun is a word that is used instead of a full name. How do we learn, for example, that in “Clint Eastwood said that John Wayne poked himself”, we interpret that the pronoun "himself" cannot refer to "Clint”, but instead refers only to "John”. Much like the old Texan-based western movies portray an experienced ranch hand wrangling up their livestock with seamless ease and expertise, adults and children appear to effortlessly understand the meaning of pronouns - despite that this task in no mean feat (see my previous blog).
Theoretical duels.
A traditional account (Chomsky, 1981) is that knowledge is driven by innate grammatical constraints, for example that reflexive pronouns ("himself") can only co-refer with the grammatical subject of its clause (i.e., John Wayne); whereas non-reflexive pronouns ("him") are only free to co-refer elsewhere (i.e., Clint Eastwood). At LuCiD, we’ve been asking whether the way we interpret a pronoun is inferred through pragmatic and meaning based properties of language, such as whichever character is the main topic of the conversation. For example, the traditional argument for grammatical constraints can actually be made redundant by the fact that reflexive pronouns are used to mark unusual scenarios where the agent “do-er” of the action (i.e., John Wayne) is also the direct recipient or target of the actions represented by the sentences. Also note that a similar formulation is that reflexive pronouns denote a referent as seen from his or her own point of view, non-reflexive pronouns from a more objective viewpoint/view from nowhere.
Our human data is tastier than a Texas BBQ.
In the earlier “poked” example with the grammatical use of a reflexive, our explanation correlates with the fact that the intended co-referent to the reflexive “John” is still the local grammatical subject. One way we disentangled these explanations was in a sentence completion task where adults were presented videos of the mentioned character being covered in a substance like mud, and we manipulated whether the verb denoted that this directly affected character was definitely performing the action (as indicated by a main lexical verb: “John SPLASHED mud all over _____”), or was not necessarily performing the action: “John HAS mud all over _____”). Adults chose to use “himself” nearly 100% after verbs like “splash” (agentive) but equivalently chose “himself” or “him” when the verb was “has” (non-agentive).
In further adult experiments we’ve investigated how adults interpret sentences already containing either a reflexive or non-reflexive pronoun, and by which it is grammatical for the respective pronoun to co-refer to two mentioned characters – e.g., either the subject "John" or object "Clint" in: “John told Clint about the picture of himself/him”. Adults rated on a scale of 0 to 100 the extent to which they thought the respective pronoun form referred to a specific character. Preferences were driven by topicality and event likelihood – that is, who it most plausibly refers to. Consider as a baseline “John told Clint about the picture of himself/him”: adults chose “John” (the subject) more often for either pronoun because, plausibly, the teller is likely to possess and therefore tell more information about themselves than the other person; whereas they chose “Clint” (the object) more often if the event was changed to “asking” because, plausibly, the asker is likely to already know information about themselves but not of the other person! To put the boot in, this object preference was even more pronounced if the characters had stereotypical power relations, for example in “The sheriff interrogated the suspect about the picture of himself/him”.
After that update, find out about my actual trip to Austin in part 2 of the blog.