hey, laurenhat, what's your research about, anyway?
I don't think I've ever posted anything directly summarizing my research here. A lot of my friends don't even know what field I'm in, exactly (answer: cognitive science) -- partly because what I do is a mishmash of psychology, linguistics, and AI, partly because I've been in various departments over the years. So, for those who are interested, here is a description, in way, way too much detail, of my main areas of research, past and present.
As an undergraduate, I majored in Symbolic Systems at Stanford. At the time, that was basically a mix of psychology, linguistics, philosophy, and computer science. I concentrated in artificial intelligence. At the same time, I wrote an honors thesis on some work that I was doing over in the psychology department with Lera Borodtisky (even back then, I was hard to classify). My undergraduate thesis was about the effects of grammatical gender systems on our understanding of inanimate objects. Grammatical gender doesn't exist in English, but in many other languages (e.g., Spanish, German, Russian), all objects are labeled with a "gender" (where there are two are more genders), and words are inflected to agree with these grammatical gender categories, in the same way that words in English have to agree in terms of singular vs. plural and other grammatical distinctions (e.g., every time you refer to the toaster using a pronoun, you have to call it "he" or "she" as appropriate, and depending on your language, you may have to mark adjectives and/or verbs to have the corresponding gender: "he is a shiny-[MALE] toaster").
So my question was, did talking about tomatoes and toasters as if they were masculine or feminine affect the way that speakers of languages with grammatical gender thought about those things over time? If so, how? We started out by finding a list of words that had opposite genders in German and Spanish (e.g., "key" is feminine in Spanish and masculine in German; "bridge" is the opposite). Then we did a silly little experiment: We asked native speakers of Spanish and German (who also spoke English) to memorize word-name pairs, like "apple -- Patrick", "bridge -- Brenda", etc. (The names were all names that were easily confusable with names of the opposite gender -- Patrick/Patricia, Brenda/Brendan, etc.) Then later we asked people to do a recognition task -- "was the apple named Patricia?" It turned out people remembered better when the name was the same gender as the noun was in their native language than when there was a gender mismatch -- even though the whole experiment was done in English.
Encouraged by the wacky result, we wondered what was actually changed for people by learning grammatical genders. Did people, for example, represent different features of the items as being more or less salient based on what gender the item had? To test that idea, we asked different native speakers of German and Spanish to list the first three adjectives they thought of to describe each noun on our list. Then we got independent subjects to rate how masculine or feminine those adjectives were as they pertained to the objects. We saw differences corresponding to grammatical gender: German speakers said keys were things like "metal, jagged, useful, hard" while Spanish speakers generated adjectives more like "little, lovely, shapely, golden" (note: I'm being lazy and typing this from memory, so I'm not sure those are all quite right). Bridges, on the other hand, were "elegant, beautiful, fragile" to German speakers and "long, long, long, hard" to Spanish speakers (again, that's an approximation from memory, but you get the idea).
Even more interesting, rather than just showing a correlation, we showed a causal effect of language on thought. We taught English speakers a made up grammatical gender system -- e.g., we taught them they had to say "oos" or "sou" in front of objects, like "oos pan" or "sou pot", and all of the females in our training set of objects were assigned to one gender and the males to another ("oos ballerina", "sou king"). After learning these made up classifications for a while, English speakers also showed a gender-based effect when they had to generate adjectives to describe the objects they'd learned about.
That was my undergraduate work -- Lera Boroditsky and Webb Phillips went on to do a whole host of really interesting and important follow-up experiments. We're still trying to publish some of them in a journal (we're bogged down in the revision process), but you can find a lot more details in the book chapter we published on the first bunch of studies.
I stayed at Stanford to get a master's degree in computer science (again focusing in AI). While I was doing that, I was also working in the psychology department, this time with Josh Tenenbaum, who would later move to MIT and become my graduate advisor. (Lera also moved to MIT and was briefly my co-advisor there, but then moved back to Stanford.)
Josh's lab takes a computational approach to cognitive science. That means that people in his lab look at a problem from psychology, like, "When people hear a new word, how do they figure out what that word means, given all the possible things it might apply to?" (E.g., "Hund" could be German for "Fido", "poodle", "dog", "animal", "thing", "dog leg", "barking", "furry", or all sorts of things that are either not that concrete or don't even have a good equivalent in English -- "that particular dog, Fido, on this particular date, while barking"). They then try to break down the question -- what kinds of different data might you expect people to have about what the word means? What kinds of prior knowledge might we hypothesize that people might have? (E.g., you might hypothesize that people have no innate knowledge at all about what words could mean; that they believe things should be nouns or verbs but not both -- no words meaning "this dog while barking"; that they have a preference for words to apply to whole objects and not object parts; that they would prefer to apply new words to things that they don't already have names for, or some combination.) And what strategies and representations might people use in solving this problem? (Do they look for categories? Do they assume that there's an underlying taxonomy organizing those categories? Do they prefer to assign new words to smaller rather than larger categories?)
This breakdown -- prior knowledge, problem solving strategy, and data -- allows us to model and play with each of these items separately. Then we can do things like give little kids a certain amount of data about what a new word means ("That is a blicket! This one is also a blicket!"), give our models the same data (or as close as we can manage), and then look at different combinations of prior knowledge plus strategies that give us similar results to what kids are doing. Often, we can get some fairly suggestive results about what seems likely to be happening, but still, our answers are usually of the form, "There is enough data for people to be learning X using some basic statistical principles, without having any assumptions except Y built in at birth". That means we're rarely, if ever, making positive statements like, "This is how people learn/do X". Still, it at the very least narrows the search space of where scientists need to look to figure out what people are actually doing. And it's awfully useful to be able to model the components of the problem this way (and it's all thanks to Bayesian statistics, for those of you know/care about that).
So, while I was a master's student, I was working on the problem of how people learn what a new noun means. Josh had already done work with Fei Xu showing how adults and kids learned what new words like "blicket" mean when they're applied to a set of familiar objects (e.g., how do you know whether "blicket" means "tow truck", "truck", or "vehicle"?), but I wanted to look at what happened when you put people in an "alien" environment where they didn't already have English words for all the objects, and accompanying expectations.
I created a set of awesome alien items (in a 3D modeling program) from the planet Gazoob. There was a taxonomy, with vaguely plant-like things, vaguely shell-like things, and vaguely
tool-like things. We gave people varying amounts of data ("this is a blicket" vs. "these three are blickets"), and we looked at what guesses they made about other items' blicketude. We also looked at how their perceptions of the objects changed after they had learned words for them (e.g., if three items were all labeled as blickets, did that make people see them as more similar afterward?).
Since this is all totally unpublished and not even fully analyzed, I won't go into the results in much detail. But people do seem to apply a couple simple statistical principles to figure out which items are likely to be blickets -- or at least, our models which do so have similar behavior to the people in the experiments -- and learning new words for things does affect how similar people think they are (at least in this particular noun-learning experiment -- some of the grammatical gender follow-up work that I was not directly involved in shows cases where this doesn't happen).
I don't know if this work is going to get published. It's kind of a small offshoot of some stuff that my advisor has done. It might fit in okay with my thesis work, though, and there's a chance I will want/need to include it in my thesis. But I need to do more analysis first. For instance, it looks like maybe if people have already learned that something is a "blicket" and decided that "blicket" means "tow truck" (or the Gazoobian random squiggly item equivalent), then if they see the same thing labeled a "zav" later, they'll be disinclined to think that "zav" also means "tow truck" (maybe it means "vehicle" instead, e.g.). (This would not be surprising given that several important word learning theories basically incorporate such a principle of mutual exclusivity, but I need to see if and how it's being applied here.) So it looks like the randomized order of the word learning trials might possibly have had an effect on what people were learning. Maybe. I need to find this out.
Whew. Okay, That was a ton of information, and I haven't even described the stuff I've done at MIT (though I have worked a little bit on both of these projects since coming to MIT). I think that stuff should be a lot easier to describe, though, now that I've explained what computational cognitive science is about. But that'll have to wait for another entry -- hopefully soon -- because I need to do some more work now, on one of my mysterious current projects. :)
As an undergraduate, I majored in Symbolic Systems at Stanford. At the time, that was basically a mix of psychology, linguistics, philosophy, and computer science. I concentrated in artificial intelligence. At the same time, I wrote an honors thesis on some work that I was doing over in the psychology department with Lera Borodtisky (even back then, I was hard to classify). My undergraduate thesis was about the effects of grammatical gender systems on our understanding of inanimate objects. Grammatical gender doesn't exist in English, but in many other languages (e.g., Spanish, German, Russian), all objects are labeled with a "gender" (where there are two are more genders), and words are inflected to agree with these grammatical gender categories, in the same way that words in English have to agree in terms of singular vs. plural and other grammatical distinctions (e.g., every time you refer to the toaster using a pronoun, you have to call it "he" or "she" as appropriate, and depending on your language, you may have to mark adjectives and/or verbs to have the corresponding gender: "he is a shiny-[MALE] toaster").
So my question was, did talking about tomatoes and toasters as if they were masculine or feminine affect the way that speakers of languages with grammatical gender thought about those things over time? If so, how? We started out by finding a list of words that had opposite genders in German and Spanish (e.g., "key" is feminine in Spanish and masculine in German; "bridge" is the opposite). Then we did a silly little experiment: We asked native speakers of Spanish and German (who also spoke English) to memorize word-name pairs, like "apple -- Patrick", "bridge -- Brenda", etc. (The names were all names that were easily confusable with names of the opposite gender -- Patrick/Patricia, Brenda/Brendan, etc.) Then later we asked people to do a recognition task -- "was the apple named Patricia?" It turned out people remembered better when the name was the same gender as the noun was in their native language than when there was a gender mismatch -- even though the whole experiment was done in English.
Encouraged by the wacky result, we wondered what was actually changed for people by learning grammatical genders. Did people, for example, represent different features of the items as being more or less salient based on what gender the item had? To test that idea, we asked different native speakers of German and Spanish to list the first three adjectives they thought of to describe each noun on our list. Then we got independent subjects to rate how masculine or feminine those adjectives were as they pertained to the objects. We saw differences corresponding to grammatical gender: German speakers said keys were things like "metal, jagged, useful, hard" while Spanish speakers generated adjectives more like "little, lovely, shapely, golden" (note: I'm being lazy and typing this from memory, so I'm not sure those are all quite right). Bridges, on the other hand, were "elegant, beautiful, fragile" to German speakers and "long, long, long, hard" to Spanish speakers (again, that's an approximation from memory, but you get the idea).
Even more interesting, rather than just showing a correlation, we showed a causal effect of language on thought. We taught English speakers a made up grammatical gender system -- e.g., we taught them they had to say "oos" or "sou" in front of objects, like "oos pan" or "sou pot", and all of the females in our training set of objects were assigned to one gender and the males to another ("oos ballerina", "sou king"). After learning these made up classifications for a while, English speakers also showed a gender-based effect when they had to generate adjectives to describe the objects they'd learned about.
That was my undergraduate work -- Lera Boroditsky and Webb Phillips went on to do a whole host of really interesting and important follow-up experiments. We're still trying to publish some of them in a journal (we're bogged down in the revision process), but you can find a lot more details in the book chapter we published on the first bunch of studies.
I stayed at Stanford to get a master's degree in computer science (again focusing in AI). While I was doing that, I was also working in the psychology department, this time with Josh Tenenbaum, who would later move to MIT and become my graduate advisor. (Lera also moved to MIT and was briefly my co-advisor there, but then moved back to Stanford.)
Josh's lab takes a computational approach to cognitive science. That means that people in his lab look at a problem from psychology, like, "When people hear a new word, how do they figure out what that word means, given all the possible things it might apply to?" (E.g., "Hund" could be German for "Fido", "poodle", "dog", "animal", "thing", "dog leg", "barking", "furry", or all sorts of things that are either not that concrete or don't even have a good equivalent in English -- "that particular dog, Fido, on this particular date, while barking"). They then try to break down the question -- what kinds of different data might you expect people to have about what the word means? What kinds of prior knowledge might we hypothesize that people might have? (E.g., you might hypothesize that people have no innate knowledge at all about what words could mean; that they believe things should be nouns or verbs but not both -- no words meaning "this dog while barking"; that they have a preference for words to apply to whole objects and not object parts; that they would prefer to apply new words to things that they don't already have names for, or some combination.) And what strategies and representations might people use in solving this problem? (Do they look for categories? Do they assume that there's an underlying taxonomy organizing those categories? Do they prefer to assign new words to smaller rather than larger categories?)
This breakdown -- prior knowledge, problem solving strategy, and data -- allows us to model and play with each of these items separately. Then we can do things like give little kids a certain amount of data about what a new word means ("That is a blicket! This one is also a blicket!"), give our models the same data (or as close as we can manage), and then look at different combinations of prior knowledge plus strategies that give us similar results to what kids are doing. Often, we can get some fairly suggestive results about what seems likely to be happening, but still, our answers are usually of the form, "There is enough data for people to be learning X using some basic statistical principles, without having any assumptions except Y built in at birth". That means we're rarely, if ever, making positive statements like, "This is how people learn/do X". Still, it at the very least narrows the search space of where scientists need to look to figure out what people are actually doing. And it's awfully useful to be able to model the components of the problem this way (and it's all thanks to Bayesian statistics, for those of you know/care about that).
So, while I was a master's student, I was working on the problem of how people learn what a new noun means. Josh had already done work with Fei Xu showing how adults and kids learned what new words like "blicket" mean when they're applied to a set of familiar objects (e.g., how do you know whether "blicket" means "tow truck", "truck", or "vehicle"?), but I wanted to look at what happened when you put people in an "alien" environment where they didn't already have English words for all the objects, and accompanying expectations.
I created a set of awesome alien items (in a 3D modeling program) from the planet Gazoob. There was a taxonomy, with vaguely plant-like things, vaguely shell-like things, and vaguely
tool-like things. We gave people varying amounts of data ("this is a blicket" vs. "these three are blickets"), and we looked at what guesses they made about other items' blicketude. We also looked at how their perceptions of the objects changed after they had learned words for them (e.g., if three items were all labeled as blickets, did that make people see them as more similar afterward?).
Since this is all totally unpublished and not even fully analyzed, I won't go into the results in much detail. But people do seem to apply a couple simple statistical principles to figure out which items are likely to be blickets -- or at least, our models which do so have similar behavior to the people in the experiments -- and learning new words for things does affect how similar people think they are (at least in this particular noun-learning experiment -- some of the grammatical gender follow-up work that I was not directly involved in shows cases where this doesn't happen).
I don't know if this work is going to get published. It's kind of a small offshoot of some stuff that my advisor has done. It might fit in okay with my thesis work, though, and there's a chance I will want/need to include it in my thesis. But I need to do more analysis first. For instance, it looks like maybe if people have already learned that something is a "blicket" and decided that "blicket" means "tow truck" (or the Gazoobian random squiggly item equivalent), then if they see the same thing labeled a "zav" later, they'll be disinclined to think that "zav" also means "tow truck" (maybe it means "vehicle" instead, e.g.). (This would not be surprising given that several important word learning theories basically incorporate such a principle of mutual exclusivity, but I need to see if and how it's being applied here.) So it looks like the randomized order of the word learning trials might possibly have had an effect on what people were learning. Maybe. I need to find this out.
Whew. Okay, That was a ton of information, and I haven't even described the stuff I've done at MIT (though I have worked a little bit on both of these projects since coming to MIT). I think that stuff should be a lot easier to describe, though, now that I've explained what computational cognitive science is about. But that'll have to wait for another entry -- hopefully soon -- because I need to do some more work now, on one of my mysterious current projects. :)