One thing I really liked about this paper was the use of a formal model of a theory (in this case, a theory of pragmatic implicature cancellation) to predict specific usages that may not have originally been considered by previous researchers working on the theory. In this case, it's a certain combination of theory of mind and implicature cancellation which leads to specific predictions when the listener knows the speaker has partial vs. complete knowledge.
One thing I was curious about was G&S's use of the term of "partial implicature". They were specifically using it when, for example, the speaker had knowledge of two of the three apples and said "one apple is red". The listeners inferred that either one or two of the three apples apples were red, but not all three. If I'm understanding G&S correctly, the partial implicature is that the listeners didn't infer all three apples were red, but did allow for more than one apple to be red. To me, this is the listener thinking, "Ah, the speaker used one instead of two, which means that one of the apples the speaker saw definitely isn't red. But the third one the speaker didn't see might be red, so either one or two apples total are red." If this is true, then it seems to me like the "partial implicature" is really a regular implicature (i.e., one = one or two) over the restricted domain of two apples (the one the speaker saw that was definitely red and the one the speaker didn't see yet). If this is true, I'm not sure this result says something particularly different than the rest of the results (i.e., partial implicature, under this interpretation, doesn't seem different from regular implicature). More specifically, there are cases where the implicature is cancelled, and cases where it isn't, and this happens to be one where it isn't. The magic comes in on the domain restriction the listener imposes, rather than being anything about the implicature itself.
Some other thoughts:
(1) Intro, p.174: I was glad G&S were explicit about who argues for a purely modularized form of implicature, because when I first read that option, my thought was, "Really? That sounds a bit like a straw man." To me, pragmatics is definitely the aspect of language knowledge that seems most amenable to incorporating non-linguistic knowledge because it's about how we use language to communicate. From what I took of G&S's summary of the strongly modular theories, I would assume that those theories haven't yet looked at incorporating this issue of incomplete knowledge on the part of the speaker?
(2) Experiment 1, p.178: I was trying to work this through for the "one" case, with "some" being used. If Laura looks at one of three letters, and then says "Some of the letters have checks inside", my intuition about what this means becomes a little wonky. For me, "some" means more than one, but of course, Laura can't know about more than one of the letters. So how do I interpret what she's saying here? My first inclination is to think she's got some kind of magic knowledge about the other letters and so knows that at least one of the other two has a check, and I wonder if some of the respondents indicated this when G&S checked for how much knowledge the listeners believed Laura to have. For the ones left over who believed Laura only knew about one, we can see in Figure 2 that they basically wibbled between interpreting some as two or three.
(3) Figure 2 results for exact number words, p.182: When we look at the "one" row (bottom of (B) and (D)), it seems like the model prefers "one" to mean two when only one object is known (1st panel of B). However, people prefer "one" to mean three (1st panel of D). Notably, if we move up one row to the "two" row (middle of (B) and (D)), it seems like here the model happily prefers "two" to mean three (1st panel of B), as do people (1st panel of D). I'm not sure how to interpret this exactly -- it certainly seems like these are different preferences qualitatively at the very least, as one matches human preferences (interpreting "two") and the other doesn't (interpreting "one"). I'm also not sure why the model mechanics would yield different predictions in these two cases. Is it something about the number word meaning "one more" somehow? If so, that would explain why the model prefers "one" to mean two, and prefers "two" to mean three. However, it's not immediately obvious to me why this would fall out of the model mechanics.
(4) Conclusion, p.183: As a follow up from our discussion last time, it seems like there's some evidence from the pragmatic inference modeling literature that there's a boundary of two on the recursion depth for theory of mind: "...we assume only one such level of reasoning...The quantitative fits we have shown suggest that limited recursion and optimization are psychologically realistic assumptions." This provides another bit of evidence that the center-embedding limit for structural recursion probably isn't specific to syntax. (Though then it becomes interesting why other kinds of recursion in syntax don't seem to have this same limit, e.g., right-branching: "This is the dog who chased the cat who ate the rat who stole the cheese." We clearly can process these three embeddings without a problem.)
Wednesday, March 12, 2014
Friday, February 28, 2014
Next time on 3/14/14 @ 3:20pm in SBSG 2221 = Goodman & Stuhlmüller 2013
Thanks to everyone who was able to join us for our exciting discussion of Levinson 2013 & the Legate et al. 2013 reply manuscript! We had some particularly interesting thoughts about potential recursion in non-syntactic domains, such as theory of mind, which relates to our article for Friday March 14 at 3:20pm in SBSG 2221: Goodman & Stuhlmüller 2013. In particular, G&S2013 discuss a rational model of how speakers interpret utterances, where listeners model the speaker's model of selecting utterances:
Goodman, N. & Stuhlmueller, A. 2013. Knowledge and Implicature: Modeling Language Understanding as Social Cognition.Topics in Cognitive Science, 5, 173-184.
See you then!
Wednesday, February 26, 2014
Some thoughts on Levinson 2013 + Legate, Pesetsky, & Yang 2013 reply (manuscript)
One of the take-away points I had from Levinson 2013 [L13] was the idea that center-embedding is not a structural option specific to syntax, since there are examples of this same structural option in dialogue. I had the impression then that L13 wanted to use this to mean this particular type of recursion is not language-specific, as dialogue is using language to communicate information and it's the information communicated (via the speech acts) that's center-embedded. (At least, that's how I'm interpreting "speech acts" as "actions in linguistic clothing".) I'm not quite sure I believe that, since I would classify speech acts as a type of linguistic knowledge (specifically, how to translate intention into the specific linguistic form required to convey that intention). But suppose we classify this kind of knowledge as not really linguistic, per se -- then wouldn't the interesting question be about how unique this type of structural option is to human communication systems, since that relates to questions about the Faculty of Language (broad or narrow)? And presumably, this then links back to whether non-human animals can learn these syntactic structures (doesn't seem to be true as far as we know) or these type of embedded interactions (also doesn't seem to be true, I think?)?
As a general caveat, I should note that while I followed the simpler examples of center embedding in dialogue, I was much less clear about the more complex examples that involved multiple center-embeddings and cross-serial dependencies (for example, deciding that something was an embedded question rather than a serial follow-up, like in example (14), the middle of (16), some of the embeddings in (17)). This may be due to my very light background in pragmatics and dialogue analysis, however. Still, it seemed that Legate, Pesetsky, & Yang 2013 [LPY13] had similar reservations about some of these dialogue dependencies.
LPY13 also had very strong reactions to both the syntactic and computational claims in L13, in addition to these issues about how to assign structure to discourse. I was quite sympathetic to (and convinced by) LPY13's syntactic and computational objections as a whole, from the cross-linguistic frequency of embedding to the non-centrality of center embedding for recursion to the not-debate about whether natural languages were regular. They also brought out a very interesting point about the restrictions on center embedding in speech acts (example (13)), which seemed to match some of the restrictions observed in syntax. If it's true that these restrictions are there, and we see them in both linguistic (syntax) and potentially non-linguistic (speech act) areas, then maybe this is nice evidence for a domain-general restriction on processing this kind of structure. (And so maybe we should be looking for it seriously elsewhere too.)
More specific comments:
L13: There's a comment in section 4 about whether it's more complex to treat English as a large system of simple rules or a small system of complex rules. Isn't this exactly the kind of thing that rational inference gets at (e.g., Perfors, Tenenbaum, & Regier 2011 find that the context-free grammar works better than a regular or linear grammar on child-directed English speech -- as LPY13 note)? With respect to recursion, L13 cites the Perfors et al. 2010 study, which LPY13 correctly note doesn't have to do with regular languages vs. non-regular languages. Instead, that study finds that a mixture of recursive and non-recursive context-free rules (surprisingly) is the best, rather than having all recursive or all non-recursive rules, despite this seeming to duplicate a number of rules.
L13: Section 6, using the transformation from pidgin to creole as evidence for syntactic embedding coming from other capacities like joint action abilities: It's true that one of the hallmarks of pidgins vs. creoles is the added syntactic complexity, which (broadly speaking) seems to come from children learning the pidgin, adding syntactic structure to it that's regular and predictable, and ending up with something that has the same syntactic complexity as any other language. I'm not sure I understand why this tells us anything about where the syntactic complexity is coming from, other than something internal to the children (since they obviously aren't getting it from the pidgin in any direct way). Is it that these children are talking to each other, and it's the dialogue that provides a model for the embedded structures, for example?
LPY13: I'm not quite sure I agree with the objection LPY13 raise about whether dialogue embeddings represent structures (p.9). I agree that there don't seem to be very many restrictions, certainly when compared to syntactic structure. But just because there are multiple licit options doesn't mean there isn't a structure corresponding to each of them. It just may be exactly this: there are multiple possible structures that we allow in dialogue. So maybe this really more of an issue about how we tell structure is present (as opposed to linear "beads on a string", for example).
As a general caveat, I should note that while I followed the simpler examples of center embedding in dialogue, I was much less clear about the more complex examples that involved multiple center-embeddings and cross-serial dependencies (for example, deciding that something was an embedded question rather than a serial follow-up, like in example (14), the middle of (16), some of the embeddings in (17)). This may be due to my very light background in pragmatics and dialogue analysis, however. Still, it seemed that Legate, Pesetsky, & Yang 2013 [LPY13] had similar reservations about some of these dialogue dependencies.
LPY13 also had very strong reactions to both the syntactic and computational claims in L13, in addition to these issues about how to assign structure to discourse. I was quite sympathetic to (and convinced by) LPY13's syntactic and computational objections as a whole, from the cross-linguistic frequency of embedding to the non-centrality of center embedding for recursion to the not-debate about whether natural languages were regular. They also brought out a very interesting point about the restrictions on center embedding in speech acts (example (13)), which seemed to match some of the restrictions observed in syntax. If it's true that these restrictions are there, and we see them in both linguistic (syntax) and potentially non-linguistic (speech act) areas, then maybe this is nice evidence for a domain-general restriction on processing this kind of structure. (And so maybe we should be looking for it seriously elsewhere too.)
More specific comments:
L13: There's a comment in section 4 about whether it's more complex to treat English as a large system of simple rules or a small system of complex rules. Isn't this exactly the kind of thing that rational inference gets at (e.g., Perfors, Tenenbaum, & Regier 2011 find that the context-free grammar works better than a regular or linear grammar on child-directed English speech -- as LPY13 note)? With respect to recursion, L13 cites the Perfors et al. 2010 study, which LPY13 correctly note doesn't have to do with regular languages vs. non-regular languages. Instead, that study finds that a mixture of recursive and non-recursive context-free rules (surprisingly) is the best, rather than having all recursive or all non-recursive rules, despite this seeming to duplicate a number of rules.
L13: Section 6, using the transformation from pidgin to creole as evidence for syntactic embedding coming from other capacities like joint action abilities: It's true that one of the hallmarks of pidgins vs. creoles is the added syntactic complexity, which (broadly speaking) seems to come from children learning the pidgin, adding syntactic structure to it that's regular and predictable, and ending up with something that has the same syntactic complexity as any other language. I'm not sure I understand why this tells us anything about where the syntactic complexity is coming from, other than something internal to the children (since they obviously aren't getting it from the pidgin in any direct way). Is it that these children are talking to each other, and it's the dialogue that provides a model for the embedded structures, for example?
LPY13: I'm not quite sure I agree with the objection LPY13 raise about whether dialogue embeddings represent structures (p.9). I agree that there don't seem to be very many restrictions, certainly when compared to syntactic structure. But just because there are multiple licit options doesn't mean there isn't a structure corresponding to each of them. It just may be exactly this: there are multiple possible structures that we allow in dialogue. So maybe this really more of an issue about how we tell structure is present (as opposed to linear "beads on a string", for example).
Friday, February 14, 2014
Next time on 2/28/14 @ 3:20pm in SBSG 2221 = Levinson 2013 + Legate et al. 2013 Manuscript
Thanks to everyone who was able to join us for our delightfully thoughtful discussion of the Omaki & Lidz 2013 manuscript! Next time on Friday February 28 at 3pm in SBSG 2221, we'll be looking at an article that argues that recursion is a central part of cognition, even if it's curiously restricted in the realm of syntax (Levinson 2013).
Levinson, S. 2013. Recursion in pragmatics. Language, 89(1), 149-162.
In addition, we'll read a reply to that article that critically examines the assumptions underlying that argument (Legate et al. 2013 Manuscript).
Legate, J., Pesetsky, D., & Yang, C. 2013. Recursive Misrepresentations: a Reply to Levinson (2013). Revised version to appear in Language. Please do not cite without permission from Julie Legate.
Wednesday, February 12, 2014
Some thoughts on the Omaki & Lidz 2013 Manuscript
There are many things that made me happy about this manuscript as a modeler, not the least of which is the callout to modelers about what ought to be included in their models of language acquisition (hurrah for experimentally-motivated guidance!). For example, there's good reason to believe that a "noise parameter" that simply distorts the input in some way can be replaced by a more targeted perceptual intake noise parameter that distorts the input in particular ways. Also, I love how explicit O&L are about the observed vs. latent variables in their view of the acquisition process -- it makes me want to draw plate diagrams. And of course, I'm a huge fan of the distinction between input and intake.
Another thing that struck me was the effects incrementality could have. For example, it could cause prioritization of working memory constraints over repair costs, especially when repair is costly, because the data's coming at you now and you have to do something about it. This is discussed in light of the parser and syntax, but I'm wondering how it translates to other types of linguistic knowledge (and perhaps more basic things like word segmentation, lexical acquisition, and grammatical categorization). If this is about working memory constraints, we might expect it to apply whenever the child's "processor" (however that's instantiated for each of these tasks) gets overloaded. So, at the beginning of word segmentation, it's all about making your first guess and sticking to it (perhaps leading to snowball effects of mis-segmentation, as you use your mis-segmentations to segment other words). But maybe later, when you have more of a handle of word segmentation, it's easier to revise bad guesses (which is one way to recover from mis-segmentations, aside from continuing experience).
This relates to the cost of revision in areas besides syntax. In some sense, you might expect that cost is very much tied to how hard it is to construct the representation in the first place. For syntax (and the related sentential semantics), that can continue to be hard for a really long time, because these structures are so complex. And as you get better at it, it gets faster, so revision gets less costly. But looking at word segmentation, is constructing the "representation" ever that hard? (I'm trying to think what the "representation" would be, other than the identification of the lexical item, which seems pretty basic assuming you've abstracted to the phonemic level.) If not, then maybe word segmentation revision might be less costly, and so the swing from being revision-averse to revision-friendly might happen sooner for this task than in other tasks.
Some more targeted thoughts:
(i) One thing about the lovely schematic in Figure 1: I can definitely get behind the perceptual intake feeding the language acquisition device (LAD) and (eventually) feeding the action encoding, but I'm wondering why it's squished together with "linguistic representations". I would have imagined that perceptual intake directly feeds the LAD, and the LAD feeds the linguistic representation (which then feeds the action encoding). Is the idea that there's a transparent mapping between perceptual intake and linguistic representations, so separating them is unnecessary? And if so, where's the place for acquisitional intake (talked about in footnote 1 on p.7), which seems like it might come between perceptual intake and LAD?
(ii) I found it a bit funny that footnote 2 refers to the learning problem as "inference-under-uncertainty" rather than the more familiar "poverty of the stimulus" (PoS). Maybe PoS has too many other associations with it, and O&L just wanted to sidestep any misconceptions arising from the term? (In which case, probably a shrewd move.)
(iii) In trying to understand the relationship between vocabulary size and knowledge of pronoun interpretation (principle C), O&L note that children who had faster lexical access were not faster at computing principle C, so it's not simply that children who could access meaning faster were then able to do the overall computation faster. This means that the hypothesis that "more vocabulary" equals "better at dealing with word meaning", which equals "better at doing computations that require word meaning as input" can't be what's going on. So do we have any idea what the link between vocabulary size and principle C computation actually is? Is vocabulary size the result of some kind of knowledge or ability that would happen after initial lexical access, and so would be useful for computing principle C too? One thought that occurred to me was that someone who's good at extracting sentential level meaning (i.e., because their computations over words happen faster) might find it easier to learn new words in the first place. This then could lead to a larger vocabulary size. So, this underlying ability to compute meaning over utterances (including using principle C) could cause a larger vocabulary, rather than knowing lots of words causing faster computation.
(iv) I totally love the U-shaped development of filler-gap knowledge in the Gagliardi et al. (submitted) study. It's nice to see an example of this qualitative behavior in a realm besides morphology. The explanation seems similar, too -- a more sophisticated view of the input causes errors, which take some time to recover from. But the initial simplified view leads to surface behavior that seems right, even if the underlying representation isn't at that point. Hence, U-shaped performance curve. Same for the White et al. 2011 study -- U-shaped learning in syntactic bootstrapping for the win.
(v) I really liked the note on p.45 in the conclusion about how the input vs. intake distinction could really matter for L2 acquisition. It's nice to see some explicit ideas about what the skew that occurs is and why it might be occurring. (Basically, this feels like a more explicit form of the "less is more" hypothesis, where adult processing is distorting the input in predictable ways.)
Another thing that struck me was the effects incrementality could have. For example, it could cause prioritization of working memory constraints over repair costs, especially when repair is costly, because the data's coming at you now and you have to do something about it. This is discussed in light of the parser and syntax, but I'm wondering how it translates to other types of linguistic knowledge (and perhaps more basic things like word segmentation, lexical acquisition, and grammatical categorization). If this is about working memory constraints, we might expect it to apply whenever the child's "processor" (however that's instantiated for each of these tasks) gets overloaded. So, at the beginning of word segmentation, it's all about making your first guess and sticking to it (perhaps leading to snowball effects of mis-segmentation, as you use your mis-segmentations to segment other words). But maybe later, when you have more of a handle of word segmentation, it's easier to revise bad guesses (which is one way to recover from mis-segmentations, aside from continuing experience).
This relates to the cost of revision in areas besides syntax. In some sense, you might expect that cost is very much tied to how hard it is to construct the representation in the first place. For syntax (and the related sentential semantics), that can continue to be hard for a really long time, because these structures are so complex. And as you get better at it, it gets faster, so revision gets less costly. But looking at word segmentation, is constructing the "representation" ever that hard? (I'm trying to think what the "representation" would be, other than the identification of the lexical item, which seems pretty basic assuming you've abstracted to the phonemic level.) If not, then maybe word segmentation revision might be less costly, and so the swing from being revision-averse to revision-friendly might happen sooner for this task than in other tasks.
Some more targeted thoughts:
(i) One thing about the lovely schematic in Figure 1: I can definitely get behind the perceptual intake feeding the language acquisition device (LAD) and (eventually) feeding the action encoding, but I'm wondering why it's squished together with "linguistic representations". I would have imagined that perceptual intake directly feeds the LAD, and the LAD feeds the linguistic representation (which then feeds the action encoding). Is the idea that there's a transparent mapping between perceptual intake and linguistic representations, so separating them is unnecessary? And if so, where's the place for acquisitional intake (talked about in footnote 1 on p.7), which seems like it might come between perceptual intake and LAD?
(ii) I found it a bit funny that footnote 2 refers to the learning problem as "inference-under-uncertainty" rather than the more familiar "poverty of the stimulus" (PoS). Maybe PoS has too many other associations with it, and O&L just wanted to sidestep any misconceptions arising from the term? (In which case, probably a shrewd move.)
(iii) In trying to understand the relationship between vocabulary size and knowledge of pronoun interpretation (principle C), O&L note that children who had faster lexical access were not faster at computing principle C, so it's not simply that children who could access meaning faster were then able to do the overall computation faster. This means that the hypothesis that "more vocabulary" equals "better at dealing with word meaning", which equals "better at doing computations that require word meaning as input" can't be what's going on. So do we have any idea what the link between vocabulary size and principle C computation actually is? Is vocabulary size the result of some kind of knowledge or ability that would happen after initial lexical access, and so would be useful for computing principle C too? One thought that occurred to me was that someone who's good at extracting sentential level meaning (i.e., because their computations over words happen faster) might find it easier to learn new words in the first place. This then could lead to a larger vocabulary size. So, this underlying ability to compute meaning over utterances (including using principle C) could cause a larger vocabulary, rather than knowing lots of words causing faster computation.
(iv) I totally love the U-shaped development of filler-gap knowledge in the Gagliardi et al. (submitted) study. It's nice to see an example of this qualitative behavior in a realm besides morphology. The explanation seems similar, too -- a more sophisticated view of the input causes errors, which take some time to recover from. But the initial simplified view leads to surface behavior that seems right, even if the underlying representation isn't at that point. Hence, U-shaped performance curve. Same for the White et al. 2011 study -- U-shaped learning in syntactic bootstrapping for the win.
(v) I really liked the note on p.45 in the conclusion about how the input vs. intake distinction could really matter for L2 acquisition. It's nice to see some explicit ideas about what the skew that occurs is and why it might be occurring. (Basically, this feels like a more explicit form of the "less is more" hypothesis, where adult processing is distorting the input in predictable ways.)
Friday, January 24, 2014
Next time on 2/14/14 @ 3pm in SBSG 2221 = Omaki & Lidz 2013 Manuscript
Thanks to everyone who was able to join us for our thorough and thoughtful discussion of the Meylan et al. 2014 manuscript! Next time on Friday February 14 at 3pm in SBSG 2221, we'll be looking at an article manuscript that argues for the need to consider the development of children's processing abilities at the same time as we consider their acquisition of knowledge. This is particularly relevant to computational modelers who must explicitly model what the child's input looks like and how that input is used, for example.
Omaki, A. & Lidz, J. 2013. Linking parser development to acquisition of linguistic knowledge. Manuscript, Johns Hopkins University and University of Maryland, College Park. Please do not cite without permission from Akira Omaki.
Wednesday, January 22, 2014
Some thoughts on Meylan et al. 2014 Manuscript
One of the things I really enjoyed about this paper was the framing they give to explain why we should care about the emergence of grammatical categories, with respect to the existing debate between (some of) the nativists and (some of) the constructivists. Of course I'm always a fan of clever uses of Bayesian inference to problems in language acquisition, but I sometimes really miss the level of background story that we get here. (So hurrah!)
That being said, I was somewhat surprised to see the conclusion M&al2014 drew with respect to their results that (some kind of) a nativist view wasn't supported. To me, the fact that we see very early rapid development of this grammatical category knowledge is an unexpected thing from the "gradual emergence based on data" story (i.e., constructivist perspective). So, what's causing the rapid development? I know it's not the focus of M&al2014's work here, but positing some kind of additional learning guidance seems necessary to explain these results. And until we have a story for how that guidance would be learned, the "it's innate" answer is a pretty good placeholder. So, for me, that places the results in the nativist side, though maybe not the strict "grammatical categories are innate" version. Maybe I'm being unfair to the constructivist side, though -- would they have an explanation for the rapid, early development?
Another very cool thing was the application of this approach to the Speechome data set. It's been around for awhile, but we don't have a lot of studies that use it and it's such an amazing resource. One of the things I wondered, though, was whether the evaluation metric M&al2014 propose can only work if you have this density of data. It seems like that might be true, given the issues with confidence intervals on the CHILDES datasets. If so, this is different from Yang's metric [Yang 2013] which can be used on much smaller datasets. (My understanding is that as long as you have enough data to form a Zipfian distribution, you have enough for Yang's metric to be applied.)
One thing I didn't quite follow was the argument about why only a developmental analysis is possible, rather than both a developmental and a comparative analysis. I completely understand that adults may have different values for their generalized determiner preferences, but we assume that they realize determiners are a grammatical class. So, given this, whatever range of values adults have is the target state for acquisition, right? And this should allow a comparative analysis between wherever the child is and wherever the adult is. (Unless I'm missing something about this.)
Some more targeted thoughts:
As a completely nit-picky thing that probably doesn't matter, it took me a second to get used to calling grammatical categories syntactic abstractions. I get that they're the basis for (many) syntactic generalizations, but I wouldn't have thought of them as syntactic, per se. (Clearly, this is just a terminology issue, and other researchers that M&al2014 cite definitely have called it syntactic knowledge, too.)
M&al2014 state in the previous work section that Yang's metric is "not well-suited to discovering if a child could be less than fully productive at a given stage of development". I'm not sure I understand why this is so - if the observed overlap in the child's output is less than the expected overlap from a fully productive system, isn't that exactly the indicator of a less than fully productive system?
In the generative model M&al2014 use, they have a latent variable that represents the unrecorded caregiver input (DA), which is assumed to be drawn from the same distribution as the observed caregiver input (dA). I don't follow what this variable contributes, especially if it follows the same distribution as the observed data.
The table just below figure 4: I'm not sure I followed this. What would rich morphology be for English data, for example? And are the values for "Current" the v value inferred for the child? Are the Yang 2013 values calculated based on his expected overlap metric?
I wonder if the reason there were developmental changes found in the Speechome corpus is more about having enough data in the appropriate age range (i.e., < 2 years old). The other corpora had a much wider range of ages, and it could very well be that the ones that included younger-than-2-year-old data had older-age data included in the earliest developmental window investigated.
There's a claim made in the discussion that "no previous analysis has taken into account the input that individual children hear in judging whether their subsequent determiner usage has changed its productivity". I think what M&al2014 intend is something related to the explicit modeling of how much of the productions are imitated chunks, and if so, that seems completely fine (though one could argue that the Yang 2010 manuscript goes into quite some detail modeling this option). However, the way the current sentence reads, it seems a bit odd to say no previous analysis has cared about the input -- certainly Yang's metric can be used to assess productivity in child-directed speech utterances, which are the children's input. This is how a comparative analysis would presumably be made using Yang's metric.
Similarly, there's a claim near the end that the Bayesian analysis "makes inferences regarding developmental change of continuity in a single child possible". While it's true that this can be done with the Bayesian analysis, there seems to be an implicit claim that the other metrics can't do this. But I'm pretty sure it can also be done with the other metrics out there (e.g., Yang's). You basically apply the metric to data at multiple time points, and track the change, just as M&al2014 did here with the Bayesian metric.
~~~
References
Yang, C. 2013. Onotogeny and philogeny of language. 2013. Proceedings of the National Academy of Science, 110 (16). doi:10.1073/pnas.1216803110.
That being said, I was somewhat surprised to see the conclusion M&al2014 drew with respect to their results that (some kind of) a nativist view wasn't supported. To me, the fact that we see very early rapid development of this grammatical category knowledge is an unexpected thing from the "gradual emergence based on data" story (i.e., constructivist perspective). So, what's causing the rapid development? I know it's not the focus of M&al2014's work here, but positing some kind of additional learning guidance seems necessary to explain these results. And until we have a story for how that guidance would be learned, the "it's innate" answer is a pretty good placeholder. So, for me, that places the results in the nativist side, though maybe not the strict "grammatical categories are innate" version. Maybe I'm being unfair to the constructivist side, though -- would they have an explanation for the rapid, early development?
Another very cool thing was the application of this approach to the Speechome data set. It's been around for awhile, but we don't have a lot of studies that use it and it's such an amazing resource. One of the things I wondered, though, was whether the evaluation metric M&al2014 propose can only work if you have this density of data. It seems like that might be true, given the issues with confidence intervals on the CHILDES datasets. If so, this is different from Yang's metric [Yang 2013] which can be used on much smaller datasets. (My understanding is that as long as you have enough data to form a Zipfian distribution, you have enough for Yang's metric to be applied.)
One thing I didn't quite follow was the argument about why only a developmental analysis is possible, rather than both a developmental and a comparative analysis. I completely understand that adults may have different values for their generalized determiner preferences, but we assume that they realize determiners are a grammatical class. So, given this, whatever range of values adults have is the target state for acquisition, right? And this should allow a comparative analysis between wherever the child is and wherever the adult is. (Unless I'm missing something about this.)
Some more targeted thoughts:
As a completely nit-picky thing that probably doesn't matter, it took me a second to get used to calling grammatical categories syntactic abstractions. I get that they're the basis for (many) syntactic generalizations, but I wouldn't have thought of them as syntactic, per se. (Clearly, this is just a terminology issue, and other researchers that M&al2014 cite definitely have called it syntactic knowledge, too.)
M&al2014 state in the previous work section that Yang's metric is "not well-suited to discovering if a child could be less than fully productive at a given stage of development". I'm not sure I understand why this is so - if the observed overlap in the child's output is less than the expected overlap from a fully productive system, isn't that exactly the indicator of a less than fully productive system?
In the generative model M&al2014 use, they have a latent variable that represents the unrecorded caregiver input (DA), which is assumed to be drawn from the same distribution as the observed caregiver input (dA). I don't follow what this variable contributes, especially if it follows the same distribution as the observed data.
The table just below figure 4: I'm not sure I followed this. What would rich morphology be for English data, for example? And are the values for "Current" the v value inferred for the child? Are the Yang 2013 values calculated based on his expected overlap metric?
I wonder if the reason there were developmental changes found in the Speechome corpus is more about having enough data in the appropriate age range (i.e., < 2 years old). The other corpora had a much wider range of ages, and it could very well be that the ones that included younger-than-2-year-old data had older-age data included in the earliest developmental window investigated.
There's a claim made in the discussion that "no previous analysis has taken into account the input that individual children hear in judging whether their subsequent determiner usage has changed its productivity". I think what M&al2014 intend is something related to the explicit modeling of how much of the productions are imitated chunks, and if so, that seems completely fine (though one could argue that the Yang 2010 manuscript goes into quite some detail modeling this option). However, the way the current sentence reads, it seems a bit odd to say no previous analysis has cared about the input -- certainly Yang's metric can be used to assess productivity in child-directed speech utterances, which are the children's input. This is how a comparative analysis would presumably be made using Yang's metric.
Similarly, there's a claim near the end that the Bayesian analysis "makes inferences regarding developmental change of continuity in a single child possible". While it's true that this can be done with the Bayesian analysis, there seems to be an implicit claim that the other metrics can't do this. But I'm pretty sure it can also be done with the other metrics out there (e.g., Yang's). You basically apply the metric to data at multiple time points, and track the change, just as M&al2014 did here with the Bayesian metric.
~~~
References
Yang, C. 2013. Onotogeny and philogeny of language. 2013. Proceedings of the National Academy of Science, 110 (16). doi:10.1073/pnas.1216803110.
Thursday, January 9, 2014
Next time on 1/24/14 @ 3pm in SBSG 2221 = Meylan et al. 2014 Manuscript
It looks like the best collective time to meet will be Fridays at 3pm for this quarter, so that's what we'll plan on. Our first meeting will be in a few weeks on January 24. Our complete schedule is available on the webpage at
On Jan 24, we'll be looking at an article that examines a formal metric to gauge productivity for grammatical categories, based on hierarchical Bayesian modeling.
UPDATE for Jan 24: Michael Frank was kind enough to provide us with an updated version of the 2013 paper (2013 version linked below), which they're intending to submit for a journal publication. It's already received some outside feedback, and they'd be delighted to hear any thoughts we had on it. Michael preferred the manuscript not be posted publicly however, so I've sent it around as an attachment to the mailing list.
Meylan, S., Frank, M. C., & Levy, R. 2013. Modeling the development of determiner productivity in children's early speech. Proceedings of the 35th Annual Meeting of the Cognitive Science Society.
Categorical productivity is typically used to determine when abstract knowledge that a category actually exists is acquired (think "VERB exists, not just see and kiss and want! Woweee! Who knew?"), which is a fundamental building block for more complex linguistic knowledge.
I think the metric proposed in this session's article is particularly useful to compare and contrast against the metric that's been proposed recently by Yang (which is based on straight probability calculations), so I encourage you to have a look at that one as well:
Yang, C. 2013. Onotogeny and philogeny of language. 2013. Proceedings of the National Academy of Science, 110 (16). doi:10.1073/pnas.1216803110.
See you on Jan 24!
Wednesday, December 4, 2013
See you in the winter!
Thanks so much to everyone who was able to join us for our thoughtful, spirited discussion today, and to everyone who's joined us throughout the fall quarter! The CoLa Reading Group will resume again in the winter quarter. As always, feel free to send me suggestions of articles you're interested in reading, especially if you happen across something particularly interesting!
Monday, December 2, 2013
Some thoughts on Nematzadeh et al. 2013
So, I can start off by saying that there are many things about this paper that warmed the cockles of my heart. First, I love that modeling is highlighted as an explanatory tool. To me, that's one of the best things about computational modeling - the ability to identify an explanation for observed behavior, in addition to being able to produce said behavior. I also love that psychological constraints and biases were being incorporated into the model. This is that algorithmic/process-level-style model that I really enjoy working with, since it focuses on the connection between the abstract representation of what's going on and what people actually are doing. Related to both of the above, I was very happy to see how the model made assumptions concrete and thus isolated (potential) explanatory factors within the model. Now, maybe we don't always agree with how an assumption has been instantiated (see the note on novelty below)- but at least we know it's an assumption and we can see that version of it in action. And that is definitely a good thing, in my (admittedly biased) opinion.
Some more specific thoughts:
I found the general behavioral result from Vlach et al. 2008 about the "spacing effect" to be interesting, where learning was better when items are distributed over a period of time, rather than occurring one right after another. This is the opposite of "burstiness", which (I thought) is supposed to facilitate other types of learning (e.g., word segmentation). Maybe this has to do with the complexity of the thing being learned, or what existing framework there is for learning it (since I believe the Vlach et al. experiments were with adults)?
I thought the semantic representation of the scene as a collection of features was a nice step towards what the learner's representation probably is like (rather than just individual referent objects). When dealing with novel objects and more mature learners, this seems much more likely to me. On the other hand, I was a little fuzzy on how exactly the features and their feature weights were derived for the novel objects. (It's mentioned briefly in the Input Generation section that each word's true meaning is a vector of semantic features, but I missed completely how those are selected.)
Novelty: Nematzadeh et al. (N&al) implement novelty as an inverse function of recency. There's something obviously right about this, but I wonder about other definitions of novelty, like something that taps into overall frequency of this item's appearance (so, novel because it's fairly rare in the input). I'm not sure how this other definition (or a novelty implementation that incorporates both recency and overall frequency) would jive with the experimental results N&al are trying to explain.
Technical side note, related to the above: I had some trouble interpreting equation (2) - is the difference between t and tlastw a fraction of some kind? Maybe because time is measured in minutes, but the presentation durations are in seconds? Otherwise, novelty could become negative, which seems a bit weird.
I was thinking some about the predictions of the model, based on figure 4 and the discussion following it, where N&al are trying to make the model replicate certain experimental results. I think their model would predict that if learners had longer to learn the simplest condition (2 x 2), i.e., the duration of presentation was longer so the semantic representations didn't decay so quickly, that condition should then be the one best learned. That is, the "desirable difficulty" benefit is really about how memory decay doesn't happen so quickly for the 3 x 3 condition, as compared to the 2 x 2 condition.
I found it incredibly interesting that the behavioral experiment Vlach & Sandhofer 2010 (V&S) conducted just happened to have exactly the right item spacing/ordering/something else to yield the interesting results they found, but other orderings of those same items would be likely to yield different (perhaps less interesting) results. You sort of have to wonder how V&S happened upon just the right order - good experiment piloting, I guess? Though at the end of the discussion section, N&al seem to back off from claiming it's all about the order of item presentation, since none of the obvious variables potentially related to order (average spacing, average time since last presentation, average context familiarity) seemed to correlate with the output scores.
Some more specific thoughts:
I found the general behavioral result from Vlach et al. 2008 about the "spacing effect" to be interesting, where learning was better when items are distributed over a period of time, rather than occurring one right after another. This is the opposite of "burstiness", which (I thought) is supposed to facilitate other types of learning (e.g., word segmentation). Maybe this has to do with the complexity of the thing being learned, or what existing framework there is for learning it (since I believe the Vlach et al. experiments were with adults)?
I thought the semantic representation of the scene as a collection of features was a nice step towards what the learner's representation probably is like (rather than just individual referent objects). When dealing with novel objects and more mature learners, this seems much more likely to me. On the other hand, I was a little fuzzy on how exactly the features and their feature weights were derived for the novel objects. (It's mentioned briefly in the Input Generation section that each word's true meaning is a vector of semantic features, but I missed completely how those are selected.)
Novelty: Nematzadeh et al. (N&al) implement novelty as an inverse function of recency. There's something obviously right about this, but I wonder about other definitions of novelty, like something that taps into overall frequency of this item's appearance (so, novel because it's fairly rare in the input). I'm not sure how this other definition (or a novelty implementation that incorporates both recency and overall frequency) would jive with the experimental results N&al are trying to explain.
Technical side note, related to the above: I had some trouble interpreting equation (2) - is the difference between t and tlastw a fraction of some kind? Maybe because time is measured in minutes, but the presentation durations are in seconds? Otherwise, novelty could become negative, which seems a bit weird.
I was thinking some about the predictions of the model, based on figure 4 and the discussion following it, where N&al are trying to make the model replicate certain experimental results. I think their model would predict that if learners had longer to learn the simplest condition (2 x 2), i.e., the duration of presentation was longer so the semantic representations didn't decay so quickly, that condition should then be the one best learned. That is, the "desirable difficulty" benefit is really about how memory decay doesn't happen so quickly for the 3 x 3 condition, as compared to the 2 x 2 condition.
I found it incredibly interesting that the behavioral experiment Vlach & Sandhofer 2010 (V&S) conducted just happened to have exactly the right item spacing/ordering/something else to yield the interesting results they found, but other orderings of those same items would be likely to yield different (perhaps less interesting) results. You sort of have to wonder how V&S happened upon just the right order - good experiment piloting, I guess? Though at the end of the discussion section, N&al seem to back off from claiming it's all about the order of item presentation, since none of the obvious variables potentially related to order (average spacing, average time since last presentation, average context familiarity) seemed to correlate with the output scores.
Wednesday, November 20, 2013
Next time on 12/4/13 @ 2:30pm in SBSG 2221 = Nematzadeh et al. 2013
Thanks to everyone who was able to join us for our feisty and thoughtful discussion of Lewis & Frank 2013! Next time on December 4 at 2:30pm in SBSG 2221, we'll be looking at an article that explores the kinds of difficulties in word-learning that can paradoxically help long-term learning and why they help, using a computational modeling approach:
Nematzadeh, A., Fazly, A., & Stevenson, S. 2013. Desirable Difficulty in Learning: A Computational Investigation. Proceedings of the 35th Annual Meeting of the Cognitive Science Society.
See you then!
Monday, November 18, 2013
Some thoughts on Lewis & Frank 2013
I'm always a fan of learning models that involve solving different problems simultaneously, with the idea of leveraging information from one problem to help solve the other (Feldman et al. 2013 and Dillon et al. 2013 are excellent examples of this, IMHO). For Lewis & Frank (L&F), the two problems are related to word learning: how to pick the referent from a set of referents and how to pick which concept class that referent belongs to (which they relate to how to generalize that label appropriately). I have to say that I struggled to understand how they incorporated the second problem, though -- it doesn't seem like the concept generalization w.r.t. subordinate vs. superordinate classes maps in a straightforward way to the feature analysis they're describing. (More on this below.) I was also a bit puzzled by their assumption of where the uncertainty in learning originates from and the link they describe between what they did and the origin/development of complex concepts (more on these below, too).
On generalization & features: If we take the example in their Figure 1, it seems like the features could be something like f1 = "fruit", f2 = "red", and f3 = "apple". The way they talk about generalization is as underspecification of feature values, which feels right. So if we say f1 is the only important feature, then this corresponds nicely to the idea of "fruit" as a superordinate class. But what if we allow f2 to be the important feature? Is "red" the superordinate class of "red" things? Well, in a sense, I suppose. But this falls outside of the noun-referent system that they're working in - "red" spans many referents, because it's a property. Maybe this is my misunderstanding in trying to map this whole thing to subordinate and superordinate classes, like Xu & Tenenbaum 2007 talk about, but it felt like that's what L&F intended, given the model in Figure 2 that's grounded in Objects at the observable level and the behavioral experiment they actually ran.
On where the uncertainty comes from: L&F mention in the Design of the Model section that the learning model assumes "the speaker could in principle have been mistaken about their referent or misspoken". From a model building perspective, I understand that this is easier to incorporate and allows graded predictions (which are necessary to match the empirical data), but from the cognitive perspective, this seems really weird to me. Do we have reason to believe children assume their speakers are unreliable? I was under the impression children assume their speakers are reliable as a default. Maybe there's a better place to work this uncertainty in - approximated inference from a sub-optimal learner or something like that. Also, as a side note, it seems really important to understand how the various concepts/features are weighted by the learner. Maybe that's where uncertainty could be worked in at the computational level.
On the origin/development of concepts: L&F mention in the General Discussion that "the features are themselves concepts that can be considered as primitives in the construction of more complex concepts", and then state that their model "describes how a learner might bootstrap from these primitives to infer more and complex concepts". This sounds great, but I was unclear how exactly to do that. Taking the f1, f2, and f3 from above, for example, I get that those are primitive features. So the concepts are then things that can be constructed out of some combination of their values (whether specified or unspecified)? And then where does the development come in? Where is the combination (presumably novel) that allows the construction of new features? I understand that these could be the building units for such a model, but I didn't see how the current model shows us something about that.
Behavioral experiment implementation: I'm definitely a fan of matching a model to controlled behavioral data, but I wonder about the specific kind of labeling they gave their subjects. It seems like they intended "dax bren nes" to be the label for one object shown (it's just unclear which it is - but basically, this might as well be a trisyllabic word "daxbrennes" ). This is a bit different from standard cross-situational experiments, where multiple words are given for multiple objects. Given that subjects are tested with that same label, I guess the idea is that it simplifies the learning situation.
Results: I struggled a bit to decipher the results in Figure 5 - I'm assuming the model predictions are for the different experimental contexts, ordered by human uncertainty about how much to generalize to the superordinate class. Is the lexicon posited by the model just how many concepts to map to "dax-bren-nes", where concept = referent?
~~~
References
B. Dillon, E. Dunbar, & W. Idsardi. 2013. A single-stage approach to learning phonological categories: Insights from Inuktitut. Cognitive Science, 37, 344-377.
Feldman, N. H., Griffiths, T. L., Goldwater, S., Morgan, J. L. 2013. "A role for the developing lexicon in phonetic category acquisition." Psychological Review, 120(4), 751-778.
Xu, F., & Tenenbaum, J. 2007. Word Learning as Bayesian Inference. Psychological Review, 114(2), 245-272.
On generalization & features: If we take the example in their Figure 1, it seems like the features could be something like f1 = "fruit", f2 = "red", and f3 = "apple". The way they talk about generalization is as underspecification of feature values, which feels right. So if we say f1 is the only important feature, then this corresponds nicely to the idea of "fruit" as a superordinate class. But what if we allow f2 to be the important feature? Is "red" the superordinate class of "red" things? Well, in a sense, I suppose. But this falls outside of the noun-referent system that they're working in - "red" spans many referents, because it's a property. Maybe this is my misunderstanding in trying to map this whole thing to subordinate and superordinate classes, like Xu & Tenenbaum 2007 talk about, but it felt like that's what L&F intended, given the model in Figure 2 that's grounded in Objects at the observable level and the behavioral experiment they actually ran.
On where the uncertainty comes from: L&F mention in the Design of the Model section that the learning model assumes "the speaker could in principle have been mistaken about their referent or misspoken". From a model building perspective, I understand that this is easier to incorporate and allows graded predictions (which are necessary to match the empirical data), but from the cognitive perspective, this seems really weird to me. Do we have reason to believe children assume their speakers are unreliable? I was under the impression children assume their speakers are reliable as a default. Maybe there's a better place to work this uncertainty in - approximated inference from a sub-optimal learner or something like that. Also, as a side note, it seems really important to understand how the various concepts/features are weighted by the learner. Maybe that's where uncertainty could be worked in at the computational level.
On the origin/development of concepts: L&F mention in the General Discussion that "the features are themselves concepts that can be considered as primitives in the construction of more complex concepts", and then state that their model "describes how a learner might bootstrap from these primitives to infer more and complex concepts". This sounds great, but I was unclear how exactly to do that. Taking the f1, f2, and f3 from above, for example, I get that those are primitive features. So the concepts are then things that can be constructed out of some combination of their values (whether specified or unspecified)? And then where does the development come in? Where is the combination (presumably novel) that allows the construction of new features? I understand that these could be the building units for such a model, but I didn't see how the current model shows us something about that.
Behavioral experiment implementation: I'm definitely a fan of matching a model to controlled behavioral data, but I wonder about the specific kind of labeling they gave their subjects. It seems like they intended "dax bren nes" to be the label for one object shown (it's just unclear which it is - but basically, this might as well be a trisyllabic word "daxbrennes" ). This is a bit different from standard cross-situational experiments, where multiple words are given for multiple objects. Given that subjects are tested with that same label, I guess the idea is that it simplifies the learning situation.
Results: I struggled a bit to decipher the results in Figure 5 - I'm assuming the model predictions are for the different experimental contexts, ordered by human uncertainty about how much to generalize to the superordinate class. Is the lexicon posited by the model just how many concepts to map to "dax-bren-nes", where concept = referent?
~~~
References
B. Dillon, E. Dunbar, & W. Idsardi. 2013. A single-stage approach to learning phonological categories: Insights from Inuktitut. Cognitive Science, 37, 344-377.
Feldman, N. H., Griffiths, T. L., Goldwater, S., Morgan, J. L. 2013. "A role for the developing lexicon in phonetic category acquisition." Psychological Review, 120(4), 751-778.
Xu, F., & Tenenbaum, J. 2007. Word Learning as Bayesian Inference. Psychological Review, 114(2), 245-272.
Wednesday, November 6, 2013
Next time on 11/20/13 @ 2:30pm in SBSG 2221 = Lewis & Frank 2013
Thanks to everyone who was able to join us for our vigorous and thoughtful discussion of Marcus & Davis 2013! Next time on November 20 at 2:30pm in SBSG 2221, we'll be looking at an article that discusses how to solve two problems related to word learning simultaneously, using hierarchical Bayesian modeling and evaluating against human behavioral data:
Lewis, M. & Frank. M. 2013. An integrated model of concept learning and word-concept mapping.Proceedings of the 35th Annual Meeting of the Cognitive Science Society.
See you then!
Monday, November 4, 2013
Some thoughts on Marcus & Davis (2013)
(...and a little also on Jones & Love 2011)
One of the things that struck me about Marcus & Davis (2013) [M&D] is that they seem to be concerned with identifying what the priors are for learning. But what I'm not sure of is how you distinguish the following options:
(a) sub-optimal inference over optimal priors
(b) optimal inference over sub-optimal priors
(c) sub-optimal inference over sub-optimal priors
M&D seem to favor option (a), but I'm not sure there's an obvious reason to do so. Jones & Love 2011 [J&L] mention the possibility of "bounded rationality", which is something like "be as optimal as possible in your inference, given the prior and the processing limitations you have". That sounds an awful lot like (c), and seems like a pretty reasonable option to explore. The concern in general with what the priors are actually dovetails quite nicely with traditional linguistic explorations of how to define (constrain) the learner's hypothesis space appropriately to make successful inference possible. Also, J&L are quite aware of this too, and underscore the importance of selecting the priors appropriately.
That being said, no matter what priors and inference processes end up working, there's clear utility in being explicit about all the assumptions that yield a match to human behavior, which M&D want (and I'm a huge fan of this myself: see my commentary on a recent article here where I happily endorse this). Once you've identified the necessary pieces that make a learning strategy work, you can then investigate (or at least discuss) which assumptions are necessarily optimal. That may not be an easy task, but it seems like a step in the right direction.
M&D seem to be unhappy with probabilistic models as a default assumption - and okay, that's fine. But it does seem important to recognize that probabilistic reasoning is a legitimate option. And maybe some of cognition is probabilistic and some isn't - I don't think there's a compelling reason to believe that cognition has to be all one or all the other. (I mean, after all, cognition is made up of a lot of different things.) In this vein, I think a reasonable thing that M&D would like is for us to not just toss out non-probabilistic options that work really well solely because they're non-probabilistic.
On a related note, I very much agree with one of the last things M&D note, which is that we should be explicit about "what would constitute evidence that a probabilistic approach is not appropriate for a particular task or domain". I'm not sure myself what that evidence would look like, since even categorical behavior can be simulated by a probabilistic model that just thresholds. Maybe if it's more "economical" (however we define that) to not have a probabilistic model, and there exists a non-probabilistic model that accomplishes the same thing?
~~~
A few comments about Jones & Love 2011 [J&L]:
J&L seem very concerned with the recent focus in the Bayesian modeling world on existence proofs for various aspects of cognition. They do mention later in their article (around section 6, I think), that existence proofs are a useful starting point, however -- they just don't want research to stop there. An existence proof that a Bayesian learning strategy can work for some problem should be the first step for getting a particular theory on the table as a real possibility worth considering (e.g., whatever's in the priors for that particular learning strategy that allowed Bayesian inference to succeed, as well as the Bayesian inference process itself).
Overall, J&L seem to make a pretty strong call for process models (i.e., algorithmic-level models, instead of just computational-level models). Again, this seems like a natural follow-up once you have a computational-level model you're happy with. So the main point is simply not to rest on your Bayesian inference laurels once you have your existence proof at the computational level for some problem in cognition. The Chater et al. 2011 commentary to J&L note that many Bayesian modelers are moving in this direction already, creating "rational process" models.
~~~
References
Chater, N., Goodman, N., Griffiths, T., Kemp, C., Oaksford, M., & Tenenbaum, J. 2011. The imaginary fundamentalists: The unshocking truth about Bayesian cognitive science. Behavioral and Brain Sciences, 34 (4), 194-196.
Jones, M. & Love, M. 2011. Bayesian Fundamentalism or Enlightenment? On the explanatory status and theoretical contributions of Bayesian models of cognition. Behavioral and Brain Sciences, 34 (4), 169-188.
Pearl, L. 2013. Evaluating strategy components: Being fair. [lingbuzz]
One of the things that struck me about Marcus & Davis (2013) [M&D] is that they seem to be concerned with identifying what the priors are for learning. But what I'm not sure of is how you distinguish the following options:
(a) sub-optimal inference over optimal priors
(b) optimal inference over sub-optimal priors
(c) sub-optimal inference over sub-optimal priors
M&D seem to favor option (a), but I'm not sure there's an obvious reason to do so. Jones & Love 2011 [J&L] mention the possibility of "bounded rationality", which is something like "be as optimal as possible in your inference, given the prior and the processing limitations you have". That sounds an awful lot like (c), and seems like a pretty reasonable option to explore. The concern in general with what the priors are actually dovetails quite nicely with traditional linguistic explorations of how to define (constrain) the learner's hypothesis space appropriately to make successful inference possible. Also, J&L are quite aware of this too, and underscore the importance of selecting the priors appropriately.
That being said, no matter what priors and inference processes end up working, there's clear utility in being explicit about all the assumptions that yield a match to human behavior, which M&D want (and I'm a huge fan of this myself: see my commentary on a recent article here where I happily endorse this). Once you've identified the necessary pieces that make a learning strategy work, you can then investigate (or at least discuss) which assumptions are necessarily optimal. That may not be an easy task, but it seems like a step in the right direction.
M&D seem to be unhappy with probabilistic models as a default assumption - and okay, that's fine. But it does seem important to recognize that probabilistic reasoning is a legitimate option. And maybe some of cognition is probabilistic and some isn't - I don't think there's a compelling reason to believe that cognition has to be all one or all the other. (I mean, after all, cognition is made up of a lot of different things.) In this vein, I think a reasonable thing that M&D would like is for us to not just toss out non-probabilistic options that work really well solely because they're non-probabilistic.
On a related note, I very much agree with one of the last things M&D note, which is that we should be explicit about "what would constitute evidence that a probabilistic approach is not appropriate for a particular task or domain". I'm not sure myself what that evidence would look like, since even categorical behavior can be simulated by a probabilistic model that just thresholds. Maybe if it's more "economical" (however we define that) to not have a probabilistic model, and there exists a non-probabilistic model that accomplishes the same thing?
~~~
A few comments about Jones & Love 2011 [J&L]:
J&L seem very concerned with the recent focus in the Bayesian modeling world on existence proofs for various aspects of cognition. They do mention later in their article (around section 6, I think), that existence proofs are a useful starting point, however -- they just don't want research to stop there. An existence proof that a Bayesian learning strategy can work for some problem should be the first step for getting a particular theory on the table as a real possibility worth considering (e.g., whatever's in the priors for that particular learning strategy that allowed Bayesian inference to succeed, as well as the Bayesian inference process itself).
Overall, J&L seem to make a pretty strong call for process models (i.e., algorithmic-level models, instead of just computational-level models). Again, this seems like a natural follow-up once you have a computational-level model you're happy with. So the main point is simply not to rest on your Bayesian inference laurels once you have your existence proof at the computational level for some problem in cognition. The Chater et al. 2011 commentary to J&L note that many Bayesian modelers are moving in this direction already, creating "rational process" models.
~~~
References
Chater, N., Goodman, N., Griffiths, T., Kemp, C., Oaksford, M., & Tenenbaum, J. 2011. The imaginary fundamentalists: The unshocking truth about Bayesian cognitive science. Behavioral and Brain Sciences, 34 (4), 194-196.
Jones, M. & Love, M. 2011. Bayesian Fundamentalism or Enlightenment? On the explanatory status and theoretical contributions of Bayesian models of cognition. Behavioral and Brain Sciences, 34 (4), 169-188.
Pearl, L. 2013. Evaluating strategy components: Being fair. [lingbuzz]
Wednesday, October 23, 2013
Next time on 11/6/13 @ 2:30pm in SBSG 2221 = Marcus & David 2013
Thanks to everyone who was able to join us for our lively and informative discussion of Ambridge et al. (in press)! Next time on November 6 at 2:30pm in SBSG 2221, we'll be looking at an article that discusses how probabilistic models of higher-level cognition (including language) are used in cognitive science:
Marcus, G. & Davis, E. 2013. How Robust Are Probabilistic Models of Higher-Level Cognition? Psychological Science, published online Oct 1, 2013, doi:10.1177/095679761349541.
I would also strongly recommend a target article and commentary related to this topic that were written fairly recently:
Jones, M. & Love, M. 2011. Bayesian Fundamentalism or Enlightenment? On the explanatory status and theoretical contributions of Bayesian models of cognition. Behavioral and Brain Sciences, 34 (4), 169-188.
Chater, N., Goodman, N., Griffiths, T., Kemp, C., Oaksford, M., & Tenenbaum, J. 2011. The imaginary fundamentalists: The unshocking truth about Bayesian cognitive science. Behavioral and Brain Sciences, 34 (4), 194-196.
http://www.socsci.uci.edu/~lpearl/colareadinggroup/readings/JonesLove2011_BayesianModelsInCogSci.pdf
(Both target article and commentary are included in the pdf file linked above.)
See you then!
Monday, October 21, 2013
Some thoughts on Ambridge et al. in press
This article really hit home for me, since it talks about things I worry about a fair bit with respect to Universal Grammar and language learning in general -- so much so, that I ended up writing a lot more about it than I typically do for the articles we read. Conveniently, this is a target article that's asking for commentaries, so I'm going to put some of my current thoughts here as a sort of teaser for the commentary I plan to submit.
~~~
~~~
The basic issue that the authors (AP&L) highlight about proposed learning strategies seems
exactly right: What will actually
work, and what exactly makes it work? They note that “…nothing is gained by positing components of innate knowledge that
do not simplify the problem faced by language learners” (p.56, section 7.0),
and this is absolutely true. To examine how well several current learning
strategy proposals work that involve innate, linguistic knowledge, AP&L present
evidence from a commendable range of linguistic phenomena, from what might be
considered fairly fundamental knowledge (e.g., grammatical categories) to
fairly sophisticated knowledge (e.g., subjacency and binding). In each case, AP&L
identify the shortcomings of some existing Universal Grammar (UG) proposals, and observe that these
proposals don’t seem to fare very well in realistic scenarios. The challenge at
the very end underscores this -- AP&L contend (and I completely agree) that
a learning strategy proposal involving innate knowledge needs to show “precisely
how a particular type of innate knowledge would help children acquire X” (p.56,
section 7.0).
More importantly, I believe this should be a metric that any component of a learning strategy is
measured by. Namely, for any component
(whether innate or derived, whether language-specific or domain-general), we
need to not only propose that this component could help children learn some
piece of linguistic knowledge but also demonstrate at least “one way that a
child could do so” (p.57, section 7.0). To this end, I think it's important to highlight
how computational modeling is well suited for doing precisely this: for any
proposed component embedded in a learning strategy, modeling allows us to empirically
test that strategy in a realistic learning scenario. It’s my view that we
should test all potential learning strategies, including the ones AP&L
themselves propose as alternatives to the UG-based ones they find lacking. An additional and highly useful benefit of the
computational modeling methdology is that it forces us to recognize hidden
assumptions within our proposed learning strategies, a problem that AP&L
rightly recognize with many existing proposals.
This leads me to suggest certain criteria that any learning
strategy should satisfy, relating to its utility in principle and practice, as
well as its usability by children. Once we have a promising learning strategy
that satisfies these criteria, we can then concern ourselves with the
components comprising that strategy.
With respect to this, I want to briefly discuss the type of components AP&L
find unhelpful, since several of the components they would prefer might
still be reasonably classified as UG components. The main issue they have is
not with components that are innate and language-specific, but rather
components of this kind that in addition involve very precise knowledge. This
therefore does not rule out UG components that involve more general knowledge,
including (again) the components AP&L themselves propose. In addition, AP&L ask
for explicit examples of UG components that actually do work. I think one potentially UG component that’s part of a successful learning
strategy for syntactic islands (described in Pearl & Sprouse 2013) is a nice example of this: the bias to characterize wh-dependencies at a specific level of granularity. It's not obvious where this bias would come from (i.e., how it would be derived or what innate knowledge would lead to it), but it's crucial for the learning strategy it's a part of to work. As a bonus, that learning strategy also satisfies the criteria I suggest for evaluating learning strategies more
generally (utility and useability).
~~~
Reference:
Pearl, L., & Sprouse,
J. 2013. Syntactic islands and learning biases: Combining experimental syntax and computational modeling to investigate the language acquisition problem. Language
Acquisition, 20, 19–64.
Tuesday, October 1, 2013
Next time on 10/23/13 @ 2:30pm in SBSG 2221 = Ambridge et al. in press
It looks like the best collective time to meet will be Wednesdays at 2:30pm for this quarter, so that's what we'll plan on. Due to some of my own scheduling conflicts, our first meeting will be in a few weeks on October 23. Our complete schedule is available on the webpage at
On Oct 23, we'll be looking at an article that examines the utility of Universal Grammar based learning strategies in several different linguistic domains, arguing that they're not all that helpful at the moment:
Ambridge, B., Pine, J., & Lieven, E. 2013 in press. Child language acquisition: Why Universal Grammar doesn't help. Language.
See you then!
Wednesday, September 25, 2013
Fall quarter planning
I hope everyone's had a good summer break - and now it's time to gear up for the fall quarter of the reading group! :) The schedule of readings is now posted on the CoLa Reading group webpage, including readings on Universal Grammar, Bayesian modeling, and word learning:
http://www.socsci.uci.edu/~lpearl/colareadinggroup/schedule.html
Now all we need to do is converge on a specific day and time - please let me know what your availability is during the week. We'll continue our tradition of meeting for approximately one hour (and of course, posting on the discussion board here).
Thanks and see you soon!
Now all we need to do is converge on a specific day and time - please let me know what your availability is during the week. We'll continue our tradition of meeting for approximately one hour (and of course, posting on the discussion board here).
Thanks and see you soon!
Subscribe to:
Posts (Atom)