50.504 53.8 uh the uh
53.8 56.861 Well the uh uh I chose
56.861 59.471 the title was in the hope that it would uh
59.471 60.748 arouse some
60.748 63.004 interest and surprise
63.004 67.25 uh it sound- the first part, should we study language uh
67.25 69.582 certainly should sound strange
69.582 71.896 uh nobody actually says
71.896 74.818 as far as I know that we shouldn't study language
74.818 76.704 uh on the other hand uh
76.704 78.448 at least in those words
78.448 82.362 uh on the other hand it's widely uh held
82.362 83.701 indirectly
83.701 86.502 that that's the right conclusion
86.502 89.02 uh that is that there's nothing to study
89.02 91.554 uh I'll uh
91.554 93.909 there're actually two versions this
93.909 97.093 one of them is a strong version
97.093 99.103 which is based on
99.103 102.755 uh hypothetical cognitive processes
102.755 104.012 uh
104.012 106.103 has no uh
106.103 108.639 evidence to support it as far as I know
108.639 110.281 uh is uh
110.281 112.926 very- is based on very serious
112.926 115.309 confusions, I think, and also uh
115.309 120.379 on a departure from the methods of science at least since Galileo
120.379 122.104 and it's very widely held
122.104 126.282 uh and probably in the fields right surrounding linguistics
126.282 130.36 and cognitive science it's maybe a majority position or
130.36 131.333 getting there
131.333 133.869 the second version, which is weaker,
133.869 135.776 uh which I think is
135.776 138.09 substantial and uh
138.09 143.16 in fact treats language just like any other biological object
143.16 148.331 and I think has been ((at)) least the way I understand it has been a kind of a guideline for the uh
148.331 150.685 pursuit of the study of language since
150.685 154.486 its modern origins in the nineteen fifties
154.486 156.473 uh so those are the
156.473 159.251 two parts of what I want to talk about, one
159.251 162.176 uh an answer that says
162.176 165.718 you shouldn't study language because it's not there, but uh
165.718 169.904 it's posing the question the wrong way, and another, second
169.904 173.598 which says something kind of similar but in a weaker form
173.598 177.532 which I think makes some sense, and does offer good
177.532 179.706 guidelines for the study of language
179.706 181.471 well these uh
181.471 183.762 comments may sound extreme
183.762 186.819 they should, so I want to give a couple illustrations
186.819 189.495 uh
189.495 193.386 I can see anything
193.386 195.313 uh the uh
195.313 197.193 (())
197.193 199.059 Well, start with uh
199.059 205.79 an article that appeared in uh the journal uh Science, Science magazine, the journal of the uh
205.79 211.468 American Association for the Advancement of Science, the premier science weekly in
211.468 214.04 the United States, maybe the West,
214.04 218.633 uh a month ago there was an article uh by uh a review, by
218.633 222.946 uh N J Enfield ((he's)) from the Max Planck Institute
222.946 226.478 ((it's)) reviewing two books on the evolution of language
226.478 230.428 and he did find one useful contribution in them
230.428 233.524 uh interesting work by Tecumseh Fitch
233.524 236.338 on the mammalian vocal tract
236.338 239.158 of deer, dogs and so on --
239.158 246.283 uh what this has to do with the evolution of language we could ask, but interesting work -- the rest of
246.283 249.55 everything in the two books, he criticizes pretty harshly,
249.55 252.353 because it assumes that language exists,
252.353 265.812 uh contrary, I'm quoting him now, "contrary to the belief of many researchers in the cognitive sciences", himself included, "that language is entirely grounded in a constellu- constellation of cognitive capacities
265.812 270.226 that each taken separately has other functions as well"
270.226 274.68 which means that there's such a thing as language only in the sense that
274.68 285.517 there's such a thing as, say, uh, today's weather, uh, it's there, you know, it's a constellation of many factors that operate independently, but it's not
285.517 290.071 an object of inquiry for the sciences, or for any rational inquiry.
290.071 296.125 Uh he does cite one source to justify his conclusion -- a pretty strong conclusion --
296.125 304.003 unfortunately the source, whatever its merits, has virtually nothing to do with the question, you can look it up and check,
304.003 306.521 and I don't think he made a bad choice,
306.521 308.886 I don't know of any others that would have helped,
308.886 313.043 uh but I think he's correct in saying that uh
313.043 319.435 uh this is what many cognitive scientists believe; in fact I suspect that's an understatement.
319.435 400.882 
400.882 404.925 of course the transition from finite to unbounded is saltationist
404.925 406.758 couldn't be anything else
406.758 411.352 uh but that's a kind of dirty word in many circles, based on a
411.352 416.022 curious but widespread misunderstanding of evolutionary biology
416.022 421.379 (()) it's been understood for forty years that uh
421.379 424.587 that slight modifications in say uh
424.587 428.275 uh regulatory circuits can lead to substantial
428.275 453.348 
453.348 455.802 in the necessity of gradualist changes
455.802 457.505 itself based on
457.505 461.926 interesting but curious misinterpretations, I won't go into it
461.926 463.773 uh however in this case
463.773 465.618 there's just nothing to discuss.
465.618 475.948 
475.948 481.034 Uh another part of Enfield's critique, which is also quite widespread,
481.034 485.227 and uh significant uh has to do with the methodology
485.227 488.533 the methods by which one should study linguistic phenomena.
488.533 491.432 the title of his article is uh
491.432 493.724 "Without Social Context"
493.724 494.819 question mark,
494.819 498.631 and he's criticizing the fact that linguists uh
498.631 502.888 often study language without looking at the social context,
502.888 506.495 in other words, abstracted from social context.
506.495 512.668 and he says linguists don't seem to recognize that communication is possible uh without syntax
512.668 514.048 and therefore
514.048 517.958 it's kind of ancillary or unnecessary
517.958 521.326 and the editors, or maybe the author, I don't know who,
521.326 523.68 uh illustrated this with a
523.68 525.381 photo, put a picture
525.381 527.745 you should really look at the original article, cause
527.745 529.814 there's a big picture on the front page
529.814 531.316 which has three
531.316 532.553 infants
532.553 533.811 who are kind of
533.811 535.309 looking at each other
535.309 536.471 and gesturing
536.471 540.121 and the caption says "Communication without syntax".
540.121 541.357 that proves it
541.357 548.741 actually they could have had two bacteria (()) they also communicate without syntax (( )) the
548.741 550.218 uh that uh
550.218 554.322 uh picture, of course it's true that the study of communication
554.322 557.822 does involve social context, I mean that's practically
557.822 559.059 true by definition
559.059 563.581 so does the study of how animals use their visual systems
563.581 566.938 yeah they use it in context, you know,
566.938 569.433 look at each other, and so on
569.433 573.095 uh but uh I mean the study of vision
573.095 575.692 of course abstracts from social context
575.692 577.685 I mean you take say the classic work
577.685 586.383 you know, say, Hubell and Wiesel, or anybody (( )) or for that matter experimental work on how
586.383 588.392 uh children identify objects
588.392 590.097 you know Liz Spelke and others
590.097 593.019 it all abstracts from social context
593.019 597.413 uh the reason is it's science, you know, it's not butterfly collecting
597.413 601.695 so you look- you try to isolate what looks important, see if you can
601.695 602.812 figure that out
602.812 605.186 and the same is true of evolutionary biology
605.186 610.438 I mean an awful lot of evolutionary biology is based on the study of fruit flies
610.438 613.424 but fruit flies aren't investigated in the wild
613.424 615.351 um nobody would try to do that
615.351 618.334 they collect them in the wild, because there's no other way to get them
618.334 623.908 uh but then they're studied under laboratory conditions in which, as is known, the
623.908 631.798 uh you know the chromosomal uh uh proportions and in fact just about everything are different from in the wild but again
631.798 639.169 that's the way you do it, naturally abstracts from natural conditions again because it's science, not butterfly collecting
639.169 643.146 but we're being told, really, that we should go to butterfly collecting
643.146 647.049 well you know all of this has been unquestioned in the sciences
647.049 650.929 I mean at least since Galileo, and actually even farther back
650.929 653.071 uh and that uh
653.071 657.046 This critique, which is pretty common these days,
657.046 659.864 uh goes even farther. It uh
659.864 664.586 it repeats a common charge that the studies of language are flawed
664.586 669.986 severely flawed because they involve uh abstraction and idealization.
669.986 673.432 And that criticism is based on a kind of intriguing
673.432 675.785 misunderstanding of science altogether
675.785 679.876 uh the argument is well, take a look at physics -- you know, real science --
679.876 683.362 it's based on observations,
683.362 687.55 you know, not observations of things in their natural state,
687.55 692.651 like you don't try to say determine the laws of motion by uh
692.651 699.585 taking video tapes of uh leaves falling and do massive statistical analysis of them, and so on and so forth.
699.585 701.263 So you do experiments
701.263 703.679 and in fact a lot of the experiments are thought experiments
703.679 705.369 including Galileo's
705.369 708.409 classic experiments,and that goes right up to the present.
708.409 712.793 But any experiment ((is)) a high level abstraction
712.793 718.113 It's uh and theory-internal, as everybody who does experimental work knows.
718.113 724.497 uh you ((put it sort of like)) same ((if you're) studying any other topic, say bee communication
724.497 729.143 I mean again, ((if you)) take a look at the work on bee science, it involves
729.143 731.029 highly contrived uh
731.029 733.718 very intricate experiments that are
733.718 736.8 radically abstracted from natural conditions.
736.8 740.45 uh nobody suggests studying bee communication
740.45 745.635 again by taking a massive corpus, you know, huge uh
745.635 747.785 library of video tapes of
747.785 749.955 bees swarming around and uh
749.955 755.268 doing statistical analysis of it, and getting some prediction about what they're likely to do next
755.268 757.803 Incidentally, as in the case of
757.803 760.561 say what's going on outside the window,
760.561 761.818 if you did that
761.818 764.252 you'd get a much better prediction for what's
764.252 768.568 going on outside the window, or what bees are going to do next
768.568 771.67 uh then physicists can give
771.67 775.906 in the outside-the-window case, or that bee scientists can give in the uh
775.906 778.218 uh the in case of bee communication.
778.218 780.066 cause that's not what they're trying to do.
780.066 783.899 They're not trying to give a fair approximation to what might happen next.
783.899 786.211 They're trying to figure out what are the principles,
786.211 788.686 what's going on, you know, what are the mechanisms,
788.686 790.309 and in order to do that, you
790.309 792.132 automatically do uh
792.132 796.411 high level abstraction and idealization, sometimes called experiments
796.411 799.809 And sometimes in fact uh thought- thought experiments.
799.809 805.445 Well you could argue, and it makes sense, that linguists should uh do better experiments.
805.445 811.273 that just, you know, the kind of informal experiments that are done all the time, maybe they're not adequate.
811.273 816.787 So for example suppose you want to -- I'll give a couple of cases and come back to them, cause they're illustrative --
816.787 819.334 uh so suppose you want to uh take
819.334 820.429 kind of an old
820.429 822.154 saw from way back
822.154 825.659 you want to establish the fact that rules are
825.659 827.057 structure-dependent.
827.057 830.553 which means that uh if you have a sentence like
830.553 833.452 "Can eagles that fly swim?"
833.452 836.291 you're asking about swimming, not flying.
836.291 843.864 It's not a trivial fact, incidentally -- why shouldn't it- should- why can't it be the other way around?
843.864 850.42 For example, on uh grounds of elementary principles of parsing, and computational simplicity
850.42 853.544 uh it would be much easier if you had,
853.544 858.889 "Eagles who fly swim", if you had "can" associated with "fly", not swim, the closest thing
858.889 860.43 In other words, if you measured
860.43 866.22 relations by linear distance rather than structural distance, which is ((much one of)) complex notion,
866.22 870.197 but uh you don't, you do it by structural distance, so it's
870.197 871.738 old question why,
871.738 875.51 So maybe we should do better experiments to explain why
875.51 877.133 uh you know
877.133 880.893 "Can eagles ((that)) fly swim" is asking about swimming, not flying.
880.893 883.127 However, it's (( )) pointless
883.127 887.308 I mean if you did experiments on this you'd really be testing the experiment
887.308 891.203 if the experiment gave the wrong result, you'd say OK I did a lousy experiment
891.203 898.858 and you'd try to refine it so it so it gives the right results, ((because)) the results themselves are much more obvious than what any experiment could do
898.858 903.481 Uh or uh take uh take another case, a binding theory case.
903.481 905.95 So suppose uh
905.95 908.613 you take- take the sentence uh they expect
908.613 910.944 to- they expect to meet each other
910.944 914.084 there's a relation between "they" and "each other"
914.084 916.711 uh "who do they expect to meet each other"
916.711 919.159 No relation between "they" and "each other".
919.159 921.723 All right, ((there's a)) conclusion from that, it uh
921.723 926.022 it tells you that binding theory relies on uh
926.022 928.518 quite uh
928.518 931.136 complex notions of uh
931.136 934.057 non-trivial notions about uh
934.057 939.618 empty elements that uh unpronounced elements that are interpreted ((and matter)) in the binding theory
939.618 942.883 Well yeah in a case like that the facts are so obvious
942.883 947.465 that it would be a waste of time to do an experiment, and again if you did an experiment it would
947.465 949.148 be a test of the experiment
949.148 951.298 But there are less trivial cases
951.298 954.341 and it's perfectly reasonable to uh
954.341 958.985 call for more serious experimentation, actually some of the cases were discussed this morning.
958.985 961.049 But the emphasis is on "serious".
961.049 964.276 If you do unserious experiments, you get noise.
964.276 966.905 uh
966.905 969.745 useless, you know, experiments are like theories,
969.745 972.448 they have to be a- adjusted so that they're
972.448 974.719 giving the right empri- empirical results,
974.719 979.593 and once you think they're giving the right empirical results you can use them for other purposes.
979.593 980.708 that's uh
980.708 982.574 often been misunderstood,
982.574 986.244 but to give a general argument against uh
986.244 987.602 abstraction and
987.602 991.92 idealization is just to say let's abandon rational inquiry
991.92 998.108 Well, actually, uh- and this critique of Enfield's is quite widespread in the cognitive sciences
998.108 1000.561 and even uh seeping into linguistics
1000.561 1002.896 so there's this current
1002.896 1007.428 issue of uh the journal Theoretical Linguistics, coming out soon, which uh
1007.428 1010.689 ((in)) which quite serious prominent linguists uh
1010.689 1012.473 endorse this critique
1012.473 1015.775 uh they say linguists are making a mistake by
1015.775 1020.211 uh by paying a- uh by uh uh adopting
1020.211 1022.888 uh abstract and idealization
1022.888 1023.943 uh
1023.943 1029.533 something that isn't even discussed in the natural sciences because it's considered to be ((a)) truism
1029.533 1035.653 uh in this case they- there're particular things they object to, one is, an interesting case, is
1035.653 1036.931 uh work on
1036.931 1038.088 embedding
1038.088 1039.326 uh goes back
1039.326 1040.502 fifty years
1040.502 1043.889 uh which abstracts from memory restrictions
1043.889 1045.978 uh we're all familiar with this
1045.978 1049.528 and they say well that's misleading, you've got to
1049.528 1051.96 consider the memory restriction- you have to
1051.96 1055.043 build your theories to include memory restrictions
1055.043 1057.863 Now that makes about as much sense as saying that
1057.863 1060.744 if you study arithmetical capacity
1060.744 1062.285 you have to
1062.285 1065.388 uh to try to figure out how people know how to multiply
1065.388 1068.836 you have to- you can't abstract from the fact that uh
1068.836 1071.443 they can only do so much in their heads
1071.443 1075.844 You can't pay attention to the fact that if you give them paper and pencil they can go on
1075.844 1079.029 up to the limits of patience or lifetime
1079.029 1080.266 That would be
1080.266 1085.515 abstraction and idealization. I- I don't think anyone's crazy enough to suggest that.
1085.515 1089.309 But in the case of language, it's a pretty widespread critique
1089.309 1092.312 in computational cognitive science too.
1092.312 1098.77 Well I mentioned two specific examples, structural- structure-dependence and binding,
1098.77 1104.929 and the reason I picked those is because they're among the very few cases where there has at least been an effort
1104.929 1109.712 uh to demonstrate that language is a constellation of cognitive processes
1109.712 1111.732 that apply elsewhere.
1111.732 1118.364 ((so this is some kind-) what people know is in fact uh just uh
1118.364 1123.086 based on statistical of a corpus of unanalyzed data
1123.086 1127.276 ((you know)) roughly like finding the laws of motion from
1127.276 1129.403 watching leaves falling or
1129.403 1135.381 replacing bee science by video tapes of bees swarming, and so on.
1135.381 1136.394 uh
1136.394 1140.735 Well, in the case of structure dependence, it's actually become an industry;
1140.735 1143.25 in computational cognitive science
1143.25 1145.686 there are a raft of papers about it
1145.686 1148.428 they're kind of interesting. And one
1148.428 1151.206 property of them, pretty easily demonstrated
1151.206 1152.848 is that every
1152.848 1154.064 proposal
1154.064 1157.331 not only fails, but fails massively
1157.331 1161.466 but in fact is regarded as a success -- I'll come back to that in a moment --
1161.466 1164.798 uh and uh furthermore it wouldn't matter if they worked
1164.798 1168.573 because if the same propos- the same proposals
1168.573 1169.668 would work
1169.668 1171.074 for a
1171.074 1171.946 you know
1171.946 1175.089 for a hypothetical language, actually non-language
1175.089 1178.492 that relied on linear rather than structural distance
1178.492 1182.566 if you- in analysis of those, you'd find ((what do you get)) those patterns
1182.566 1185.339 so it wouldn't matter even if they worked, and they don't work
1185.339 1186.965 uh
1186.965 1193.562 but that doesn't seem to bother anybody, in fact, as I've mentioned, it's regarded as a success, and I'll come back to that
1193.562 1195.914 uh in the case of binding theory
1195.914 1202.197 the trivial example that I mentioned actually suffices to refute a proposal in the
1202.197 1205.156 current issue of the journal Cognitive Science
1205.156 1210.568 which is unusual in that it at least tries to deal with some property of language
1210.568 1211.868 uh but
1211.868 1214.975 it claims that uh
1214.975 1217.55 the binding principles rely on a
1217.55 1220.798 a simple general cognitive process
1220.798 1227.995 whether it's a cognitive process I don't know, it's just asserted, but essentially when you look at it, it says find the closest antecedent
1227.995 1233.616 uh which is exactly kind of a crude restatement of the principle that they're trying to refute
1233.616 1242.076 uh but in order to apply that cognitive principle, as this and numerous other examples show, it's necessary to introduce the
1242.076 1246.043 syntactic mechanisms in this case that yield the missing element
1246.043 1251.279 and if you proceed to introduce uh a great deal more of the
1251.279 1252.15 uh
1252.15 1257.608 sytactic uh operations that- the syntactic mechanisms that they
1257.608 1259.919 they claim to be eliminatable in favor of
1259.919 1261.684 general cognitive processes
1261.684 1264.971 well I don't think these examples are
1264.971 1272.449 uh badly chosen, I think they're among the few cases where there's even been an attempt to deal with some real property of language
1272.449 1276.929 and as I said, it's it's it's some- it's of kind of some interest that these
1276.929 1279.539 consistent failures are
1279.539 1281.769 cited as successes
1281.769 1286.214 in the uh literature of computational cognitive science
1286.214 1288.341 and the difference of judgment,
1288.341 1290.592 it has to do with the notion of success
1290.592 1293.814 There's a new notion of success that's entered
1293.814 1297.668 that I don't think existed in the sciences before, so
1297.668 1300.268 go back to the falling leaves and uh
1300.268 1302.477 you know, the bees swarming
1302.477 1305.758 I mean again, if you wanted to approximate
1305.758 1308.797 unanalyzed observations
1308.797 1313.838 statistical analysis of massive data will give much better predictions than-
1313.838 1318.058 of what's happened outside the window, or you know, bees swarming than uh
1318.058 1319.985 than scientists could give.
1319.985 1321.426 Much better.
1321.426 1325.98 Uh but nobody suggests that the scientists go out of business for that reason
1325.98 1334.113 you know what's taken for granted is you don't do that work, nobody cares whether you can approximate unanalyzed data
1334.113 1337.883 uh but in this field uh you do, that's
1337.883 1340.603 success, if you can approximiz-
1340.603 1344.162 approximate you know something in the corpus, or unanalyzed data
1344.162 1345.015 without
1345.015 1349.052 analyzing out various factors that enter into it, say the uh
1349.052 1352.342 memory restrictions on embedding, or anything else.
1352.342 1355.405 It's also kind of interesting that in the many
1355.405 1362.367 the rather triumphalistic reviews, you can read them in journals like Topics in Cognitive Science and so on
1362.367 1365.734 a lot of triumphalist reviews of the successes of
1365.734 1370.038 computational cognitive science which treat the failures as successes
1370.038 1371.945 I- I've never seen
1371.945 1373.632 a mention of the
1373.632 1375.317 the authentic successes
1375.317 1377.978 for example Charles's work on
1377.978 1380.412 how uh UG principles
1380.412 1382.72 uh and an independent
1382.72 1385.824 probabilistic learning theory that has
1385.824 1388.506 independent justification, how they interact
1388.506 1392.329 to yield real results on the language that's acquired
1392.329 1395.493 As far as I know those things like that are ignored
1395.493 1398.342 And I guess that the neglect uh
1398.342 1400.554 traces back to uh
1400.554 1402.44 the fact that uh
1402.44 1408.769 (()) Charles's work is guilty of the crime that uh Enfield and others are
1408.769 1409.783 trying to
1409.783 1413.008 get rid of, namely assuming that language exists
1413.008 1415.016 and is an object of study
1415.016 1418.392 of course by abstraction and idealization
1418.392 1419.386 ((like)) everything else.
1419.386 1421.09 Well there's another
1421.09 1422.57 influential
1422.57 1424.246 idea that uh
1424.246 1429.606 uh lead to essentially the same conclusion, that language doesn't really exist
1429.606 1431.046 uh it's uh
1431.046 1432.831 so there's nothing to study
1432.831 1434.981 it's illustrated in a
1434.981 1437.212 the entry on language
1437.212 1440.945 in a current Handbook of Child Development
1440.945 1443.526 This one happens to be by Michael Tomasello
1443.526 1446.937 In his view, language is just -- I'm quoting --
1446.937 1450.427 "a structured inventory of linguistic constructions"
1450.427 1452.03 it's structured
1452.03 1454.244 only in that it is acquired
1454.244 1458.97 by processes of pattern finding, schematization
1458.97 1462.793 uh abstraction, common to all primates
1462.793 1467.612 and a few other general processes, all of them left quite obscure.
1467.612 1472.691 So from this point of view there's a focus construction, and there's a passive construction
1472.691 1475.792 and the child just learn them kind of like a
1475.792 1477.158 big inventory
1477.158 1480.444 and it would be a mistake to try to uh
1480.444 1481.316 uh
1481.316 1483.633 pick out of them
1483.633 1487.641 to see if you can identify in them some processes of much more
1487.641 1489.802 uh general ap- uh
1489.802 1494.051 s- scope that uh apply here and elsewhere, like uh
1494.051 1497.419 you know that ((maybe the)) displacement conditions and so on
1497.419 1499.921 island conditions, you shouldn't look at that.
1499.921 1504.571 They just learn the constructions one by one, kind of a big library
1504.571 1505.828 uh
1505.828 1510.279 These approaches, which are again pretty widespread, even in linguistics
1510.279 1511.232 uh
1511.232 1515.488 invariably ignore the simplest questions,
1515.488 1518.917 for example, two that I mentioned, structure dependence and
1518.917 1523.574 uh binding-theoretic problems when there's an unpronounced element
1523.574 1527.83 or else they provide what are called solutions that just beg the questions
1527.83 1531.707 and these ((approaches)) too, if you look at them, are kind of like the Enfield
1531.707 1533.555 paper, they're concerned with
1533.555 1534.873 communication
1534.873 1537.009 and hence, sensibly,
1537.009 1538.876 the social context
1538.876 1541.25 you can't study communication without that
1541.25 1547.553 uh as is communication among organisms generally, from the bacteria up to humans
1547.553 1551.045 well, as I said, I half suspect that Enfield
1551.045 1563.0 understates his conclusion, that these views are accepted by many researchers in the cognitive sciences, from my own reading, I suspect that "many" should be changed to "most"
1563.0 1565.902 uh and the uh growing number
1565.902 1568.709 and also if these tendencies persist
1568.709 1571.175 the study of cognitive phenomena
1571.175 1574.69 I think is going to depart
1574.69 1580.005 even more from biology and the natural sciences, in a rather strange way
1580.005 1582.824 and that efforts to keep to the norms
1582.824 1585.512 of the sciences just
1585.512 1588.515 uh will appear to be some kind of quirky cult.
1588.515 1594.483 Well there are more dangerous developments in the world, but within the domains that
1594.483 1598.499 concern us, that seems to me a rather unfortunate one.
1598.499 1605.514 uh so to go back to the question that (()) raised in the title, I think we should study language, but
1605.514 1607.81 by the usual methods of the sciences
1607.81 1612.641 and if so, ((from)) the second part, how should we proceed, ((I))
1612.641 1616.698 have to become a little more technical here, and presuppose quite a lot
1616.698 1620.187 and also to skip quite a lot, which I assume is
1620.187 1621.568 familiar to
1621.568 1623.758 many of you if not most.
1623.758 1625.159 Uh
1625.159 1626.477 I'll try to keep it
1626.477 1628.727 reasonably brief.
1628.727 1631.65 Well by "language" here
1631.65 1636.123 uh uh I mean what's called "internal language", i-language.
1636.123 1638.64 an internal property of an individual
1638.64 1640.022 I think that
1640.022 1645.765 as far as I can see every approach to language, whether it's sociolinguistic, or you know
1645.765 1648.307 study of communication, whatever it is,
1648.307 1653.462 presupposes that in some form, usually tacitly given those who deny it
1653.462 1656.375 So, some property of an individual
1656.375 1659.349 uh the core principle of it of course is
1659.349 1662.506 discrete infinity, that's why you have to have
1662.506 1664.493 what are called saltationist
1664.493 1666.196 evolutionary accounts
1666.196 1669.537 you can't get away with them, any more than for arithmetic
1669.537 1670.368 uh
1670.368 1675.179 which means that uh a language, an internal language, an i-language
1675.179 1679.713 is- we can basically think of as being a generative procedure
1679.713 1688.873 some procedure that yields an infinite uh hierarchy of uh structured expressions- an infinite array of structured expressions
1688.873 1690.447 uh which have to have
1690.447 1692.84 uh uh several properties
1692.84 1695.805 for one thing it has to be unbounded, for another it has to be structured
1695.805 1700.329 uh hierarchically structured is has to uh
1700.329 1702.013 yield information
1702.013 1706.572 that's used by at least two systems that make use of language --
1706.572 1714.052 interfaces, they're called -- one is uh the sensory-motor system and the other is um systems of thought, sometimes called uh
1714.052 1716.736 conceptual intentional systems
1716.736 1718.642 ((so)) that's essentially it
1718.642 1721.783 refomulation of uh Aristotle's
1721.783 1722.823 uh
1722.823 1726.843 slogan that uh language is sound with a meaning
1726.843 1729.523 notice that the "with" part of that has
1729.523 1732.349 rarely been studied until quite recently
1732.349 1735.86 uh but we want to study the "with" part too.
1735.86 1740.079 Uh so uh how do we proceed, well we should
1740.079 1742.371 proceed in the way of uh
1742.371 1745.86 the investigation of any biological organ.
1745.86 1750.836 uh for one thing we should regard this generative procedure, we should study it
1750.836 1753.595 in technical terms, in intension
1753.595 1755.003 intension with an 's'
1755.003 1758.043 meaning what the algorithm actually is
1758.043 1765.091 and not just s- the set of things that it characterizes, that's a fundamental difference often misunderstood
1765.091 1777.823 uh and uh at least that's what we should be doing if we want it to be part of biology, or studying it as uh some kind of what's informally called an organ, a subcomponent of a
1777.823 1781.575 more complex system that has enough internal integrity so that
1781.575 1783.461 you can investigate it
1783.461 1787.688 in abstraction, you can look into its interactions with others in the
1787.688 1789.19 life of the organism.
1789.19 1791.199 uh this work
1791.199 1793.888 began to take off in the 1950s
1793.888 1799.131 it was part of what's been called the "cognitive revolution" -- I don't think that's a
1799.131 1802.217 very good term but uh that's what it's called
1802.217 1806.299 uh it departed pretty sharply from prevailing
1806.299 1809.771 uh structuralist and behavioralist conceptions
1809.771 1814.463 which are now incidentally being renewed, interestingly that's the first
1814.463 1816.747 approach that I mentioned
1816.747 1823.427 uh and it uh beg- turned to the procedures of basically normal- normal biology
1823.427 1830.414 uh well if uh we pursue that we ask how the system say develops in the individual
1830.414 1834.834 uh there's going to be at least three factors involved, like every
1834.834 1835.808 uh
1835.808 1837.493 organic system
1837.493 1839.807 uh one factor is uh
1839.807 1841.206 uh whatever
1841.206 1843.722 uh genetic information there is
1843.722 1845.488 now there has to be some or
1845.488 1846.693 you can't do anything
1846.693 1849.129 uh so for example if uh
1849.129 1853.959 uh say a- a bird or a
1853.959 1856.899 or a kitten or a chimpanzee or whatever is-
1856.899 1859.648 gets exactly the same experience from
1859.648 1861.415 birth, or even before birth
1861.415 1863.1 as my granddaughter
1863.1 1865.357 they'll get nothing out of it
1865.357 1868.078 as far as language is concerned
1868.078 1872.636 but my granddaughter in fact even before birth as is now known
1872.636 1876.298 is already picking out of it language-related material
1876.298 1878.41 and very quickly uh
1878.41 1881.011 developing the capacities that we're all using
1881.011 1887.692 Either that's a miracle or it's some kind of genetic information crucially involved, as in the case of everything else.
1887.692 1889.316 So that's one factor.
1889.316 1892.722 Actually there we can divide it into two components
1892.722 1896.561 uh going back to Enfield and others, we can ask
1896.561 1897.843 to what extent
1897.843 1902.571 do these uh genetic factors are- to what extent are they just general cognitive
1902.571 1903.768 capacities
1903.768 1905.149 shared by
1905.149 1906.996 you know birds
1906.996 1908.702 chimpanzees and so on
1908.702 1910.447 and to what extent are they
1910.447 1912.862 uh human and language specific.
1912.862 1914.1 Surely
1914.1 1918.309 it has to be some of the second, or else what I just described would be a miracle.
1918.309 1923.492 but that's a important conceptual distinction, and work like say Charles Yang's uh
1923.492 1929.591 uh points the way to trying to isolate them, of course by abstraction and idealization
1929.591 1933.632 So that's one, the first, the language-specific ones are
1933.632 1937.456 what's called uh universal grammar, U.G.
1937.456 1941.049 Well that's uh the first uh the second
1941.049 1945.381 element, the second factor, is whatever data there are, so
1945.381 1949.06 if you end up speaking Swahili you have different data from English
1949.06 1951.315 and that obviously has an effect
1951.315 1957.416 and the third, which has got to be there, and is all over the theory of development and evolution,
1957.416 1962.799 but really hasn't been studied much until recent years because it's just too hard to figure out
1962.799 1965.558 uh is uh just general laws of nature
1965.558 1970.762 uh how do general laws of nature shape the course of development and evolution
1970.762 1974.531 and by now in the biological sciences, there is a
1974.531 1976.114 relatively new field
1976.114 1977.108 uh
1977.108 1980.049 ev dev, it's called, which deals with
1980.049 1982.009 the role of the
1982.009 1982.926 uh
1982.926 1984.311 general
1984.311 1987.314 biological or maybe just physical laws
1987.314 1988.755 that enter into
1988.755 1994.348 shaping the course of evolution and shaping the course of development -- not a new idea, people like
1994.348 1999.924 Alan Turing, ((Bert Gonnick)) and others, but it's kind of taken off in the last thirty forty years
1999.924 2001.853 and roughly at the same
2001.853 2007.727 during the same period when these considerations have been studied more carefully in the case of language
2007.727 2008.905 uh
2008.905 2010.631 it's uh
2010.631 2014.918 it's sometimes called the "minimalist program" in language, but it's a continuation
2014.918 2019.916 of normal research in language, shouldn't have a separate name, that's confused people
2019.916 2022.17 And in fact it's just ordinary science
2022.17 2023.693 Yeah, you want to figure out
2023.693 2029.455 how uh general principles enter into say the acquisition of language, principles that are
2029.455 2033.892 uh not language-specific, not cognition-specific, uh
2033.892 2034.656 uh just
2034.656 2038.881 have to do generally with the way things work in the physical world.
2038.881 2042.616 Now language is a computational system, unquestionably,
2042.616 2044.099 at least at its core
2044.099 2050.958 so the obvious place to look is uh principles of computation, in particular principles of computational efficiency
2050.958 2053.563 uh should be that uh
2053.563 2056.699 language like other systems we know
2056.699 2059.826 uh is based on minimal computation
2059.826 2061.308 do as little as possible
2061.308 2067.221 uh but we don't have a complete theory of minimal computation by any means, there's a lot to be learned, but uh
2067.221 2070.572 even very simple ideas like say
2070.572 2072.277 less is better than more
2072.277 2074.538 uh carry you quite far
2074.538 2076.204 uh so that's
2076.204 2079.193 there the- there're three factors
2079.193 2081.549 uh genetic endowment, uh
2081.549 2083.721 external data, uh
2083.721 2088.195 uh general principles ((there factor principle)) and there're others
2088.195 2094.92 uh if we ever- if we knew enough- enough about them, we might be able to introduce them, like the uh the structure of the brain
2094.92 2100.431 uh whatever it is surely imposes constraints on how language could develop, but
2100.431 2106.856 and someday it may be- maybe enough will be known to introduce the effect of that factor, but for the moment, it's
2106.856 2108.094 kind of a dream
2108.094 2111.075 So but at least these three factors, uh
2111.075 2113.248 we can investigate
2113.248 2114.307 uh
2114.307 2115.346 well the
2115.346 2118.411 early approaches to this, back in the fifties
2118.411 2121.714 uh pretty much borrowed from traditional grammar
2121.714 2126.282 so one core part was phrase structure grammar, has traditional origins,
2126.282 2127.948 uh and uh
2127.948 2135.502 it was very quickly abandoned, by the nineteen sixties phrase structure grammar was abandoned, pretty much, for good reasons
2135.502 2139.096 For one thing it was much too strong; for another it was too weak.
2139.096 2143.24 It was too strong because it was based on massive stipulations
2143.24 2149.66 and so for example should you have a rule uh S arrow NP VP, and not a rule
2149.66 2153.928 uh S arrow P uh NP, or something
2153.928 2156.181 ((and)) all of that is just pure stipulation
2156.181 2160.427 And in fact some of those stipulations have survived, unnoticed
2160.427 2161.706 right to the present
2161.706 2167.239 (( )) mentioned some of them, and they're pretty lethal when you look at them carefully.
2167.239 2169.086 So it's worth looking at.
2169.086 2170.162 uh
2170.162 2174.933 So it was too strong, had the- that had to be abandoned. It was also too weak
2174.933 2178.811 It had no plausible way to deal with uh
2178.811 2181.349 relationships that are at a distance
2181.349 2183.563 like displacement phenomena
2183.563 2184.293 uh
2184.293 2188.181 structure-dependent rules, for example, I mean you can make up complicated
2188.181 2193.381 rules to deal with it, but they don't make any sense, they're not capturing the phenomenon properly
2193.381 2195.391 or- or just agreement, for that matter
2195.391 2197.807 So it's both too strong and too weak
2197.807 2203.392 uh by the nineteen sixties, it was pretty much abandoned, in favor of X bar theory
2203.392 2208.97 uh it was supplemented by principles that dealt- that tried to deal with uh
2208.97 2212.958 remote relationships called transformational grammar
2212.958 2216.287 uh and in the uh through the following years, I won't trace it,
2216.287 2220.69 uh the major work, the major theoretical work, has been to
2220.69 2222.478 try to pare down
2222.478 2225.263 the initially very rich assumptions
2225.263 2227.252 about the nature of these systems
2227.252 2229.179 to try to show that they were
2229.179 2234.66 ((related to different)) unacceptable stipulations that could be replaced by
2234.66 2238.213 broader generalizations, deeper principles, and so on
2238.213 2240.121 Now there're two reasons to do this.
2240.121 2242.327 uh one reason is it's just
2242.327 2252.645 if you want explanations, you have to get rid of stipulations, every time you stipulate something, you're saying I can't explain it
2252.645 2255.532 (( )) (( )) something
2255.532 2262.263 So you want to get rid of stipulat- you want to identify stipulations, which is not so simple, they're hard to see sometimes
2262.263 2265.108 as I mentioned, some of the stipulations of
2265.108 2267.159 phrase structure grammar are still
2267.159 2271.911 there in current work, and still unnoticed, but you want to get rid of them
2271.911 2275.651 for the normal reasons of rational inquiry
2275.651 2279.237 uh there's a second special reason in the case of language
2279.237 2282.397 uh ultimately out there on the horizon
2282.397 2287.536 there is a question which maybe can be addressed and maybe can't be addressed
2287.536 2290.258 and that's the question of evolution of language
2290.258 2294.268 Now there are some first-rate evolutionary biologists
2294.268 2298.254 who have argued pretty persuasively that you should give up
2298.254 2302.115 that there's never going to be an explanation for the evolution of cognition
2302.115 2304.1 altogether ((let alone)) language
2304.1 2308.855 because of the limitation of the tools that are available for inquiry
2308.855 2311.596 and of the kind of evidence that's available
2311.596 2316.306 uh the most uh important case is uh Richard Lewontin, a very
2316.306 2319.6 outstanding evolutionary biologist, who's been
2319.6 2323.641 trying to convince cognitive scientists of this for years
2323.641 2329.846 nobody ever looks at his articles, and nobody ever refutes the articles, but I think they should be
2329.846 2331.551 ((taken)) pretty seriously
2331.551 2333.686 and I think maybe he's
2333.686 2335.772 a little too strong but
2335.772 2338.453 not insignificant arguments
2338.453 2342.173 anyway off there on the horizon there's that possible field
2342.173 2344.448 and uh any
2344.448 2350.067 stipulation, anything that's in universal grammar, anything that's assumed about the nature of language
2350.067 2354.151 poses a further difficulty to this eventual field
2354.151 2356.222 uh and that's
2356.222 2357.927 something to keep in mind
2357.927 2363.147 uh maybe we can do no better, but uh we should remember what we're doing
2363.147 2365.41 Well, phrase structure grammar
2365.41 2367.766 uh when it sort of broke down
2367.766 2372.09 as it should have for the reasons I mentioned, say by the nineteen sixties
2372.09 2375.931 it- it's kind of disentangles three different issues
2375.931 2382.949 now it's sometimes argued that they shouldn't be disentangled and that's a very controversial question right up till today, but at least
2382.949 2385.63 conceptually we can distinguish three parts
2385.63 2387.865 uh one is constituency
2387.865 2389.927 and the second is order
2389.927 2391.778 and the third is
2391.778 2393.768 what's called projection
2393.768 2400.141 how do you know that some unit is featured as a nominal rather than a verbal element, say
2400.141 2401.664 So those are three
2401.664 2406.627 uh components of phrase structure grammar which should be taken apart
2406.627 2409.593 uh though as I said it's controversial
2409.593 2412.13 uh with regard to order
2412.13 2414.099 uh there is a general
2414.099 2418.034 this- it's very interesting, the set of considerations here
2418.034 2420.957 just on pure conceptual grounds
2420.957 2423.414 we would expect- I think we would expect
2423.414 2424.428 order
2424.428 2426.479 to be one of the properties
2426.479 2430.864 that is involved with mapping to the sensory motor system
2430.864 2434.118 uh the sensory-motor system plainly requires order
2434.118 2440.778 uh or other- depending on the modality you're using it requires some external constraints on
2440.778 2444.175 what you can do, it's different for sign and sound, let's say
2444.175 2446.675 but some constraints are coming from the
2446.675 2448.888 sensory motor system, you can't
2448.888 2451.405 talk in parallel, so you've got to
2451.405 2454.55 order things, and other kind of constraints
2454.55 2457.372 so that's one reason to expect
2457.372 2460.618 what's a nice word externalization, mapping of
2460.618 2464.54 the internal system to the sensory-motor system to regard
2464.54 2469.05 order as a property of that system, not of the internal system itself
2469.05 2473.296 and a second complementary part of that is that it appears that
2473.296 2476.058 for core- the core semantics
2476.058 2478.263 uh order doesn't matter
2478.263 2480.417 what matters is hierarchy
2480.417 2481.979 not order
2481.979 2484.759 and so those two considerations
2484.759 2487.339 lead to the expectation
2487.339 2489.248 that order ought to be a
2489.248 2491.177 peripheral property of language
2491.177 2494.222 a property of externalization to the sensory-motor system
2494.222 2497.289 shouldn't be part of the internal system of
2497.289 2500.03 syntax and semantics and pragmatics
2500.03 2501.39 formal pragmatics
2501.39 2503.909 uh well there's plenty of evidence against that
2503.909 2507.765 uh so - empirical evidence - so we're faced with a
2507.765 2510.163 standard problem in the sciences, there's
2510.163 2513.148 conceptual reasons to expect something
2513.148 2516.54 and there's empirical evidence that says it doesn't work
2516.54 2521.31 well how you go on from there depends on your research intuitions
2521.31 2522.162 uh
2522.162 2523.784 scientists
2523.784 2526.765 go all different ways but ((that)) you make up your own mind then
2526.765 2531.433 anyhow that's uh an issue to be addressed, that was actually discussed this morning
2531.433 2532.447 indirectly
2532.447 2535.209 Now that's order. What about uh
2535.209 2536.468 constituency?
2536.468 2538.763 Well the simplest
2538.763 2540.386 answer to how to
2540.386 2540.732 uh account for constituency
2540.732 2546.712 uh account for constituency, and of course you always pick the simplest one until it's proven wrong
2546.712 2548.416 is that there's some
2548.416 2554.224 elementary combinatorial principle which is embedded somewhere in every generative procedure
2554.224 2556.317 that just says if you've constructed
2556.317 2558.977 two elements, call them X and Y,
2558.977 2560.723 uh make up a new element
2560.723 2563.201 call it Z out of X and Y
2563.201 2568.908 And in the simplest case, that is, meeting the general condition of minimal computation
2568.908 2571.773 you won't change X and Y in the process
2571.773 2572.849 that's
2572.849 2575.429 the best possible case, so what you pick
2575.429 2579.36 uh you know absent uh contrary evidence
2579.36 2583.432 uh that means- and- and for the reason I mentioned, you won't include order
2583.432 2585.192 at least in the simplest case
2585.192 2587.413 well that simply means that this
2587.413 2589.952 operation which is now called "merge"
2589.952 2595.559 just takes X and Y and forms a set X Y, no order, no change
2595.559 2599.197 uh that's the simplest form of constituency
2599.197 2600.972 well that uh
2600.972 2603.125 though it wasn't realized at first
2603.125 2605.774 been understood for over a decade, that
2605.774 2607.338 gives you a ways of
2607.338 2609.227 carrying forward the
2609.227 2614.677 task of simplifying the two components of grammar, the descendents of
2614.677 2617.196 phrase structure and transformational grammar
2617.196 2620.04 uh carrying it to the next step, namely unifying them
2620.04 2622.617 uh obviously the next step
2622.617 2624.282 so if you just have
2624.282 2628.849 merge uh just as a matter of logic there's going to be two kinds
2628.849 2631.393 uh one is called external merge
2631.393 2633.322 you pick X and Y which are
2633.322 2635.477 distinct from one another and uh
2635.477 2637.306 make up the set X Y
2637.306 2638.463 external merge
2638.463 2641.103 uh which is what you get if you say uh take
2641.103 2644.607 uh see and uh the man and you construct
2644.607 2647.086 the abstract form of "see the man"
2647.086 2648.385 but there's another kind
2648.385 2651.899 uh one of them, let's say Y, could be inside the other
2651.899 2654.662 could be a term of the other, in the technical sense
2654.662 2658.763 and if you then merge it you get displacement
2658.763 2660.71 so if you have uh
2660.71 2661.732 ((...))
2661.732 2665.976 "you saw what" and you merge "what" you get "what you saw what"
2665.976 2672.315 notice that you get the copy theory of movement, because you're not supposed to modify X and Y
2672.315 2673.473 uh
2673.473 2675.098 and in fact
2675.098 2678.532 that yields the major displacement phenomenon
2678.532 2682.574 uh it yields many copies if you look at a
2682.574 2684.422 complicated structure
2684.422 2688.078 so you do unify the two
2688.078 2689.114 major
2689.114 2691.247 aspects of generation
2691.247 2694.091 in terms of the simplest possible operation
2694.091 2696.197 now it's possible to
2696.197 2699.612 stipulate that this operation's not allowed
2699.612 2701.817 uh that's a stipulation
2701.817 2706.089 that enriches universal grammar to stipulate that this is barred
2706.089 2710.484 and if you do stipulate it you have to make up other mechanisms to yield
2710.484 2715.278 you know the same empirical consequences so it's a double stipulation
2715.278 2716.212 uh
2716.212 2718.426 and uh as far as I know that's
2718.426 2723.739 never been addressed by- alternatives they kind of skirted, but it's kind of important
2723.739 2725.396 uh think how the
2725.396 2728.751 the simplest approach, the approach you take unless there's
2728.751 2730.006 evidence against it
2730.006 2734.426 uh is the- unified and consistent ((it's just)) one system
2734.426 2735.886 in terms of the
2735.886 2740.108 simplest operation, which has to be embedded in any generative procedure
2740.108 2743.678 uh namely just simple merge in the simplest form
2743.678 2745.786 all right back to
2745.786 2750.246 uh looks like the best answer to- for constituency, it has
2750.246 2753.312 direct consequences which are worth considering
2753.312 2754.326 uh
2754.326 2758.222 the form, say, you know, "what do you see what"
2758.222 2762.568 uh that happens to be the right form for the conceptu- for the semantic system
2762.568 2764.253 for the thought system that
2764.253 2765.874 gives you the right interpretation
2765.874 2767.131 and if
2767.131 2770.194 the ((pressures)) are more complex
2770.194 2775.972 as is well known it gives you a lot of what are called reconstruction {cough} spots
2775.972 2781.284 which is the right form for more intricate semantic interpretations
2781.284 2785.705 and on the other hand its completely the wrong form for the sensory-motor system
2785.705 2788.393 you don't say "what do you see what"
2788.393 2792.89 you know "which book did you see which book" (( )) nothing like that in any language
2792.89 2793.761 there's some
2793.761 2795.647 marginal residues
2795.647 2796.778 uh
2796.778 2800.949 which we could talk about if there's time, but there's just nothing like spelling out the copies
2800.949 2802.375 uh
2802.375 2806.74 why? well cause there's another principle of minimal computation
2806.74 2810.249 which says that you should articulate as little as possible
2810.249 2811.402 uh
2811.402 2814.16 articulation is an expensive operation
2814.16 2817.363 lot of the brain is involved in it
2817.363 2819.736 uh so you want to do as little as possible
2819.736 2820.77 well
2820.77 2826.122 "as little as possible" means you erase all the copies except for the hierarachically highest one
2826.122 2829.549 cause when you erase that one, there's no evidence that anything happened
2829.549 2830.988 so you've got to
2830.988 2835.449 for interpretation, you've got to have at least the
2835.449 2838.364 uh the highest one hierarchically
2838.364 2839.296 most
2839.296 2840.594 prominent one
2840.594 2842.297 but you eliminate all the others
2842.297 2846.047 Well so far that's just following the principle of minimal computation
2846.047 2847.893 It has some consequences
2847.893 2853.969 what is could- means is that language is very well designed in its nature just for
2853.969 2855.654 computational reasons
2855.654 2860.363 where "design" is kind of a metaphor, but designed for thought
2860.363 2863.181 but very poorly designed for communication
2863.181 2867.094 in fact anyone who's worked on a parsing program knows this
2867.094 2872.444 uh one of the hard problems in parsing programs is called ((the)) filler-gap problem
2872.444 2876.791 When you get- hear ((with)) what at the beginning of the sentence you've got to figure out what it is
2876.791 2878.534 which can be a very tricky
2878.534 2883.421 operation, particularly if you're trying to find all the positions where it isn't pronounced
2883.421 2887.496 gave one example before, they get much more intricate
2887.496 2888.895 That's a hard problem
2888.895 2892.882 uh it's not just for building parsers, it's also for perception
2892.882 2895.902 uh that would be overcome if you pronounced all the copies
2895.902 2901.175 but you know so minimal computation and communicative efficiency
2901.175 2902.317 conflict
2902.317 2905.522 and minimal computation wins hands down
2905.522 2907.022 overwhelmingly
2907.022 2911.324 and not just in this case but in every other case that's known, and there are quite a few
2911.324 2915.812 uh ambiguities, uh garden paths, uh islands, uh
2915.812 2918.305 to the extent that we understand any of them
2918.305 2921.529 uh they all seem to involve minimal computation
2921.529 2922.767 but uh
2922.767 2925.061 impediments to communication
2925.061 2929.162 and it just from this massive considerations
2929.162 2931.25 it looks as if uh
2931.25 2936.724 uh not only that externalization altogether is a peripheral part of language
2936.724 2942.019 and communication in particular, which is just one aspect of externalization, is very peripheral
2942.019 2943.534 uh which means that
2943.534 2945.041 you know
2945.041 2947.648 substantial of the work done in
2947.648 2950.589 computational cognitive science, or uh
2950.589 2955.678 you know the speculations about the evolution of language, is just off on the wrong track
2955.678 2964.153 it's trying to look at the evolution of communication, or the properties of communication, which of course involves social context, but that's
2964.153 2971.601 it's a marginal part of language; (( )) everything we do, communication involves almost everything we do
2971.601 2973.169 hair style and
2973.169 2975.094 motor behavior, (()) so on
2975.094 2977.932 so of course language is involved, but it looks like a
2977.932 2981.945 peripheral part of it, just from the nature of the design itself
2981.945 2985.411 uh well that kind of suggests that we should uh
2985.411 2986.505 invert
2986.505 2989.504 uh Aristotle's maxim, language is
2989.504 2991.957 should not be regarded as sound with a meaning
2991.957 2994.43 but as meaning with a sound
2994.43 2996.01 and that uh
2996.01 3000.043 leads to quite uh different ways of studying it. Well
3000.043 3006.172 you know at this point the third thing we have to talk about is projection, I'll just say a word about it, but it's
3006.172 3009.938 gets interesting at this point ((but)) that's more complicated so I'll go on
3009.938 3013.365 uh projection is a serious problem
3013.365 3015.596 uh how do you decide
3015.596 3019.63 that one of these syntactic objects formed by merge
3019.63 3025.713 whether internal or external, or both of them, usually, how do you decide what it is?
3025.713 3027.173 I mean is it a
3027.173 3029.262 clause, is it a noun phrase
3029.262 3031.513 is it a prepositional phrase, and so on
3031.513 3036.041 Well that information is necessary for further computation
3036.041 3041.87 ((Ok)) you have to know that in order to compute further in the
3041.87 3044.385 syntax, in order to
3044.385 3045.911 uh interpret it
3045.911 3047.555 at the two interfaces
3047.555 3050.839 in fact it has to correlate at the two interfaces
3050.839 3053.929 so if something's going to be a noun phrase at the
3053.929 3058.085 thought interface it has to be a noun phrase at the sensory-motor interfaces
3058.085 3062.202 so somewhere along the uh in the internal system
3062.202 3067.591 before you hand information up into the interfaces uh these things have to be identified
3067.591 3073.535 uh and that identication optimally should be part of the computational procedures
3073.535 3075.009 uh
3075.009 3076.815 uh it shouldn't be
3076.815 3078.275 represented
3078.275 3082.328 it's my mistake fifteen years ago, it shouldn't be represented in the
3082.328 3090.499 syntactic objects themselves, the way I proposed years ago, but uh just should be part of the computational algorithm
3090.499 3091.877 which assigns
3091.877 3094.593 what's called labels to things
3094.593 3098.892 so that they can then be interpreted as the computation proceeds
3098.892 3101.912 uh well how do you assign labels?
3101.912 3106.416 uh again we'd like it to work by minimal computation
3106.416 3109.4 this general prin- third factor principle
3109.4 3113.101 ((that is)) non-linguistic, in fact probably non-biological
3113.101 3115.454 uh so how would that work?
3115.454 3118.085 well that would say you have to have a procedure which
3118.085 3120.967 searches the object you're investigating
3120.967 3122.184 and finds
3122.184 3124.436 as quickly as possible
3124.436 3125.754 an element
3125.754 3127.789 which is as small as possible
3127.789 3129.716 that gives you all the information
3129.716 3132.005 necessary to tell you what it is
3132.005 3134.021 and that's basically projection
3134.021 3140.679 so you want an algorithm which will find some designated element by minimal computation that'll identify it
3140.679 3143.113 Well often there's an easy answer
3143.113 3322.441 
3322.441 3325.403 uh anyhow just to summarize there are
3325.403 3331.22 uh there's a serious question about whether you should study language, it's not a frivolous question
3331.22 3337.411 uh and in a certain sense the answer is you should study it as little as possible
3337.411 3342.519 meaning whatever- we should try to show that whatever is specific to language
3342.519 3344.748 is as limited as possible
3344.748 3349.455 There're very good evolutionary reasons for believing that on the basis of what we know
3349.455 3350.894 and it's just part of science
3350.894 3355.049 if you want to get rid of stipulations, including the ones you haven't noticed until now
3355.049 3356.792 So yeah we ought to try to
3356.792 3359.549 show that there's as little as possible to study
3359.549 3366.198 But it seems very clear that there is something important to study, and you learn a lot about it when you study it hard
3366.198 3372.232 uh there's an alternative approach, which is increasingly sig- you know
3372.232 3378.262 I would say even dominant in the surrounding disciplines, cognitive science, computational cognitive science,
3378.262 3381.852 which applies the same conclusion, carries much further
3381.852 3383.496 so you just throw out everything
3383.496 3385.28 uh on the basis of
3385.28 3389.577 illusions, total illusions about evolution
3389.577 3391.948 uh misunderstandings of
3391.948 3393.874 the nature of language
3393.874 3395.414 uh and a
3395.414 3398.495 deep misunderstanding of the nature of science
3398.495 3402.571 namely the belief that science means trying to approximate
3402.571 3404.252 unanalyzed data
3404.252 3406.746 not trying to figure out why it works
3406.746 3410.192 as in say physics or bee communication or whatever
3410.192 3413.274 So there are these two answers to the question