Speech Transcript - Conversational Eliciture in a Bayesian Model of Language Interpretation

0:00:14	she could mining at b one next get static
0:00:18	i think we have i where i know keynote speaker an extremely glad to that
0:00:24	coming can
0:00:25	a and detail is a perfect setups linguistics i've invested californians ninety eight
0:00:32	and if you see that you like is that he had a broad the second
0:00:37	one thing that out succumbs
0:00:39	from at the expense of a
0:00:42	a professor of linguistics like is also then if at the sri international in the
0:00:47	past their arts and sun microsystems
0:00:51	and he is associated celtic general of logic and complication and he has and then
0:00:56	at the executive both
0:00:59	i think ideas and is what actually e focus the and it's "'cause" interpretation so
0:01:06	you guys a lot of computational modelling luck but also experiment a lot and which
0:01:12	can be is stealing and think that i is to listen relationship you're that idea
0:01:18	i'd and like to nominate
0:01:20	as the sink a pragmatic and delta psycholinguistic features of language
0:01:25	and some of the things he's not that trinkets trumpet on and it's this collecting
0:01:31	instruction on a events
0:01:33	and ten
0:01:35	and his book a coreference utterance and that the idea of grammar isolate and a
0:01:41	and b citation
0:01:43	and any maybe come up with a set of speaker is a who would broadly
0:01:49	at the cable i think i understand dialogue in an ideal is a very a
0:01:55	backchannel an ideal choice
0:01:57	and in have been talking to a bunch of people like yesterday and today before
0:02:02	and you come from a variety of backgrounds comments and makes it might makes that
0:02:08	the psycholinguistic at least at the end you have something to say tell a few
0:02:13	and i just like and take get from there
0:02:17	okay thank you mean and you
0:02:21	being we okay
0:02:23	alright
0:02:25	well thank you very much
0:02:28	for having you here and four
0:02:32	a median in this morning
0:02:35	so
0:02:37	so that
0:02:39	famously positive to competing desiderata and language design right one he called the auditors economy
0:02:47	right which is kind of biased towards here
0:02:51	right that languages should enable here's to get the speakers intended a message
0:02:57	with minimal interpreted in inferential after right so that
0:03:02	pushes mine which stores having more products the unless the ambiguity
0:03:06	is that what we would like
0:03:08	language to have when we're building system right we want the information right there where
0:03:12	we can grab
0:03:13	unfortunately for systems there's a competing does it around
0:03:17	which is the speakers a kind of which has the languages should allow speakers to
0:03:21	get
0:03:21	their message across with minimal articulatory effort
0:03:25	right so that pushes towards
0:03:27	less felix the and greater amounts of ambiguity
0:03:32	kind of the limit if you see variants of a galaxy
0:03:35	it's kind of the group language right good so always says i am group and
0:03:39	then everybody has been for well what it means by that
0:03:44	so
0:03:46	one way to speakers can be economical
0:03:48	i and still be expressive in getting them there'll a message across
0:03:53	it's a designer utterances would take advantage to be here's
0:03:56	cognitive apparatus mental state incapacity for inputs
0:04:00	so is to be able to convey more information than what they explicitly say
0:04:04	and this voices problem week constantly face when we're building discourse and dialogue systems because
0:04:10	the systems don't have that same apparatus thing capability that languages kind of wrapped itself
0:04:16	around
0:04:17	now of course the source of these pragmatically determine aspects of meaning
0:04:23	also been kind of the focus
0:04:26	in pragmatic since its birth and it's become an industry of its own since the
0:04:29	seminal work of rice
0:04:32	what i'm gonna focus on this part is a type of actually semantic enrichment that
0:04:39	i'll clean them fit neatly into any the other kind of enrichment of interest custom
0:04:44	the list linguistics and philosophy literatures
0:04:47	so let's illustrate by treating right in the some examples
0:04:51	like a jogger with it by far out about the last night
0:04:55	you're probably getting that the victim was india somebody who jobs
0:04:59	but was actually jogging at a time
0:05:02	right
0:05:03	the sentence doesn't entailed
0:05:05	right and you can see that by comparing with one be a farmer it was
0:05:08	hit by acquired how about the last night
0:05:11	it's for less inevitable where you get an inference that the victim was four and
0:05:15	half the time right in fact if you're knowing about how afterwards pretty unlikely even
0:05:20	though it could be that require veered off the road one so far in the
0:05:23	field of markup for guy also extractor right
0:05:27	you're probably not getting so that you don't need
0:05:31	to get that inference
0:05:32	in a case like one b would cause one to ask what why are you
0:05:36	getting it and one a
0:05:37	it's not limited to choice of a nominal
0:05:40	you get a with adjectives as well
0:05:42	so
0:05:43	the drug addled undergrad fell auditory pints clips
0:05:47	you probably getting
0:05:49	not only that
0:05:50	the victim follow the clips
0:05:52	and
0:05:53	was on drugs
0:05:55	but fell off the cliff
0:05:56	or because they were on drugs
0:05:59	but if you get you compare with to be the well liked undergrad about the
0:06:03	storyline squareds you're probably not saying ty why would being well like to call somebody
0:06:07	the fall off
0:06:11	into c
0:06:12	the normally with skippers undergrads of auditory by waves
0:06:15	you're probably getting kind of a contrary to expectation inference there wondering why somebody who's
0:06:21	risk of course would find themselves in such a document
0:06:25	finally you get it would relative clauses and referring expressions as well
0:06:29	so the company fire the manager who was embezzling money
0:06:33	again you probably getting narrowing that they were embezzling money they were they were fired
0:06:37	and embezzling money
0:06:39	but they're fired because they were embezzling money
0:06:43	you can compare that the three be the company fire commander whose tired in two
0:06:47	thousand two again doesn't send you off on the search for a while being hired
0:06:51	two thousand would cause one
0:06:53	to be fired
0:06:54	and
0:06:57	i then three c is another case of the a bilabial expectation kind of inference
0:07:01	right so
0:07:02	mean you think about a dialogue system
0:07:05	i be perfectly natural to respond the freebie
0:07:09	by saying y
0:07:11	right but it would be a little i to respond that way to three a
0:07:14	well that's a speaker was trying to convey the reason for the fire
0:07:19	use if you ask why haven't picked up on the inference that the speaker intended
0:07:22	to get across
0:07:24	so for one of and interpret it appropriate term of or i'm gonna brand x
0:07:30	as
0:07:31	conversational elicited right it's meant to kind of play on use other terms and pragmatics
0:07:36	implicature explicate sure imps the structure and so forth which are we talking about the
0:07:41	moment
0:07:42	to get at the idea that what you have is a speaker who is choosing
0:07:46	her referring expressions among alternatives
0:07:50	so as to trigger inferences on the part of her here that wouldn't otherwise be
0:07:55	drawn
0:07:56	so wanna do one is talking is that i'm first gonna
0:07:59	the gap a little bit on the kind of linguistics and philosophy side so a
0:08:04	topic
0:08:05	they're with me on that
0:08:06	and kind of talk about why this is a new type of richmond in the
0:08:10	literature and what are the car what are the aspects of
0:08:14	people's cognitive ask apparatus
0:08:17	that the speakers taking advantage in being able to communicate this extra content
0:08:22	and that's can be largely joint work with jonathan how women the philosophy department use
0:08:26	est
0:08:27	then i'm gonna go experimental
0:08:30	with joint work with honda roller at university of edinburgh in talk about how a
0:08:36	list features are just important for getting all the content however at of the message
0:08:42	but also actually impact
0:08:44	the interpretation of language an unexpected places in this case illustrated with pronoun interpretation
0:08:50	and then
0:08:52	i will conclude with some slides on the ramifications of the model that will build
0:08:55	for computational work in the area
0:09:00	so if you are you know
0:09:03	from hollywood pragmatics
0:09:05	you probably react
0:09:08	to beat examples pressing one that sounds familiar that sounds like could be
0:09:12	in cases of the gradient implicature
0:09:14	right so
0:09:15	i think a lot of we only know what implicature is that won't going to
0:09:18	detail but the important thing is that
0:09:21	a according to grice's implicature results from assumptions of a rationality and whopper activity among
0:09:28	me in the lock interlocutors
0:09:31	you can just at out in terms of for maxims i well i will read
0:09:34	them but will be most interested in
0:09:36	the first quantity maxim
0:09:38	which says you know say is much improved
0:09:40	information is required
0:09:42	the third some maxim of manner that says to be brief
0:09:46	avoid unnecessary fill actually and then finally the relation maximum says be relevant
0:09:53	so that the important thing i want to focus on is that implicature is a
0:09:56	failure driven process
0:09:59	meaning
0:10:02	the here and encounters a problem and ralston implicature to fix it so basically what
0:10:07	happens is the speaker
0:10:09	says something that has the literal meeting say color p
0:10:13	and the here says gee it just a really means peace you wouldn't be very
0:10:18	what order
0:10:21	but
0:10:22	rather than that
0:10:24	this one
0:10:26	but i identify some after information for q
0:10:30	i assume that she's can trying to convey
0:10:33	then she becomes cooperative again
0:10:36	and so mean
0:10:37	i think in fact she intended that i do this whole calculation and draw the
0:10:41	inference q in addition to the content
0:10:43	p so to illustrate right we're gonna be talking about referring expressions amiss talk
0:10:49	and grace was the first denote the choice of referring expression
0:10:53	so i can in some cases have hallmarks of implicature so he's kinda
0:10:57	rather dated example was for actors meeting a woman this evening
0:11:02	which would normally implicate that the woman being mentioned is not ex's wife
0:11:07	sister mother and so even know those are all when
0:11:12	so the idea is that
0:11:13	if you're talking the speaker was talking about acts as y
0:11:17	then
0:11:18	she what is said white
0:11:19	but n and n in accordance with the maxim of quantity give as much information
0:11:23	is required
0:11:25	since the speaker didn't do that
0:11:27	we're gonna draw the inference that in fact
0:11:30	the referent the space of four of possibilities for a woman don't include these other
0:11:36	kind of salient possible a reference that would be denoted by terms like a system
0:11:44	otherwise and so
0:11:48	so implicatures
0:11:49	right
0:11:51	or kind of we would those out with standard tasks
0:11:56	basically when you have implicata content
0:11:58	you could do a few things with it you can actually asserted input on the
0:12:02	record that's a reinforcement
0:12:04	they can say x of meeting of a woman this evening in implicatum out his
0:12:08	wife and then you can actually save
0:12:10	but not his wife
0:12:11	and that doesn't have a strong sense of redundancy
0:12:15	you can select
0:12:17	in fact ceases wife
0:12:18	a or you can actually get on the record that you don't know that the
0:12:22	two status of the imply could consist
0:12:25	impact that's in fact possibly twice
0:12:27	well you or times are examples satisfy these tests as well right you can say
0:12:33	the company fired manager whose embezzling money in fact that's why you got fired
0:12:37	that's a reinforcement
0:12:39	but that's not widely got fired
0:12:41	cancellation
0:12:42	and that mainly why he got five which is the suspension
0:12:45	so our researchers just implicatures
0:12:49	there's one
0:12:50	one person who i think is really given a serious pragmatic example of exam analysis
0:12:55	of examples of the kind a general character
0:12:58	that i'm talking about here and have a and we profile
0:13:03	so i took this
0:13:05	this is a kind of a an example come
0:13:08	from the first
0:13:11	hillary clinton donald trump presidential debate in the us lester whole
0:13:16	is the moderator from n b c
0:13:19	and what is in trouble starting to ask a question
0:13:24	he did not say seven a
0:13:26	right research on for five years you perpetuated of false claim that brought about what
0:13:30	was
0:13:30	not an actual word that is
0:13:32	it's not only said what he said instead with seven b
0:13:36	it should run for five years you perpetuated of false claim that the nation's first
0:13:39	black right
0:13:41	was not a natural born citizens
0:13:43	those two sentences are extensively equivalent
0:13:46	right they differ in these over for rain expression that denote the same individual
0:13:51	but seventy goes beyond seven a
0:13:54	right in kind of
0:13:55	giving rise to this idea that there could be some kind of causal relation between
0:14:00	drama hassling a one man and his status as the first
0:14:04	why are present
0:14:07	fortunately nothing had happened the sense that you make its worry about rampant racism
0:14:14	sarcasm
0:14:16	and if we compare that with seven same as for example five years to perfect
0:14:19	way to false claim that the first part of the place to one of women
0:14:22	on that some key where was i do not report susan that gets a little
0:14:25	kinda confusing
0:14:27	i
0:14:28	using the one
0:14:31	explain actually to referring expression even know
0:14:34	that possible first to a bomb
0:14:36	so
0:14:38	compels ideas that you for you see that these referring expressions are longer and more
0:14:42	descriptive the need to
0:14:44	they violate the product c d sub maxima and the maxim of quantity
0:14:49	and what i and basically what you do is happens with some kinds of implicatures
0:14:53	is
0:14:53	you rescue it
0:14:56	by way of another max
0:14:58	in this case relation you find
0:15:00	this relevancy relationship
0:15:03	that justifies the use of the more probable x more informative referring expression there's a
0:15:09	lot of technical detail here that i'm just gonna gloss over
0:15:12	okay so now making it case so far that a list features or a species
0:15:16	of the implicatures
0:15:18	but the in general
0:15:20	these cases do not pattern with template
0:15:23	so
0:15:24	can maybe try triggered by the maxim of manner
0:15:29	not really right probably studies
0:15:31	an issue
0:15:33	probably the use my
0:15:35	require so and he'd age on fire the employee who was always late
0:15:38	you get the elicit your
0:15:40	john fired employee who is read here we generally don't
0:15:44	the relative pauses just picking out one
0:15:47	salient employee
0:15:48	and there's no real difference meaningful difference in perplexity between those two referring expressions
0:15:54	and ac john fire the employee who is right here appeared in glasses
0:15:58	is more products but you still don't get the a causal inference
0:16:03	so what the maximum and at elvis is that e c
0:16:06	might be side in a situation where at what it's advice right of there's only
0:16:11	one employee would right here why going on about to be are in classes but
0:16:14	its orthogonal to the existence of a causal inference in again like eight
0:16:20	a another reason for doubting
0:16:23	maxim of manner being
0:16:25	relevant here
0:16:26	is that
0:16:27	these examples lack kind of the canonical
0:16:32	the heat here
0:16:34	implicatures driven by mail or so
0:16:37	what
0:16:38	larry horn call the division of pragmatically or so
0:16:41	if we compare john kill bill with john "'cause" tilted i
0:16:45	those essentially have the same view notation
0:16:49	but you get this division where the shorter version tends to describe the more typical
0:16:54	situation and a longer version them or a typical situation so
0:16:59	you know when i say john hospital did i
0:17:03	you was probably be surprised if you wanna john just one often shot building
0:17:08	you can get the sense that do exist
0:17:10	might a bit indirect causation or accidental killing or something like that
0:17:16	because only because
0:17:18	if gunshot build a it probably would just said john killed
0:17:23	so in of are cases you don't have this you just talking about competing referring
0:17:28	expressions of all denote the same reference
0:17:30	there is no this characteristic division of the do you notational space
0:17:36	so what about the maximum relevance you might be thinking relation it might be thinking
0:17:40	these are just kind of relevance implicatures
0:17:43	but that doesn't really work
0:17:45	either "'cause" the problem is relatively more restrictive relative clauses there
0:17:50	can stream
0:17:52	the dean the
0:17:55	the reference a b and p to which they attach
0:17:59	are kind of by definition relevant
0:18:01	so it can be a couple if i really am manager whose higher in two
0:18:05	thousand and two
0:18:06	that relative clause
0:18:08	is fine
0:18:10	even though it doesn't give rise any pair of causal inference
0:18:14	so by then relation you don't have a an explanation for why you go beyond
0:18:18	that draw comp a causal inference in the case like a ten day
0:18:23	really what the feeling is that the these inferences are not
0:18:27	triggered by gracie in maxim violation
0:18:30	it's the it's are already are machinery for recognizing relevance
0:18:34	thank
0:18:34	gives rise
0:18:36	so the inference right
0:18:38	by the time you would think of in terms of triggering the maximal relevance
0:18:42	you've already identified the relevancy relation
0:18:46	it's a more automatic process
0:18:50	there's a number of other types of pragmatic enrichment that have been discussed in the
0:18:53	literature you know i'll go to just
0:18:56	cut on this we use quickly
0:18:59	you know from rice
0:19:01	you know it's a pretty simple picture right you would
0:19:05	hearer's
0:19:08	interpret sentences do a little work we on that in terms of fixing reference i
0:19:13	index tickles tends interpretation
0:19:15	and b ambiguity resolution
0:19:18	and then everything else is left to implicature other researchers have argued that there's other
0:19:24	types of enrichment that go beyond
0:19:27	what's literally said but
0:19:29	we wouldn't wanna call implicatures so
0:19:32	it's is box implicit sure and part of what constitutes a explicate your relevance this
0:19:38	so these are cases like lemonade i'm always true crazy
0:19:44	well we don't really can't even decided to value to that unless we know you
0:19:48	know to pray six or what
0:19:49	so that's called a completion
0:19:52	in a way of other cases like second class cases like eleven b i haven't
0:19:56	had breakfast
0:19:58	which you know that usually need ever it just means today
0:20:02	right so you can compare that to a sentence like
0:20:04	i haven't headset
0:20:06	which usually means ever and not today
0:20:10	unless you live image society of course where people typically have sex every morning but
0:20:13	very rarely have breakfast and then presumably the justice record of slot
0:20:18	so the crucial thing there's a lot to be said about of these but
0:20:23	crucial thing is these all constitute
0:20:25	developments
0:20:27	expansions completions to the logical form of a single
0:20:33	utterance
0:20:33	where and it again their failure trip
0:20:37	either the sentence is an even complete enough to assign a to a value or
0:20:41	it is complete
0:20:43	but it can represent something that the speaker would plausibly once the same as in
0:20:48	the breakfast example
0:20:50	so you have to narrow it's t d notation
0:20:52	elicited don't have a characteristic in all right the sentences are perfectly well formed
0:20:57	without
0:20:58	the inferences in question
0:21:01	they're not triggered
0:21:02	by any
0:21:04	communicative any risk of communicative failure
0:21:09	okay
0:21:10	and the and then they involving inference of
0:21:13	then do not the completion of a logical form but they it it's an additional
0:21:17	inference additional proposition so the company fired employee who's always late
0:21:23	and another obstacle
0:21:24	it was the lateness because the five
0:21:28	so i there's a lot this is said in terms of other types of enrichment
0:21:32	and but i'm not i won't
0:21:35	i think you get the picture so then the question is where do these a
0:21:39	list features come from
0:21:42	and i'm gonna argue that they come from
0:21:46	part of our contributions apparatus
0:21:49	that
0:21:49	many of you actually in this audience will be familiar with
0:21:53	less so for other audiences the type of presented this at
0:21:57	presents two
0:21:59	which basically it's or it's the same machinery that we used to establish
0:22:03	or world is coherent
0:22:05	right so
0:22:05	it's well known that we interpret when we interpret our world we go well beyond
0:22:09	what our perceptions give
0:22:11	right so
0:22:12	if we're working at more or something and you see this chronically tardy employee show
0:22:18	up late for work
0:22:19	and then witness a few minutes later
0:22:22	and getting fired
0:22:23	you probably draw inference
0:22:25	that there's a call a causal inference between the two
0:22:28	the feasible you know could be wrong
0:22:31	but you draw all these kind of inferences
0:22:34	anyway
0:22:34	but if you see a party employ articulatory employee coming late again
0:22:39	and couple minutes later class for what to say where is the automotive department
0:22:44	you don't draw causal relation between those two it's just two events that happened in
0:22:48	the world is perfectly coherent otherwise
0:22:50	so if we make these kinds of enrichment
0:22:53	were not as a situation
0:22:55	so guess
0:22:57	as we interpreter world
0:22:58	it only makes sense that we would make similar kinds of inferences
0:23:02	when we understand natural language descriptions
0:23:05	of the world
0:23:06	right so which is why we see the boss fired employee who came in late
0:23:10	again you might draw this inference
0:23:13	and when you see a customer s employee who came in late again with the
0:23:16	automotive department is
0:23:18	you want draw a causal inference
0:23:20	so many ways he's inferences of the or the most pedestrians or right there just
0:23:25	the kind of inferences we draw
0:23:27	to establish the coherence of our environment
0:23:31	and as a argue it's a very different kind of process
0:23:34	then the other kind of more value driven processes that underlie other kinds of pragmatics
0:23:40	enrichment
0:23:42	so what are these cognitive principles well
0:23:45	yes there will be
0:23:46	familiar to a lot of them you lot of you
0:23:49	they're the same kind of principles that underlie reestablish of establishment of coherence
0:23:55	in discourse between set s
0:23:58	so
0:23:59	in seven a the boss fired him for you came in late again its essentially
0:24:03	the same kind of inference that
0:24:05	you will get to establish an explanation coherent relation
0:24:09	for seven b
0:24:10	the boss fire the employee
0:24:12	it came in late again
0:24:14	it typically infer causal relation we've also seen by a labial expectation relations
0:24:20	the company fired to manage to is a long history what words same inferences if
0:24:24	you break it up between sentences
0:24:26	the company filed the manager he long history of corpora towards
0:24:30	we've also seen cases a better non-causal or maybe just like enable my relations like
0:24:36	i with very hard collocation
0:24:39	we employ you want to the still the but we employed want to the store
0:24:42	bought a bottle of scotch
0:24:43	for the authors part i have somebody said that to you and them
0:24:47	somebody later ask so where the employee get this sky
0:24:51	you probably say at the grocery store
0:24:54	not probably not notice that sentence doesn't never says
0:24:58	it's just an inference that you draw to connect the going to the grocery store
0:25:02	and the binary files got
0:25:05	just like you would draw for across causes the employee went to the store she
0:25:11	bought a wildcard
0:25:12	for the office party
0:25:14	the crucial difference how ever
0:25:16	is that
0:25:17	when you're establishing coherence between
0:25:23	these sentences
0:25:24	that's a failure driven process right language mandates that when you have sentences within the
0:25:30	same discourse segment you have to find some kind of relevancy relation between
0:25:35	less we be satisfied for discourse is like not dale's twenty be replicated adaptation so
0:25:42	you know the employee broke his leg
0:25:44	you like models
0:25:46	we'll probably strike you as a kind of a discourse right you don't to say
0:25:52	you know i just i just one two things about the employee great
0:25:56	move on
0:25:57	right now you might object and say
0:26:00	well wait a sec
0:26:01	i think that could be coherent may the employee happened upon a problem tree try
0:26:06	to climate to get a aplomb and so i'll broke is like
0:26:11	now it's hard pointed out many years ago that they are shows you
0:26:15	right that you are within two car in its to by the search for coherence
0:26:21	right you
0:26:23	you know is interpreted has to check this sense you want to search for coherence
0:26:28	between the utterances and you willing to accommodate a certain amount of a context to
0:26:32	do that that's totally different from twenty eight
0:26:35	by say that what we employ a would like one broke is laying does not
0:26:39	send you off on the search for coherence
0:26:42	it's just employee broke his leg which one o one like ones among we others
0:26:47	okay
0:26:48	so
0:26:49	same time of the machinery twenty eight feet is so tell your free try nothing
0:26:55	in the sentence is explicitly telling you have to search for coherence in a way
0:27:00	that twenty be does
0:27:04	so really what's happening here it's just like other kinds of pragmatic enrichment
0:27:10	right where the speaker is taking advantage of her here's cognitive some aspect of our
0:27:15	current cognitive apparatus in constructing a referring are utterances
0:27:22	so the case of implicature again its reasoning about you know rationality poverty the right
0:27:27	a five assigning grades and as soon as me about the grades in my class
0:27:31	and i say some students will get an eight
0:27:33	i'm not being cooperative
0:27:35	if it turned out that it every student
0:27:38	even though it's like able students in a actually gives students and
0:27:43	a you have cases like indirect speech acts right where we know like
0:27:48	these are all over dialogue and you have the reason about
0:27:52	the plan-based goals of the interlocutors
0:27:55	beliefs desires and intentions and all that kind of thing
0:27:59	it's the same kind of thing except the aspect of cognitive
0:28:02	here's cognitive apparatus taking advantage of this is more basic
0:28:06	kind of associative a reasonably kinds of reasoning that can extract the last in a
0:28:14	in a temporally extended convolutive
0:28:18	sequence
0:28:18	so basically we have this machinery for understanding coherence in our world we use that
0:28:23	for understanding coherence across utterances in
0:28:27	dialogue
0:28:29	and discourse
0:28:30	and then the speaker takes advantage of that into using a referring expressions within a
0:28:34	sentence to give rise to these inferences even though they're not mandated by anything that's
0:28:40	explicit in the utterance okay
0:28:43	so
0:28:45	so i think this is the structures are particularly difficult challenge problem
0:28:51	when you're building computational systems precisely because
0:28:55	there's yes right we build systems we think of
0:28:58	triggering interpretation problems we see an utterance
0:29:01	and we have to you know we have to interpret it we see a problem
0:29:05	in we have to search for reference
0:29:07	we see multiple sentences and we have to find a coherence relations
0:29:12	cases of the list of judges just nothing there
0:29:14	that's saying hey you have to try to search for you know every possible any
0:29:18	kind a causal relation that could occur between the content of any two constituents right
0:29:22	it's something that a rises automatically
0:29:27	when you have the cognitive apparatus that we
0:29:30	so
0:29:32	hopefully
0:29:33	at this point can be into solicitor's arms
0:29:36	important part of
0:29:37	extracting the for meaning out of utterances now minutes which years and to
0:29:41	experimental mode
0:29:43	with the joint work with on a roller
0:29:46	but in argue that
0:29:48	i can elicit yours is an important part of
0:29:53	tracking discourse meaning
0:29:54	and ultimately can affect
0:29:57	interpretation of downstream linguistic some i'm gonna do that
0:30:01	make that case with respect to
0:30:03	a particular problem pronoun interpretation
0:30:06	so i think it's
0:30:08	the safe to say
0:30:10	but is in a common wisdom in the field
0:30:14	reference for
0:30:16	for decades which is that there's this unified notion
0:30:19	of energy salience or prominence mediates between pronoun production interpretation
0:30:26	speakers
0:30:28	use pronoun to refer to salient reference
0:30:31	and then hearer's users think use the salience
0:30:34	to interpret
0:30:35	the reference
0:30:36	they're mirror images of each other
0:30:39	happen to be any other way
0:30:41	and then so then you know the pulse stress discourse terrace
0:30:45	it's just identify what are these different contributors to energy scaling some i put you
0:30:50	know
0:30:51	a very
0:30:52	i a partial list their own
0:30:54	in the bowl
0:30:55	don't and it's fifty minutes or so i'm gonna kind of this completely this is
0:30:59	used to of this idea
0:31:01	the experiment so i'm gonna describe all about implicit causality concept so
0:31:06	let me take a moment to tell you what those are
0:31:09	right these are
0:31:09	is or verb student very well studied in the psychology literature
0:31:13	and their said to impute causality to one
0:31:17	of there are two of an artist a tense
0:31:20	such a ten that
0:31:23	computing of causality then affect
0:31:25	downstream referential by a six
0:31:28	so if you run a little experiment
0:31:30	in your lab or on mechanical turk us people to complete the sentence
0:31:34	amanda mazes britney because she
0:31:38	right there it completions you have three annotators tell you what you refers to i
0:31:41	can tell you what's gonna happen
0:31:43	by enlarge the vast majority are gonna write something about amanda
0:31:48	we just found that amanda is amazing
0:31:51	and we're gonna here we
0:31:52	okay so those are subject biased implicit causality verbs
0:31:57	you can compare that to the second case amanda detest britney because she
0:32:02	now we're gonna hear about britney
0:32:05	we just for different means detestable
0:32:08	and we're gonna find out what those are updated by implicit causality verbs
0:32:13	now a couple things worth mentioning here if you run in experiment
0:32:17	where you don't include
0:32:19	but well so the but that was here usually experiments as like a linguistics literature
0:32:24	use the cars and of course that indicating a particular type of coherence relation an
0:32:28	explanation relation you're gonna hear a cause
0:32:31	or reason that follows
0:32:33	and
0:32:34	that's really what these strong bias a user or try to
0:32:38	so if you ran a study that we just adam animations britney
0:32:41	and let people write the next sentence a couple things will happen
0:32:45	one is
0:32:46	you'll still get the biases but they won't be a strong
0:32:49	because you're gonna get some other coherence relations decides explanation you're not gonna have the
0:32:53	same by sees if somebody tells you know what happened next or something like that
0:32:58	but the other interesting thing that happens
0:33:01	and i wrote a showed years ago
0:33:04	is that you will get
0:33:07	many more explanation relations
0:33:10	in an implicit causality context
0:33:12	then
0:33:14	for other kinds of content
0:33:16	so it should make some sense if i say amanda just has britney
0:33:20	what do you thinking
0:33:22	why
0:33:23	you can tell me why i need to know why provides a you know amanda
0:33:26	solver e
0:33:28	you're not thinking
0:33:29	wow i need to know why okay well what happened next right so they generate
0:33:34	god greater expectation you're gonna get a cause or reason
0:33:38	in an icy context and i'm foreshadowing that's gonna become important a couples slides yes
0:33:45	so
0:33:46	to give some background there was this study is very influential
0:33:51	in my thinking by rosemary stevenson and colleagues and nineteen ninety four
0:33:56	where they did set task completion studies vary across a different context types including the
0:34:02	two implicit
0:34:06	and they compared
0:34:08	what happens if you give people a pronoun prompt verses no problem
0:34:13	so in the first case you get my pronoun it's ambiguous between the two then
0:34:16	participants and you see how they assign to run
0:34:20	in the in the three prompt condition you find out to things
0:34:23	you find out who they mention next
0:34:26	and
0:34:26	what form a reference to they choose
0:34:28	do they use a pronoun where they use any
0:34:32	they found to really interesting facts one is that
0:34:36	when you given the problem now
0:34:38	you always get more references to the previous okay
0:34:41	then when you do
0:34:43	across all a context types
0:34:45	now the overall is might not be to the subject
0:34:50	might not be in an object by simplistic causality context
0:34:53	but you still get more to the subject
0:34:55	when you let them take the referring expression
0:34:58	the second thing that happens is that
0:35:00	again across all context types there is a strong production tendency when they're referring to
0:35:06	the previous okay
0:35:07	they like to use the pronoun
0:35:09	maybe that at each one
0:35:11	and when they were for to the previous non stop
0:35:15	they like to repeat and me
0:35:18	so that is computed for a little while
0:35:20	well of people clearly have
0:35:22	this production bias it says for normalize the previous subject
0:35:26	don't problem lies in previous object
0:35:29	why would you have
0:35:31	ever get an object
0:35:33	found out as and but not by simple the called out of context
0:35:38	in terms of the that's actually not paradoxical
0:35:41	at all
0:35:43	once you can ask the relationship
0:35:47	between interpretation and production in terms of bayes rule
0:35:51	so this term on the left
0:35:53	is the interpretation problem
0:35:56	or interpreter see the pronoun and has to figure out what the reference
0:36:00	the first time in the numerator is
0:36:02	are production is the production bias
0:36:06	our speaker knows what you want to refer to and has to decide whether use
0:36:10	of pronoun or not
0:36:12	bayes rule tells us that these two one your images of each other
0:36:15	there's another term there in the numerator
0:36:18	the prior
0:36:19	the prior probability that a particular referent is going to get mention next
0:36:24	regardless of the for linguistic form
0:36:26	other speaker chooses to do it
0:36:29	okay
0:36:31	so there's nothing paradoxical about having a production bias
0:36:36	that says pre-normalized the subject
0:36:38	much more than minimizing the object
0:36:41	and then interpretation by s
0:36:42	they close to the object
0:36:45	as long as the prior probability of who's gonna get mentioned next
0:36:48	is weighted strongly enough
0:36:50	towards the arc as it is interrupted by simply the called out
0:36:56	now
0:36:57	theory can comes into forms kind of the weak formant a strong for the week
0:37:02	form just as
0:37:03	we expect interpretation production to be related by bayesian principal
0:37:08	but we posit that the stronger form "'cause"
0:37:11	all the evidence that we have seen at a time
0:37:14	pointed to the fact that the to use the types of contextual factors the condition
0:37:19	the two terms in the numerator
0:37:21	seem to be very different
0:37:22	all the semantics and pragmatics stuff semantics like verbs i
0:37:27	implicit causality
0:37:29	pragmatics like coherence relations
0:37:32	seem to be affecting not problem interpretation directly but the prior
0:37:36	those are pushing you your expectations it's about who's going to get mentioned
0:37:43	the production via seemed much more basic based on things like grammatical role some get
0:37:47	a or probably more probably information structure what's the top
0:37:53	you know pronouns like a lot centering theory basically say hate i think i was
0:37:57	talking about before and still talking about it
0:38:02	no when you can see like this makes in extremely counterintuitive prediction
0:38:07	which is that the speaker in deciding whether she's gonna use a pronoun or not
0:38:12	is ignoring a rich set of semantic and pragmatic pisces
0:38:17	that's those conditioning the prior
0:38:19	that the interpreter is nonetheless going to bring the bear in interpreting the problem
0:38:24	i think very a
0:38:26	but despite its honest
0:38:27	a number of experiments have provided evidence that is in fact
0:38:32	the case
0:38:34	so
0:38:35	that's it you're is a and experiment from
0:38:38	and a rotors thesis the three by two
0:38:42	should look familiar this twenty
0:38:43	the three way to three waiver five
0:38:47	comparison
0:38:48	subject by a simplistic all value added biased
0:38:51	i see verbs
0:38:52	and an icy verbs
0:38:54	and in the from you and affiliation
0:38:57	three problem versus pronoun problems
0:39:00	so the prediction is that verb phrase verb type should affect the prior
0:39:06	and imagined of the effect in the prior for a cascade to affect interpretation
0:39:11	but that verb type
0:39:12	will not
0:39:13	affect production
0:39:15	right so
0:39:16	again in the in the three prime condition we get to measure two things we
0:39:21	see who they mentioned next
0:39:22	that's our measurements of the prior
0:39:24	and we see what number of reference way to get that you
0:39:29	they choose whether use a pronoun and so we get the production bias
0:39:33	and then down here we wanna given the pronoun we get direct access to their
0:39:36	interpretation
0:39:38	giving them a pronoun how to interpret the
0:39:41	okay so
0:39:43	we're predicting an affect
0:39:44	a verb type on both the prior and
0:39:49	pronoun interpretation and that's exactly what we
0:39:52	so you see more subject references the subject i z condition
0:39:57	the least in the object i c condition
0:39:59	and then on ice verbs or somewhere in between
0:40:02	and then you see that the light or light
0:40:05	blue bars those of the pronoun problem condition
0:40:08	data
0:40:09	are always a little higher than
0:40:11	the prior the dark blue bars and that's the actor production bias coming in the
0:40:17	production term that's tilting everything towards the subject from the baseline presented by the prior
0:40:23	okay so that works out
0:40:25	now did verb type affect
0:40:27	production when speakers to use pronouns verses names any answers no not at all
0:40:33	only thing that matters
0:40:35	is grammatical role lot of pronouns for subjects not a whole lot for objects
0:40:40	right so to put a fine point on this
0:40:43	right people or no more likely to use a pronoun
0:40:47	to refer to the direct object
0:40:50	in a biased implicit causality context
0:40:56	then in this update bias implicit causality
0:40:59	and then one or more likely to use a pronoun to refer to the subject
0:41:03	and a sub device context and a bias context there is a dissociation between production
0:41:09	by sees and interpretation
0:41:14	so
0:41:15	and a noun take the last two parts of the talk
0:41:19	and bring them together and one b new tiny little experiment it's a two by
0:41:23	two
0:41:24	when a very prompt i as before
0:41:26	and we're gonna have a model that manipulation that involves and the literature
0:41:31	so you compare the boss widely employed was hired in two thousand two verses of
0:41:35	all so far we employ was embezzling money
0:41:38	now most there is a condom interpretation and i pretty much all the taurus i
0:41:42	think
0:41:43	don't predict any difference and pronoun by season those two cases
0:41:47	the same subject the same for the same object
0:41:50	the relative clause is a little different
0:41:52	that's and introduce any new reference who cares
0:41:55	but are analysis the bayesian analysis does predict the difference
0:41:59	based on this interconnected sheen
0:42:02	of referential incoherence driven dependencies
0:42:06	so here's
0:42:07	gives a crucial slide
0:42:10	what are we expecting that
0:42:12	when you have
0:42:13	the when you have
0:42:15	you know at in the literature
0:42:20	in the relative clock so we call that you split at all
0:42:24	or three condition
0:42:25	right the relative also gives you an explanation
0:42:28	versus the control condition when it doesn't
0:42:30	i told to first that
0:42:32	when you have a these are all gonna be uttered by simplistic causality verbs when
0:42:37	you have an icy context
0:42:39	you're really expecting an explanation to come
0:42:41	we exploit the lot of a
0:42:44	exhalation coherence relations exact
0:42:47	in the explanation or c condition
0:42:49	we are defined explanation
0:42:51	it was in the relative cost
0:42:53	so we predict that you're gonna get fewer explanation coherence relations
0:42:57	after those cases then in the control condition
0:43:01	why give an explanation when the proper already have one
0:43:05	batch and then can say to affect the prior the next mentioned bias
0:43:09	user i've requires verbs we expect a lot if we have a lot of explanation
0:43:13	relations you expect a lot of
0:43:14	object references
0:43:16	but then we have you have fewer exclamation relations in the explanation or c condition
0:43:21	then you're gonna get fewer object mentions
0:43:23	because
0:43:25	the object biases try to there being an explanation relation
0:43:29	so we expect an effect on the prior
0:43:31	we also expect
0:43:35	and effect of the production by this what we seen before
0:43:38	in interpretation we expect to see more pronouns
0:43:42	referring more mentions of the previous subject when you get more prone then when you
0:43:46	down
0:43:47	i'm sorry the production by we expect people to produce more pronouns to refer to
0:43:52	subjects
0:43:52	then objects
0:43:55	and then when you put those two together at the bottom
0:43:58	both terms the prior and the likelihood term should affect interpretation
0:44:03	more or fewer references to the object that is more to the subject
0:44:08	in the exclamation rc condition
0:44:10	and also within the pronoun problem condition compared to the free problem condition
0:44:17	the crucial thing about this slide right is that
0:44:21	here's a little graphical model for influences on pronoun interpretation
0:44:26	and all the interesting stuff is on the right-hand side
0:44:30	all the stuff that's completely independent
0:44:33	a pronunciation
0:44:35	that
0:44:35	all building on the right is about predicting
0:44:38	the message who's going to get mention next
0:44:43	the most boring part of the slide is the part
0:44:46	over here where a pronoun comes into play
0:44:51	so
0:44:52	notice that this part of the a pop years possible to affect
0:44:55	on interpretation directly only indirectly
0:44:58	okay
0:44:58	so first predictions do we get
0:45:00	fewer explanations
0:45:02	in the when the relative clause already gives you one yes
0:45:05	people
0:45:06	do you still get some explanations but
0:45:10	not as monies in the control condition
0:45:12	people one explain why the person higher than two thousand and two got fired more
0:45:17	than they wanna explain
0:45:18	why the person who was embezzling money got fired
0:45:23	does that affect the next mention biasing yes
0:45:25	as we expected you get more mentions of the direct object
0:45:30	in the control condition than in the explanation or c condition
0:45:35	the and the existence of a causal literature in a relative clause
0:45:40	affect production or not
0:45:42	not at all
0:45:43	same pattern we seem before
0:45:45	all a matter was grammatical role
0:45:47	and then when you put these two things together you get expected interpretation patter
0:45:54	you get the existence of the literature pushes around
0:45:59	the prior when we so like to slide to go about those of the white
0:46:02	blue bars
0:46:03	and i map object references here so when you give people pronoun prompt
0:46:08	those parts go down because you get
0:46:10	the production by given by using everything towards subject reference so fewer object references when
0:46:17	you give them a pronoun
0:46:20	okay so
0:46:23	this idea that again production and interpretation
0:46:26	are mirror images of each other
0:46:29	is clearly not happening and something is kind of subtle is the existence of the
0:46:33	list such are way up here
0:46:35	you can see how often cascades to affect
0:46:38	several other things and ultimately down here then tweaks your by sees for how you
0:46:43	would interpret
0:46:46	quickly we can do little model comparison
0:46:49	you know passes completion studies don't really
0:46:52	rate that's highly on the sex appeal meter and cycle
0:46:56	but i want doing them because they give us actual fine grained
0:47:00	numerical
0:47:01	measurements for biases
0:47:03	and so we can use that to compare different models so again what we can
0:47:08	do
0:47:10	we can estimate interpretation by using our free prompt condition
0:47:15	we get can measure
0:47:16	really mentioned next that gives us the prior we get to see whether they use
0:47:20	the pronoun are not that gives us the production bias we can plug them into
0:47:23	this equation get interpretation by s
0:47:26	then we can compare that with the actual interpretation by s
0:47:30	there we
0:47:32	c in the pronoun prompt condition
0:47:34	right so we're estimating
0:47:36	we coming up and the estimated bias from the free from condition using this formula
0:47:40	in comparing it to the actual one we find in the pronoun condition
0:47:45	we can compare this with two kind of competing models that are out there one
0:47:49	is
0:47:50	and of the what i've been calling them your model
0:47:53	that's where in there so
0:47:55	what we reference
0:47:56	was the speaker most likely to use a pronoun to refer to
0:48:00	so we can calculate that by taking the production bias and normalizing
0:48:05	i wrote it this way
0:48:07	just to point out that its essentially like bayes rule set without the prior
0:48:12	the other model is the agenda for arnold expectancy model she said look what's happening
0:48:18	is
0:48:18	you greater generating expectations about who's gonna get mentioned that
0:48:22	and if you have a
0:48:24	you see a pronoun
0:48:26	and it matches and gender number or not i think
0:48:29	that tells you say that's the thing
0:48:31	it's the thing you're expecting get mention x
0:48:33	that's essentially just the prior now the priors already probability distribution soapy referent would have
0:48:39	sufficed
0:48:40	but i wrote it this way to show you that this is
0:48:42	basically bayes rule except without the production bias
0:48:45	and
0:48:46	when you compare the numbers basically the bayesian model when so these in the actual
0:48:51	column or the actual numbers we get
0:48:54	for article percentage of object references
0:48:58	in the problem i'm condition
0:49:00	and then you see three sets of numbers
0:49:02	four where we plug in the frequencies that we get in the free problem condition
0:49:08	into those different equations and you see
0:49:11	the bayesian members of predictions are actually pretty close and have a higher degree of
0:49:19	correlation
0:49:20	we expect all
0:49:21	the other models to have some correlation because
0:49:24	as i just showed you
0:49:25	essentially those models of being combined in the bayesian model but it's the combination of
0:49:30	the two that
0:49:31	that makes the best predictions
0:49:33	so to summarise this part of the talk
0:49:40	we see that you know pronoun temptation is
0:49:42	since it is very kind of subtle
0:49:45	coherence prevent factor where
0:49:47	production isn't
0:49:48	which
0:49:49	is counterintuitive but is exactly the dissociation that the bayesian model
0:49:53	would project
0:49:55	so contrary to this is that there is no unified notion of salience it's between
0:50:00	production interpretation
0:50:02	there's always in this problem in the pronoun interpretation literature right where
0:50:07	you know you read somewhere in the first paragraph of the paper it says you
0:50:11	know pronouns refer to salient reference
0:50:14	you say okay well what are the contributors the salience
0:50:17	as well that
0:50:18	go look at a corpus and see look identities pronoun to refer to
0:50:22	i two basic unit variance pronouns before to the kinds of entities that pronoun to
0:50:27	refer to its completely circular right so bad i have any meaning
0:50:31	right you're notion of salience has to be treated derived from
0:50:36	something that independent of choice of referential form which is what we're trying to predict
0:50:40	so for me i don't follow
0:50:43	l email to clocking here and three it's this next mention buys the prior
0:50:49	that's the best measurement we have for salient
0:50:52	right who you're expecting to get mentioned
0:50:54	but as we've seen pronoun vices don't know one directly
0:50:57	with that notion of salience
0:51:01	okay
0:51:02	so let me conclude with this a few quick slides oaks i think there are
0:51:06	some lessons for computational work here
0:51:11	ideas that i wanted to follow up on a long time adjust can ever get
0:51:13	a student interested enough so
0:51:15	i hope somebody here doesn't step
0:51:17	i think it's safe to say that when we've done computational work on reference
0:51:22	if you look over the last number of years
0:51:24	using a lot more progress on them on the mission the modeling side
0:51:29	that
0:51:29	the feature engineering side right
0:51:31	many new machine learning method
0:51:33	not aligned in terms of new
0:51:35	linguistic features right people still can be used the same three dozen or so features
0:51:39	gender number
0:51:40	distance maybe little grammatical role information that kind of thing
0:51:44	and for good reason because retraining these and systems unsupervised mode
0:51:49	you can ask people to annotate morton to three thousand pronouns
0:51:53	and so you can never ask questions
0:51:55	in your features that like is this an implicit adopted by some close to causality
0:52:00	you never have enough data to it to you know
0:52:04	to do something like that
0:52:06	well this
0:52:08	the bayesian model contest you don't need that indicate
0:52:13	because
0:52:15	it
0:52:16	you know the prior doesn't care all the semantic and pragmatic stuff
0:52:22	conditions the prior
0:52:23	and apply would you can calculate
0:52:26	doing so reference
0:52:28	for cocoa reference in general and not just for pronouns
0:52:32	you can go into data and have your system fine
0:52:35	case of the car reference that is really sure about
0:52:38	right repeated proper names
0:52:41	definite descriptions with substantial
0:52:43	lexical overlap
0:52:45	with their antecedents and pretend that human when and said that's co reference
0:52:52	you could get calculate millions of get millions of examples like that of the corpus
0:52:56	and then have a model that can has seems very fine grained features now you
0:53:01	might have a hundred
0:53:03	two hundred thousand
0:53:04	implicit causality verbs in there and be able to model that get some predictive power
0:53:09	added
0:53:10	all you need annotated data for is the pronoun specific part the production price and
0:53:15	a couple thousand pronouns is going to be plenty
0:53:18	to learn
0:53:19	that people pre-normalized
0:53:22	subjects the most and then less and less as you move down the oblique this
0:53:26	hierarchy
0:53:28	so
0:53:30	it was not at all obvious before that you could take
0:53:34	apply the factors that you learn for co reference in general using only like kinda
0:53:39	high probability cases the co reference and the teleport directly onto the pronoun interpretation problem
0:53:46	so
0:53:46	the situation is entirely analogous to bayesian models of other
0:53:51	kinds of things right now
0:53:54	machine translation in those the or in this case speech recognition right
0:53:58	you doing speech recognition with a bayesian model you could write well we could try
0:54:03	to train
0:54:05	a you know a model can directly that maps from acoustic signal to work
0:54:09	but we don't do that because
0:54:11	then when somebody says to
0:54:13	you've no idea they said t o
0:54:16	t o where t w well
0:54:18	right so we don't do that
0:54:20	instead we reverse it into production model given the words we predict
0:54:25	what's the likelihood that the speaker produce that acoustic signal
0:54:28	for that word
0:54:30	and then we can plug in the prior a language model like an n-gram model
0:54:34	imac and help tell us
0:54:35	where in that context
0:54:37	it is at o p w or well
0:54:41	the same idea right pronouns just like ambiguous words are used
0:54:46	underspecified signals a place strong constraints on their interpretation
0:54:50	but you need context a fully resolved
0:54:55	so
0:54:57	is it would have an efficient language should allow speakers to take advantage of
0:55:02	whatever aspects of or interlocutors cognitive apparatus you can get our hands on basically
0:55:07	for implicature that
0:55:10	collaboratively rationality for an indirect speech acts that plan planning and satisfying goals least designers
0:55:17	intentions for literatures it is more basic
0:55:22	aspect of our cognitive abilities that is
0:55:26	inferring relations have you do with
0:55:28	causality
0:55:29	com security
0:55:31	and the more
0:55:33	basic associated principles
0:55:37	so
0:55:40	you know when it when we know we build systems and easy the think of
0:55:44	language interpretation as a
0:55:45	as there is a reactor process right overall scheme
0:55:49	i need interpreter is a pronoun i need to a search
0:55:53	right everything happens when you see
0:55:55	the trigger right
0:55:57	on the other hand that the bayesian model
0:56:00	right is a more directly captures what is become
0:56:04	a more modern view of interpretation of
0:56:08	not as a reactive process but
0:56:10	one where interpretation is what happens
0:56:13	when you're top-down proactive
0:56:17	expectations about the ensuing message
0:56:19	commonly contact with the bottom-up linguistic evidence
0:56:23	by the by utterances
0:56:25	right
0:56:25	and so it's important i think of the case of the literature in a really
0:56:32	spells out the important
0:56:33	of doing that proactive modeling
0:56:35	right recognising these kinds of inferences and having that discourse update occur
0:56:41	so it's ready by the time you get
0:56:44	particular linguistic forms in the input like problem right you don't wanna wait to you
0:56:48	see a problem down a to run around a context and try to figure out
0:56:52	whether there's some of the list a true if you the
0:56:55	and i will
0:56:58	stop there thank
0:57:17	thanks very inspired design and i
0:57:21	definitely agree with a the
0:57:23	kind of approach to these kinds of inferences and the bayesian status great i had
0:57:29	a
0:57:30	couple of questions so
0:57:34	i guess you made a distinction about the
0:57:38	understands rely coherence relations versus intra-sentence and i don't think there's really a difference there
0:57:44	that i
0:57:45	i think you're sentences we're not really parallel and twenty and twenty b and
0:57:49	exactly the same kinds of coherence issues whether it's within one sentence or cross
0:57:54	to that
0:57:56	so
0:57:58	so that twenty seven twenty be like that if it was you know the employee
0:58:02	the likes plan c broke his leg that's
0:58:05	that's fine
0:58:07	and similarly the employee
0:58:10	likes plans and broke his leg this is just as weird as twenty p
0:58:13	so it's
0:58:15	the thing is i think issue thing so maybe
0:58:19	well let me
0:58:20	i are you are you commenting on my characterisation is intersentential versus intra sentential right
0:58:28	so i
0:58:28	i would nine i probably
0:58:31	there's is not a good term for that
0:58:33	i think it's exactly of what i want
0:58:35	right to compensate intra clausal purses inter-pausal
0:58:40	because the cases where you have been here
0:58:42	i'm not i'm those are still intersentential from me
0:58:47	and i'm but once you start saying intra clausal wow you now relative clauses the
0:58:51	clause and everything so
0:58:53	if you put
0:58:56	you know
0:58:59	a like a because in here or you know one hand or something i'd still
0:59:03	treat those as
0:59:05	intra sentential i intersentential
0:59:09	right we need we need to have
0:59:11	are coming here and
0:59:14	machinery come along and tell us
0:59:17	well i think i might have employed work is late because you like models
0:59:22	you don't a well
0:59:24	okay there must be causal relationship where it's been asserted i'm happy
0:59:27	no you need to establish the causal relation
0:59:30	right you're not happy until you
0:59:32	see you know
0:59:35	so that the crucial point is that in twenty a
0:59:38	right although it needs to happen to this to be explicit this is that we
0:59:42	can pick you know which employed we're talking about it doesn't trigger
0:59:46	this search well process but you know what i think i think it is that's
0:59:50	that is very search process that
0:59:53	you know the reason for this kind of free markets to identify particular employee and
0:59:57	that is a coherence relations
0:59:59	this you know identification purse or that i mean i depends on your area
1:00:05	coherence
1:00:06	but the crucial thing is you you're
1:00:09	you're not often running
1:00:11	here
1:00:11	this twenty we send you are running trying to figure out
1:00:14	how liking problems
1:00:16	could relate costly or otherwise right to breaking or like in a way that
1:00:22	twenty eight does most of the time
1:00:25	you know use a
1:00:28	this morning sense of the relative also there's no causal researcher
1:00:33	and it doesn't mean that where confused by all of those utterances right so the
1:00:41	question you know in the theory of pragmatics then is
1:00:44	when there is one why would you ever draw and that's
1:00:48	what's problematic for just about every type of enrichment that's out there
1:00:53	the that the triggers that day these different
1:00:56	if you know implicature implicit your
1:01:01	explicate sure for relevance theory and there's a mother's two
1:01:05	the comedies work in
1:01:08	you know local
1:01:09	pragmatic strengthening things like that
1:01:12	none of them had the triggers a need to give rise the inferences
1:01:27	i joint time
1:01:30	when doing a little bit of several works
1:01:34	so that you know it's well parents don't that early work on this is a
1:01:39	constant for
1:01:42	i really interesting they predict properties
1:01:46	it can send your last few slides out like where this is the computational approaches
1:01:50	is that
1:01:51	we need a corpus there surely the second sounding words rule
1:01:56	work are
1:01:57	and then we just the sse probabilities the implicit causality where cases where
1:02:04	we have all referring expressions we could
1:02:07	actually in there at least that's causality spend any system theory and
1:02:13	there i from a norse
1:02:15	somewhere right
1:02:17	so i displayed if we kind of ways
1:02:20	done it is part of your time and not just
1:02:24	i get my corpus blog stories
1:02:27	and i think a you know for all these lexicons nodding or any see what
1:02:32	happens next
1:02:36	and maybe i don't need to look for implicit causality groups need to adjust
1:02:41	wow
1:02:42	g you just have a lexical
1:02:44	probabilities with every
1:02:46	so i see you like that
1:02:49	yes that that's exactly right i mean you could
1:02:52	if you had enough data you could calculate
1:02:56	a probability
1:02:58	you know for every kind of or some or all four and of n participant
1:03:02	complex so there's no reason
1:03:05	you know to employers causality is a very weird
1:03:08	kind of concept in terms of
1:03:10	it really a cover term for a set of verbs that tend to have solar
1:03:14	by a six
1:03:16	there is no
1:03:17	deeper
1:03:19	definition for what implicit causality for is there are there are consistent subclasses so
1:03:26	experience or a stimulus for so you know annoyance surprise and you know the test
1:03:32	and
1:03:34	those kinds of verbs tend to be impose a causality but there are others it's
1:03:38	just you know like hit you know or things like that
1:03:43	have
1:03:47	just have strong by season with sailors and thus causality
1:03:51	so
1:03:52	there's a reason i think if you're gonna do modeling you know anything if you
1:03:56	have enough data
1:03:58	to just limit yourself to those kind of verbs because in fact
1:04:01	all verbs
1:04:04	have
1:04:05	some kind of biased you might want to account for is just gonna be more
1:04:09	meaningful when it's stronger one way or the other
1:04:12	i mean this hits on like a real problem all
1:04:15	in the cycle
1:04:20	you know the very one of the very first week
1:04:22	one the very first experiments we when that we random it can been that's what
1:04:27	we have these none of the so called reality
1:04:31	we had twenty of them
1:04:32	an hour now we forty of
1:04:34	and we calculated what the next mentioned by sees or for those that the prior
1:04:39	and only in it with
1:04:41	every spot
1:04:42	in a from between zero and one
1:04:46	okay you know starting from verbs it really should be considered impose causality concern so
1:04:52	far towards the end even though there's no real causality involved
1:04:55	there's ones in the middle in every point
1:04:57	help in
1:04:58	so i don't the cycle linguistics literature on column permutation
1:05:01	in people just make up a bunch of examples and saying
1:05:04	there's no pragmatic bias
1:05:06	i don't have any pragmatic bias
1:05:09	there's no such thing is a sentence it doesn't have
1:05:11	i'm having is
1:05:13	and some the verbs
1:05:14	i can take my three verbs it you want me to show you to a
1:05:17	noun setup a subject is
1:05:20	i'm gonna run a study with these twenty
1:05:22	you want me to show that there's no such that is
1:05:25	i'm gonna run at least one
1:05:27	right
1:05:28	now they are i'm gaming it "'cause" i know with verbs have you know
1:05:32	but not based on problem interplay to five is only the next mentioned by c
1:05:37	i can take one it has an x men a prior that eighty percent of
1:05:40	the object
1:05:41	run in a pronoun interpretation study
1:05:44	that pronoun is gonna pool be eighty percent to fifty percent under sail there's no
1:05:50	you know there's no bias it exactly what happens with transfer possession groups you know
1:05:54	john handle booked ability you will get fifty john a bill you don't put the
1:05:59	he there it's eighty five percent of bill so this is a huge problem in
1:06:04	the literature "'cause" nobody's and warming everybody treats the baseline like it fifty as the
1:06:09	baseline between subject and object
1:06:11	that's not the baseline
1:06:13	the baseline is
1:06:15	the prior
1:06:16	and two there's always confusion could people say well
1:06:19	pronouns are
1:06:20	or something biased
1:06:21	except when they're not i can transfer possession of verbs wonder fifty m and when
1:06:26	they're towards the object like a not device simplistic a value for
1:06:30	all that is wrong
1:06:32	every context
1:06:35	when you give a problem
1:06:37	it a few contribute a subject bias
1:06:40	over
1:06:43	it suggests that
1:06:44	the it's over the baseline
1:06:46	of what the next mentioned bias would have
1:06:49	so the it may appear there's a fifty bias but it's a
1:06:53	it's a strong subject bias because if you don't given the pronoun it's eighty five
1:06:58	percent to the up
1:07:00	it does that make sense
1:07:01	so
1:07:02	this is a long winded way of saying
1:07:04	yes
1:07:05	not only would you want to capture these by things are gonna be important for
1:07:09	your statistical model for all urban context i one computational systems
1:07:15	but also what's really important
1:07:18	in psycho-linguistic work
1:07:20	and you know we talking about this for a decade and you still just i
1:07:24	guess get papers review every year that you sell here's my you know
1:07:29	and adapted normal may happen they have to control for next mention by z having
1:07:33	control for coherence relations none of the stuff isn't
1:07:41	i'm sorry are talking too much
1:07:44	the just like a
1:07:48	with the image plane is a real injuries
1:07:54	i think that the weights consonants symmetrical because both used for use with this here
1:08:01	is model
1:08:02	used to model distribution for the extension to rotation nice right that one problem that's
1:08:13	this was more time
1:08:17	it is perhaps
1:08:20	the problem of using it probably have a museum or any reference maybe it should
1:08:28	probably it is different for the for the speaker so that we find no
1:08:37	something like that
1:08:40	right in which i like watching what i like to
1:08:47	see
1:08:48	this way
1:08:49	the speaker
1:08:51	if a probability distribution o is a real image rain rate
1:09:01	so i missed my and health
1:09:05	you know that's
1:09:07	and discussion and from which is
1:09:12	you guys i and i
1:09:14	o where there's right so that there's two notions of asymmetry here
1:09:23	so than the but the one i was talking about was really there
1:09:27	the production and interpolation are really based on different factor so
1:09:31	i'm saying that even know the speaker could have a model of
1:09:36	the hearer's prior that is not coming endeavour decision to produce a problem i'm saying
1:09:41	that at the at that i that symmetry is not there
1:09:45	now
1:09:45	you're pointing out also that
1:09:49	be here
1:09:50	doesn't have direct access to the speaker's production bias is a he has to estimate
1:09:55	her production by season put that into his interpretation equation
1:09:58	and that could be off to
1:10:01	but
1:10:03	when we have these mix that's where we're not tracking each other right you each
1:10:06	other's reference
1:10:07	it could be due to either of those asymmetries it could be you know i'm
1:10:11	not tracking the discourse right on the speaker's perspective
1:10:14	work could be that the speaker she's
1:10:17	i'm being little of thinking the discourse is going in one direction
1:10:21	and she's taking in other directions is producing
1:10:24	pronouns
1:10:25	based on you
1:10:26	and has a positive and i get i get messed up because
1:10:30	i'm not tracking
1:10:32	the prior right and she's not even using the prior to produce or problem to
1:10:36	begin
1:10:37	so what it could be either
1:10:43	so we take that it is in the next

Conversational Eliciture in a Bayesian Model of Language Interpretation

Keynotes

Andrew Kehler