0:00:15so first thank you for coming to my presentation
0:00:18in the first part i'm going to introduce a if you a notion of all
0:00:22of the problem i want to tackle
0:00:24in the second part i'm going to show how to model and solve this problem
0:00:29in linguistic terms
0:00:31and finally i'm going to implement a solution
0:00:34in a formal system
0:00:36so here
0:00:38we have a very simple example involving a discourse relation that you didn't come to
0:00:44the party to one out of style so we expected to infer an explanation relation
0:00:49between the two sentences
0:00:50obviously this explanation can be made explicit with a discourse connective
0:00:55jane didn't come to the problem because she one out of style
0:00:59so in english you have three main categories of discourse connectives the first are coordinating
0:01:05conjunctions with and all other than subordinating conjunction with because or even though
0:01:12and obvious instead of the white ball for
0:01:15so one of the phenomena i'm interested in it what we call the none alignment
0:01:21of syntactic and discourse structure
0:01:23so there is an alignment or mismatch
0:01:26when there is no mapping between the syntactic arguments of a discourse connective and the
0:01:30arguments of the discourse relation that is lexical
0:01:33so a very basic example
0:01:35a concern at the bills because they have only one syntactic arguments
0:01:39but you have to discourse arguments for the discourse relation
0:01:43for example blah than jane went to paris
0:01:47and so this
0:01:50for those women under we also consider does not structure all
0:01:54and are often a mobile and of course
0:01:57but i'm interested in other kind of mismatch so to understand that is first look
0:02:03at this discourse
0:02:05jane is very talkative as she told me for example that you had an apple
0:02:09for access to the there is no mismatch here because an example of j being
0:02:13very talkative is telling me about breakfast and so
0:02:17but now
0:02:18i see a trend is are lucky jane told me for example but he won
0:02:22the lottery last week now one is an example of friends being very likely is
0:02:26not change holding anything but that he one of the let's stick to here we
0:02:29have a mismatch and on the line
0:02:32yes in this body blocking the discourse argument respond to a corpus of the syntactic
0:02:37so you have this is actually supporting conjecture
0:02:42so here no mismatch between the j thing that here in europe
0:02:47so them becomes less
0:02:51i think you indicated here
0:02:53but a fragment time for a fragment of parallel data think that you know which
0:02:57will email
0:02:59this kind of sentences are
0:03:01and the light with usually with a mismatch because we
0:03:08the speaker intends to
0:03:11express the contrast between going to pair and not going to t
0:03:15so those high enough examples are discussed
0:03:17like the intention context
0:03:19no relation to the penn discourse treebank so this kind of example annotated with that
0:03:24such mismatch in the pdtb
0:03:28but it has been question by a hard to support their uniformity hypothesis according to
0:03:34uniformity there is no such mismatch
0:03:36according to you do uniformity the them the contrasting for be would be between found
0:03:41when comparing and for jane think that knowledge
0:03:45but the reasoning in that
0:03:47you need if a and b are in contrast you can say that a and
0:03:51someone saying we are in contrast but the problem in that this kind of preventing
0:03:56the times in to extend to all the kind of a discourse relation
0:04:01if an example of a is b
0:04:04you don't wanna say that an example of a is someone saying be the women
0:04:08going to follow the uniformity purpose of this work considering kind of mismatches
0:04:14you have also very similar in a minute with implicit relations
0:04:17a friend did not come to the party jane told me he was taking care
0:04:21of his gets
0:04:25so when you have such a mismatch
0:04:28you have something and income taxes or according to don't listen rambo in i there
0:04:35is a commitment of the speaker to once the content of the report
0:04:39you agree with one jane told it is what we call the religion affected information
0:04:46but for sensor and after there is no official religion faculty but you expected to
0:04:53hedge relations or possible explanation and because explanation is a very the correlation that it
0:04:58may be that it is possible that afraid was taking care six get
0:05:03i so
0:05:05i don't know exactly what is true do we
0:05:10do you like in five when you want to find you necessarily agree with what
0:05:15jane told you or do you
0:05:19at least come to their it's possible i don't know exactly so we're going to
0:05:23talk about that like
0:05:25and see how we can almost constant
0:05:28bit of technology so those mismatches in the whole action verbs like to be used
0:05:34to know all those of people would like to say and tell a these four
0:05:40and with distinguishing between the parenthetical user but it's okay one predicate is not possible
0:05:46of the discourse argument so when there is an on the alignment and on the
0:05:50other hand the and attentional use of a t v so when the predicate is
0:05:55included in the discourse segments so when there is no that the line
0:06:00yes very important not all not all connectives are compatible with a session of alignment
0:06:07for example we are cool lexicalising an explanation you can say that did not come
0:06:13to the party because he was taking care of its good at you cannot instead
0:06:17an evidential weight you cannot say predicting outcomes of partly because jane says he was
0:06:21taken care
0:06:22and the problem comes from the connective because you can do that with an implicit
0:06:27relation you can say for a given that comes partly j said he was taking
0:06:31care of is good
0:06:32the problem comes from the connective blocking their the infinite shown in fig
0:06:39so our we're going to mortars so as to the to kind of other we
0:06:44also is that are distinguished by a high amount so on the one hand you
0:06:51have sensible a direct orders so they show good syntactic integration tool a measure close
0:06:55and they provide some time for local information on the intensity in the matrix
0:07:02for example jane left of the written because you have an appointment this is a
0:07:07sensible close according to like them
0:07:10on the other hand you have paris for and the realtors as they are much
0:07:13less into were integrated to the matrix level and that's function is more to structure
0:07:19the discourse so they can express concession or some but one relation for interpretation of
0:07:25the matrix roles for example in eight
0:07:28fragile estimates for less the rest and even though here not finished eating so you've
0:07:33got some the concession this is a powerful and real close to record
0:07:39so she wants many differences in syntax-semantics probability and so on between those two kind
0:07:45of data is that what we see that another different it was performed so that
0:07:50are compatible with last year's while sensible one or not
0:07:54so we already thing that in the previous example but
0:07:58please consider you know and b goes
0:08:01so you can you can have
0:08:03the intangible like in nine so no mismatch for the case the policy
0:08:09even though jane said he'd and should be one thing that you can have a
0:08:13mismatch like in we fight change the party even though j said you hadn't finished
0:08:17writing his research couple
0:08:19but you can and do that we because so you can have to be like
0:08:24in ten eight contains the partly because jane said she wanted to talk to him
0:08:29but you cannot have any and then eventually likely to be fed into the partly
0:08:34because james as he finished writing is research cable
0:08:41so how could one expenses so well stream no i tell you audio that are
0:08:46middle connective a considered and of course not to what i'm going to say that
0:08:51in fact even conjunctions are not for the structure of the kind of selection mechanism
0:08:55able syntax but with more the that i'm going to model as an unable to
0:08:59the y about syntax
0:09:02we couldn't better use examples of all useful so a connective sse
0:09:07and so i don't want to postulate any syntactic n b greedy in this kind
0:09:13of sentence that sentences so that the if there is a mismatch or not
0:09:18the this syntactic structures same but there is this selection mechanism involved in with the
0:09:24connective that will explain the different kind of interpretation
0:09:29yes and obviously this selection mechanism always different constraints force and for every from all
0:09:34those can find a
0:09:37a discourse arguments must have been introduced by the corresponding syntactic argument but you cannot
0:09:43anything like when you're discourse connective you need to select something that is related to
0:09:47you that syntactic arguments and a central connective cannot they compose a closed headed by
0:09:53a navy so that you to access the report but to parry for one can
0:09:58so with us to a constraint you can really account for the data we would
0:10:04seen that before
0:10:05but a in addition are considered as a something that is related to sequences of
0:10:11a congestion one b conjunction juicy but i'm not going to dwell
0:10:18yes so i'm going to implement that it is now i mean in a formal
0:10:23system called a continuation semantics so it's based on lambda calculus it has been used
0:10:29to model the syntax-semantics interface and
0:10:33also some aspects of the semantics pragmatics interface and that's the kind of it we're
0:10:37going to use right now
0:10:39so and you are then used to model f one for nominal for
0:10:46so in well you got a very standard high priced for proper noun like jane
0:10:51is a function with an argument pete's and individual predicate and j just take that
0:10:57these two of the consulting
0:10:59now for context sensitive
0:11:02so like of the pronoun she
0:11:04you need some expected the context arguments
0:11:09and now p is not true of any specific constant a viable it's true and
0:11:14individual that has to be retrieved from the context
0:11:18you think what people selection function
0:11:21and in this case the output context use the same every
0:11:25for a predefined expressions such as cats
0:11:30what happened these we introduced by a factor the cat is to this
0:11:34and we aren't information that x is a cat in the context so it can
0:11:38be this information is from might be relevant for subsequent enough
0:11:43i don't and the context c
0:11:46is transmitted from sentence to sentence for this the discourse of the operator
0:11:51a very simple example she's with things that we wish additionally get something like that
0:11:58it's quite funny or just
0:12:01the individual corresponding to she in the complexity
0:12:04i think on the whole this proposition
0:12:11so we're going to we use this model of anaphoric expressions of to model the
0:12:15behaviours discourse connectives to do that's it we're going to add the basic type woman
0:12:21from the for propositional reference so that matters the different from the g the choose
0:12:26volume type
0:12:28because these markers are going to be the arguments of discourse relation
0:12:33so i consider any sentence describe a propositional market that is given as argument but
0:12:39can also introduce watermark as for example when there is a number that roles
0:12:45and so now he's we think is something like that you get this condition e
0:12:49liable all the proposition law and what happens is based at that
0:12:54i e is a proposition that use this thing and
0:13:00and of course because propositions control we need a true a pretty eight that just
0:13:05a tad bit hes an argument denotes a superposition
0:13:10is somehow the discourse update it looks like that so what how do you continue
0:13:16to discourse d with the new sentence s
0:13:19so what's encountered here but we introduce a new proposition along with played that this
0:13:25proposition one denotes the true population and that without recognition of s
0:13:30given this the potential market
0:13:33for and maybe such a thing
0:13:35what is important in the we introduce a positional marker you problem that for the
0:13:39embedded everything that
0:13:42e interpretation that s
0:13:44things in prior and we start executing the and the embedded close given prime
0:13:49we don't like that in practice true because
0:13:53i think it does not give them so someone can think something that
0:13:57now the last time
0:13:59in the this pretty big channel for the because
0:14:05so what's importantly we have to augment a and b for the two
0:14:10very first oppositional a quite a that is introduced and passed to the recognition and
0:14:17a propositional a kiwi is introduced and passed to the application of me and then
0:14:21the main property should he is described as an explanation holding between to do things
0:14:27to augment that has to be selected
0:14:30and so you got for instance this term select an appropriate propositional markets in the
0:14:36context the system
0:14:37that is really do not have to satisfy the three constrain the this section six
0:14:42and seven see for a central
0:14:46a connective so they it needs to satisfy the this we construct a for a
0:14:53note that connect points and although it's gonna be very similar
0:14:57the only difference
0:14:58is going to be the lexicalized discourse relation and the selection function is gonna be
0:15:03this cell speech section function
0:15:05that is satisfied only to constrain
0:15:09so if we consider the sentence print came to the party although jane things will
0:15:16think what we get it this a big thing
0:15:20so what happens is we expect that for k and this is okay
0:15:27in trying to be a prime
0:15:30he is think
0:15:32and we have a conversation between holding between those two times
0:15:35and in this particular case there is only one
0:15:40a positional marker related to a in this context so the first argument is going
0:15:45to be necessarily the case for making the possible
0:15:49but for this argument
0:15:51there is actually to a different position long as related to me satisfying the three
0:15:56that the constraint you have
0:15:58maybe so anything that was sick and it prior
0:16:02is applied
0:16:03which is only a few was so it is but two possibilities are available and
0:16:10based on the world knowledge this lp algorithm will be able to decide that the
0:16:16correct one is just a in primes
0:16:20not all there is less than that with since like you so french state on
0:16:25because j if you will see
0:16:28in this case
0:16:30instead of lp you will have to see so about satisfying different constraint and the
0:16:37only possibility for those of an argument you think you primes a little or jane
0:16:41think you will see and in this case it it's and natural interpretation
0:16:46so we can explain
0:16:49this is correct and this is not
0:16:54us a lot about this mapping consequences that i mentioned so if we think they
0:16:59consider a some very elite the actual so before and there exist the true confession
0:17:05holding between two proposition then the to proposition must be true
0:17:10if we do that's we get just from where jennifer actually information
0:17:15this by dollars in mimo
0:17:18but if we decide to slightly change that sounds for the connective so that they
0:17:23introduced hedge relations when then selecting the structural argument
0:17:27we get the interpretation proposed by hands and eyes
0:17:33so what have been done so we found that the distinction between a parenthetical and
0:17:39parenthetical you deserve at this is explained in terms of discourse unfolds and that the
0:17:45distinction between sends one powerful conjunction include specific argument selection constraints of those is thoughtful
0:17:51and we show that
0:17:54the average height or discourse connectives can be implemented compositionally using continuation
0:18:01thank you for your attention and please ask me any questions
0:18:35so we have to mention that this is very theoretical work so i'm not trying
0:18:41to actually implement the system that is going to do that because it relies on
0:18:47like good syntactic analysis and it's very hard to get
0:18:52and plus if you want to analyze the full discourse you need to have like
0:18:56to affect syntactic analyses for all the sentences so that the and so it's
0:19:02two days not practically doable but its aim of this work was to explain the
0:19:08why some connectives a compatible with mismatches wine some not like more linguistic or not
0:19:34so you can is selection function so you can plug any kind of algorithm you
0:19:39want so probably something related to get the resulting system if you're improve the that's
0:19:47so the competition like the conditional computational aspects are going to be related to what
0:19:53kind of specific function you implement for this we can do you gotta you could
0:19:57have a very basic heuristic or input other much texas
0:20:58so the thing is when
0:21:01when you like the way it's and this kind of us a discourse on a
0:21:06light in the literature and i will not have once you infer the explanation relation
0:21:12you're kind of committing to the actual is a possibility of the other two something
0:21:16in particular
0:21:40but in this case you are not going to infer the explanation relation or you're
0:21:43going to attribute this explanation to jane not to use all
0:21:57actually it's less obvious when you have implicit discourse relation we could you can be
0:22:01page could be attributed someone else but if we go back to example with for
0:22:06example like a friend is lucky jane told me for example that he wanted the
0:22:12last we like we last week it seems very hard to say that she
0:22:17not an example was lying or so in this case you will kind of committing
0:23:01i know if you use this kind of children you've prepositional phrases like a according
0:23:06to j so they behave differently but they are really syntactically parenthetical and so you
0:23:11can always i whose for any within a connective
0:23:29but there is there something we're going on when used according to jane because sometimes
0:23:35it seems like the relation is actually needed to a the like you it the
0:23:39whole the input the edge between phrase is coping of a discourse connective but sometimes
0:23:44not it's
0:23:46a bit more complex
0:23:54i've been here sorry
0:24:01no well
0:24:04i know exactly like
0:24:06what would be according to me difference is like the scoping relations
0:24:12how it's going to scope of the discourse relation
0:24:17which is
0:24:18both because is syntactically complete very different so you expect that the scoping relations are
0:24:24going to be different and that's probably what