hello everybody again and welcome to class tokenization
i would like to say
talk about the data driven model of explanation for chat about that helps to practise
conversation in a foreign language
this work has been done known as i was a at the university of maximum
that's why see here to those but
and no one with the different this situation
this is a different kind of data driven it differs a lot from those data
driven approaches that
the first keynote speaker at this conference presented a novel but there's still
level c l what we can do with the data
besides that was the statistical one is just approaches
but first let me
outline what
tends to happen
in the next twenty minutes at first i would like to give you a little
bit more background about to start itself the nets with was written in the people
are just that the it it's a extra lights a premium service for the participants
of the conference
and then i would just there
explain what y dot and the weighted that this way
i will present the data a more of a piece of the data and
just explain all the empirical findings and then we will go to the maybe more
interesting part
for you to the computational modelling to a all the race generalization of these empirical
findings and to the
to the case study psychology implementation case study i will explain why it's like a
and
then we will finish what the overview of the results of the huge to the
big battle field
where it was a time
is started with the
artificial companions would be ideal to the machine that interacts with language learners
just you know as it is an artificial for and to have a kind of
france in your instant messenger
it was two thousand eleven it was before the chequebook around
and then you just and this check whether the context into account at least and
just
right it's start talking
practise for language and this
in this and
the way
but then later i found out okay the wrong on the people what to simplify
things and they work in the area
cold computers i intelligent computer assisted language learning and so these two things are somehow
overlap on the intersection between those fields and
we you can imagine how many people from different disciplines already were very natural language
processing for language teaching
second language acquisition computer science
journal corpus research computational linguistics in general
and don't
on the other hand they so many publications in conversation analysis
which exactly focus on the learner
interactions between language learners one that non-native speakers nineteen speakers
for between only two speakers
and no idea you just look what the query
at one or
one and conversation analysis these buttons to than they require domain we see what
can what within
okay model
because
i had initially this idea of
having a machine that
i don't i
behaves like a language experts in the channel buttons it is not a teacher
because i
do you have a clue you can do about what is not exactly i was
not in table two
so top of what we loaded experiment for data collection because it didn't have any
idea
about
what exactly this person these operators there was a lot that's to me to behave
like a language expert in an informal chat
and i for like the dataset
examples of future work
and text
you can take it for free it's on the language resource a repository it's in
germany
a dataset of truman evident needed only two speaker conversations it's seventy two dialogues it's
about
for now that wasn't
turns
and
that was my treasure
so i to this data and
a lot okay what is the
i met that's of conversation analysis because i didn't have anybody this all
what to look for and that's what they call unmotivated looking
it just look at a guy without any idea what
will you will find
and then you may collections of
interesting sequences of typical
sequencers and then you try to generalize describe prototypical structure of this
sequences
and then
is a computer scientist
i then looked at these prototypes and transform them into grammars and roles
and
sometimes it was even possible to do very simple machine learning
and then i set up this implementation case that i is a call that case
study because you can take a dialogue system
any complexity but i two the simplest one
i took and ai ml-based chat but
that and based language understanding
and so how far can go
just to give an overview of what have found
there are different and directional practice
of how
participants of an interaction can orient to the air
linguistic identities all
language learners or language experts in chat
it includes a different forms of face working of negation where language learners six q
is made matrix you made excuses for they are insufficient knowledge for errors timit health
assessment but that was not real self assessment it was on the very beginning of
the interaction that was more like
you know fishing for compliments
or
they got brace for excellent language learners for their
talk one
during the during this data collection and then what you are far different types like
me to talk about language lorna learning and collaborative learning the people
practice
like in the role playing
i data x m situations for instance or
they compared grammatical systems of their native language used
so it was talk about the language
and then we have this
very prominent type of
a positioning
hum is not obvious are expecting some bins and somebody writes
a different kind of creepy a in this case it was rupiah would linguistic troubles
of still that this also problems
in all grace repair sequences was
because what was what's due to
insufficient knowledge of the foreign language
and the focus of this talk is marked the rat
both their explanations upon request is only one type of text while i'm
one subtype of this one type of all the
possible
incarnations of a language expert
and this is
this is the research objective of this paper so i wanted to create computational models
of interactional practice
where only two speakers of truman in chat the of what troubles in comprehension in
a chat but conversation for learning would native speakers
why conversation for learning because it was an informal chat but it wasn't this
yes the bit the participants met because of the they are status of native and
non-native speakers of the with rubber but together because they have these different statuses that's
why it was a conversation for learning it was not just the naughty a conversation
in this sense
why is that challenging i said in the beginning
i had forty five thousand about forty five thousand turns
and maybe you remember all of elements that in the in his you know talk
a eleven reappears
that there are challenging for speech recognition or core approximately every two and half torrance
i had only thirty
so i can i can forget all the machine learning
and
ideally an example of
these
what i five sequences
so that the data original data are in german let there are a translations
he did not need not
non-native speaker has the difficulty to understand or not the i do magic expression and
how can
request
a clarification how this clarification is formatted it's just repeat
all these
probably might think it's not what
there is no
did you mean how what is a it's just a repetition in the question mark
off the dock
and
this is only one
format of a repair initiation but there are many others
and then
after to really be initiation
the unknown speaker
provides the explanation so we it carries out the repair that the but this is
the prototypical structure
of repair sequence we have what troubles source
which can be everything
it never know what will corset
problem and in comprehension then there is a rip initiation which can theoretically your occur
everywhere even have to silence it has been shown already
and then it can be followed by a repair carry out but it doesn't have
to
and
okay the empirical part
would be
finished in this place
what the what i found was
questioning is the praxis but it was not really my finding i just conform to
what has been found before for oral interaction but it
what the same in chat
and
the right different
devices
specific the in the interaction resources
that we have a unit chat to signal that we have trouble
and there are also a specific interaction of resources
that well we half an hour these pet a disposal
two point to the trouble source also every pair initiation contains
kind of signal and the kind of
reference to the trouble source
only repeat initiations the time talking about l corresponding
to the second position
repair initiations
so it's the first structurally defined place where the other speaker can initiate but they
can still
immediate or delay because it and this is because of the of the specific
structure of chat because we can
just you know have mount multiple threads or
in certain things in between and but
that is they steal the su of all the same type of second position
and
but some of them come directly after trouble source or and some of them a
little bit later
and the this has an influence on the resources that need to be
employed for the area a pointing to this trouble source
then
i am used on this example
there was a repeat as a as a as a as a reference to the
trouble source used
but their own
because we have to deal with non-native speakers
but cannot say that
only
as syntactic i syntactic you can be repeated
i it regardless of unit boundaries so a piece of
trouble source to an can be copied and pasted
so we i do you cannot rely on the on the completeness of the second
syntactic structure
and then
what is very common for all interaction
but it's
i can find it in chat
that when you didn't understand something completely i just acoustically but because it's difficult to
follow the overall talk native speakers of mandarin native speaker sometimes
and
then i is the repair
and just the representation of the troubles source
is it is okay is acceptable you don't find it in chat case you can
just really the everything
but still i have was surprised at
some people really
i read it in the wrong way but it becomes usable not through the rip
in each iteration
but there through
i don't things where people try to repeat so that every time the that things
and you see from these retyping that they we applied
role labeling
and there are there are also
things that are typical for money non-native speakers
and if we have very much from the native speaker talk it's to the design
of the repair itself so it's
it's more about the sense of the word that it explains the meaning of the
word order the meaning of the of the of the use this
yes semantic unit
and their it's less it's less about it
something like functional or a foreign
the intention or something like that it not an intention but with the meaning of
the word was
repeat or explained
for their repair
carry out
the of the
participants you was used a different direction results again
like it just looking synonym so paraphrases
but sometimes they also just
you know use google translate
and translated everything in the native language of the l two learners
not to be added one going out to be funny or something they translated that
really with machine translation and that not explanation
and
or they just the arm
because it was difficult explain some of the phenomenon a like what is that what
is a
lapsed we it was difficult explain than words and they just
pasted linked one example
and then it was clear somehow and
again to the same as a rip initiation survey carry out can be delayed or
immediate but the same reasons
and we have a distinct is
so that it was type of repeat
very pi carry out here is a
and i so i called it's speech reap here if
l where
utterance is unclear
or a longer part of a longer utterance is unclear the and not
every word is explained somehow but
only something that is supposed to be difficult
so with that it is clear that didn't
units in each difficult unit is explained but not everything is rephrased all par for
a store
elaborated somehow
so what we need to know for the chat what's your
and first
what does the chat but
need to be able to
do the same joke was a native speaker do you hear the first to chatbot
needs to recognise we can initiate and then detect what is to extract a trouble
source and then generate a repair proper because you cannot predict
what it will be you cannot just used
scripts for ep is forty packing it needs to be generated from what linguistic database
maybe
and is what i've done so why i just used
dictionary
as the linguistic resources and a field templates with the knowledge from the dictionary
and the interactional resources at which my machine looked where
all these signals that are found in any
corpus and with question marks dishes and
a quotation marks and then lexical and things like unclear or i don't understand
the directional resources not allowed to print the trouble sources include repeats but also just
the adjacent addition because lp initiation may consist of only three question marks and then
only the position of this trip initiation points to the trouble source target it's exactly
the previous turn so these
but for instance this type of
pointing cannot be used in the delay position
for the implementation case study that said i used
and i ml based chat about it was
the program d its name a limb interpreter for german and their use the as
a baseline this german the emails that
we take standard by several categories allow that element is to render the
rip here
carry out
based on the island imaginary
now let's and i added to processors the processors in the in problems
process different tasks and i added to different tax that the law to do with
three pairs that was down explanation and meaning tag why this three because
we have a
two different types of questions
that there are kind of baseline questions to which all the rip initiations can be
mapped it's
apple are questions requiring a yes or no
hence there were it's a content question and out that requires an explanation like synonyms
of paraphrases
which translation and then
i need to distinguish between
two of down
i automatically and that's why all the all the request were mapped only two
to this functions and there's white
i had only these two processors
what does that mean for the linguistic knowledge that we need for not
it to recognize repair initiations it might be sufficient just to have this pattern based
language understanding
and
and determine formats that o can be used to initially a creepy a
can be described as patterns
but
we have still real related nlp problem sets are really hard for either princess referring
expression generations because our pointers to the to the trouble source
are referring expressions
but only the domain is a different one we have don't have the whole conversation
only in this a local rupiah domain low in the local bps sequences what we
need your
and in contrast to
to the other two d or overall problem of they're referring expression referring expression generation
there we are normally nouns and
pronouns i seen as the main the main results for that here we can see
also
entire sentence or sentences or phrases or works because a repetition of labor
points to the trouble source of them were
and
then for the
repair carry out
we can use
as a set their definitions paraphrasing synonyms translations and demonstrations and you know probably that
paraphrasing is a hard problem
synonyms is hard problem
finding it automatically
it's also hard to say if the if you're
to in a
confirmation in a in
i mean exact situation but to use it things are expressed
that's this one mean the same as this one it's hard to say
yes or no just
without specific resources
but not worse
low numbers
is not the only challenge other challenges contingency
so
utterances form as rip initiations can have also different accents on their functions like jokes
or error correction
or rejection of surprise are many others
and that's why
it remains still challenging because i don't have a solution
and
it is so i have i don't only one minute but maybe the time is
over again
so i i'm just i just a finishing we have different
results forget regarding the complexity of rip initiations and their repair carry out i compared
with literature that i us in before with work well by david
that's line and
work from conversation analysis like documents and it by the way we are this for
like described
rip initiation formants
across languages and their own
i think that it's quite
language-independent
and that's why
for me to the most of the most
and
positive outcome of this work was that they can use this model
first the cover other languages and second to cover other domains because definition talk works
in the same way in engineering and model and in every other domains what i
need to explain something
and then
a so i can go beyond duty cycle
application case
just to zooming out ic not
conversational this
method helps understand what's going on in human interaction and help to
ground
our conditional models and them into built on a
but we need datasets good data set of
good quality is really large
but of a specific quality
not
we take a to speech sixteen systems that we want to simulate in the and
so we i want to simulate a
dialogue between line i learner and an artificial friend i want to see first how
it works in a similar thing i cannot take
an interview for that
as an days
and the
maybe we can have
just simple chat bots is amenable waibel product in this case but
if you want to cover everything it's but it becomes
very quickly and a complete on we need all the end of knowledge that
well that the that people had produced you know
to
cover all the phenomena that interest
okay we have the two and half minutes for questions
so i'm also interested in computer mediated human interaction
and i wonder if did you serve in these interactions some kind of the interleaving
of comments
"'cause" i imagine that would be the problem with two humans having a conversation over
messenger rather than a human robot because they're we would be more interleaving
in like the manner that people do in
spoken conversation
a about fitting
how much interleaving is there between the utterances of your computer mediated dialogs and rt
similar to spoken
it was between can be that can be eliminated in spoken
is there a lot of interleaving of
i didn't compare datasets i only compared what i found to go to define it
will findings that are described in literature
okay
and the
the right
there are things that i the same
like
formats all replay initiations
some of them are the same as an oral interaction
but
the because we have different directional resources available in chat
we don't have the prosody for instance we don't have data the phase we don't
have the voice
i am that they are they are somehow replaced internet by other things like a
motion a multi consider and instead of laugh
or
when determining when you want to twenty participants wanted to emphasise something
they made uppercase a word stretches
or i had
one example
the data collection that took place in two thousand twelve a what is it
european some cocoa
football cup and the at this time and sometimes participants that just typed at the
same time or in front of their t
and watched again
incremented
and that's how i don't know the word german work goal for what
sixty two holes in the high
and this is really what you say well what the what a reply data and
then to relative to these oral while and when they screen
and so it's
i would say
there are the same things but the expressed by different directional resources
that's the first thing and the other thing is some of the things
cannot be replaced because they become
irrelevant because
but don't we don't have the voice for instance because that's why i didn't find
any repair initiation that require the repetition after that because it's not necessary you can
read everything but these are two differences that i would describe
okay so what to do this
one of things you just informed about the database is that it's montague that you
mentioned supervector doodle for one straight or something but effective some just curious with mobility
longitude minimum assessment luminance
perhaps not increase the learning used to do something like that because but also potentially
useful project work because you want so what kind of increasing importance density distribution of
the material is thanks so that evaluating was not the focus that just adding but
that
whistle normally when you talk about a talk about learning or at least
i'll with this is second language acquisition theory to your in the background
well normally people look at error corrections as a sign for learning
or any kind of a meeting negotiation a call it may negotiations all these repair
sequences that it or are we explain to date technical it meaning negotiations you know
what
and then
this may be costly also obvious
normally only these two things are an online but
i so also
the learning for all
i'm sorry at
i forgot the word
in this
but they wouldn't but there is just the
the null something or didn't use a structure and then
based on example of that and repeated that without any rate wer so that you
want to say that but
you know but
not observation but
making likely making a native speakers
and then and then no i found also that
they learn from implicit corrections which are really hard to capture which are normally not
use the bathroom research
or not the not no not the that they are not use the use of
the wrong word but normally people don't pay attention to that because it's not evident
enough it's not
a node in there is no evidence that people butlers notice these corrections
but i have evidence that
in the data
because they've repeated things that have been corrected through implicit embedded corrections later in later
sections for instance then repeated that's an incorrect wait for it
it's more than just
i'm afraid that drifted a little different direction
anyway changes over time so that the why i explain this thing with artificial companions
in the beginning i guess that posterior have these artificial friendship knows the user and
userspreferences and everything and
and that's why i set up to study the data collection in this way that's
why i'm talking about a specific speech actually systems every participant of the study was
put out every load it was would wherein appear within a speaker and they directed
in pairs for a longer time
and i wanted to see the development
and i can say
the development in learning was not only because they interact longer but because some of
them engage in these corrections and in this evident obvious selling sequences in the beginning
and that's why it developed somehow more intensively later and in either appears it was
not relevant
they just
so i don't have a so we can continue offline five minutes once it uses
this isn't the speaker