right name of lena
i am from the natural language dialogue systems
you see characters
i'm presenting the paper modeling linguistic and first now an adaptation for natural language generation
i
each our view
in the parse tree and marilyn walker
and that the first author is unable to attend its content could be the issue
linguistic adaptation is the phenomenal
and human can first thing in dialogue adjust their be a here is to their
conversational partners
it is also referred to as entrainment for me
here is an example
of linguistic adaptation in the dialogue
if one person said this is why make a less specific apply all users to
specific expression are make a lot
to express the direction
which is considered in certain
feature of an utterance
we call this utterance the prize in response is dialogue are there
can either choose to that on this feature i think yes make allow a specific
avenue or not so that i think yes sir that is the fact that until
we call the response utterance
the target
we will use prime and target rat presentation
linguistic adaptation can happen on different features
for example words referring expressions syntactic structures and discourse markers
previous research has shown that there exists on the correlation between adaptation and task success
and that users prefer
adaptive computer agents
with the right of the intelligent
computer agents such as and of on the lattice a realistic stint in theory it
really need to adapt well certainly make user experience even better
however how much adaptation is needed in natural language generation remains an open question
this paper attempts model linguistic and personality adaptation
for natural language generation
it challenges of this work are first of all we don't really have a good
measure
for turn by turn adaptation
second while most work in this area used human annotated corpora we want to automatically
measure adaptation without human annotation
there
we also want to work with different feature subset
especially feature sets that reflect linguistic style such as lexical and syntactic features hedges referring
expressions and personality features
that's we propose a new venture dialogue adaptation score yes
comparing it with other measures on data requirement
can be computed on-the-fly
almost other measurements
require complete dialogues to calculate
in terms of linguistic features ta yes measures adaptation of feature sets
almost
measures focus on single linguistic features
the goal that is to model adaptation and dialogue progress
whereas most measures thing to find out
what happens after dialogues complete
here's the definition of the s
any dialogue turn can be considered as i
and target if it's following dialogue turn i different speaker
we form a prime target pairs using window size one for most of our experiments
which means only one target dialogue turn back to the prime is considered
in the following equation and it is the number i'm target her s u i
p the number of features in crime of the prime target here and that
a p to the t is the number of features in both prime and target
we can use we describe da
as follows
the numerator is the number of features but i am and target and the denominator
is the number of features and right
within a feature set yes reflects the average probability of feature is in prime better
at that point
in target across all prime target paris
here is an example showing how the a s works
you have to be calculated with different adaptation target
if we used articles
for the dialogue excerpt on the right we have read prime target source
and gas is calculated as follows
using that word based features if we used articles e
there is only one prime target pair f one hundred and d one o one
we apply yes one for dialogue corpora the first one is our walk corpus are
walking the pedestrian direction task corpus or director instructs the follower define public or in
downtown santa cruz the telephone conversation
there are sixty dialogues total approximately four hundred fifty turn each
i walk in it or is designed to test series of adaptation in the y
the second is the walking around corpus
walking around can also work that's free interaction
time task corpus
where director instructs follower
to eighteen destination on campus yourself well
there are thirty six dialogues in total
the hundred and seventy five to eight hundred and eighty five turn each
is the director combat destinations together with labels in pictures of destinations
destination labels are provided to participants
for example twenty however memorial
and shown in section
this is a corpus designed to elicit adaptation behaviors
the third corpus is the map task corpus
widely used in much previous work
a task is a collection of cooperative task oriented dialogs where director instructs a follower
to reproduce a route on a paperback
error hundred and twenty eight dialogues total with thirty two to four hundred and thirty
for each
the map says
a line drive the landmarks with labels for example are there
the last corpus is the switchboard corpus a collection of topics and
direct
it's one thing is telephone conversations an example topic is to find out what kind
of the other colour path
switchboard had very skewed dialogues where two speakers take turns in the main speaker
for example a parsed first
be back channels than a and b switch
a dialogue a
annotated subset in
available for switchboard
them example the annotations are
e for gas pedal that the for statement opinion
as shown in the sample here
there are over a thousand dialogues and the subset with fourteen to three hundred and
seventy three different each
some theories discourse defined dialogue turns as extending over backchannels and we think that is
but a lower ones
to measure adaptation more faithful
so we utilize as the u e d dialogue act
tasks
to filter turns that only contains backchannels
filtering process removes forty point one percent of original dialogue turns only twelve point six
percent of the works
next i will introduce the feature sets we use well-matched previous work on measuring adaptation
use human annotation here we use automatic annotation and feature extraction
it's the end for nlp we extracted unigram lemma for example building and n
and the sample dialog turn and bigram i'm up for example the brick
the thing syntactic features
i taking all subtrees from a parse tree
for example and he structure shown here for than a building
we referring expressions by taking all subtrees with route and he
and removing articles for example a building
we build the dictionary of hedges in discourse markers for example you know and like
we also use linguistic and yuri enquiry and word count
the
to extract higher level features
we automatically classifies words in over at linguistic and logical and topical categories for example
a sample dialog turn has words in categories second or skin and informal
previous work has also shown that the big five personalities correlate highly with certain look
categories for example
extra version correlates highly
i think the motion in swear words
use all my personality
luke feature sets introduce our previous work
we performed for experiments using in the first experiment we validate our measure
i comparing da f-score is of original version is randomized dialogues
randomized dialogue turns within speakers
we calculate the is scores for each dialogue starting with lexical features
and had other low-level features
syntactic structures referring expressions and hedges and finally at the look features
we performed a paired t-test on all pairs of original and random dialogues
scores
the table shows the average dis scores of different p sure that in four corpora
numbers in bold indicate a statistically significant difference in the paired t-test
in almost all paired t-test
scores the original dialogues for significantly higher in the randomized dialogues
the results indicate that the gas measure sensitive to dialogue turn order
in the second experiment
we aim to compare adaptation models across corpora and feature set
you calculate average gas scores for each corpora and feature set results are shown in
the table
comparing columns we have adaptation models across corpora
one or walking around have similar yes scores which indicate that the yes is able
to reflect the similarity of that you corpora
to map task has higher lexical adaptation
and lower personality adaptation
possibly due to more controlled experiment setting
three switchboard corpus have the highest adaptation scores
despite the fact that switchboard is filtered
it is also possible that the social chitchat setting allows why adaptation to can occur
comparing rows we have adaptation models across features
we have highlighted the is course
well they're linguistic features are largely contents if it consists of higher level features that
have a broader categories
so it's high da f-score
are extracted
gas scores for the limit feature range from zero point one forty zero twenty nine
followed by syntactic structure hedge and bigram
referring expressions have the low with es score
among personality features emotional stability agreeable this and in its to experience
traits are adapted more than extra virgin conscientious this
it yes that are shown in the table here is the same are more column
in the previous table
adapted models
and that adaptation model learn dialogue corpora can be represented as the is there like
is in used to control adaptation
e javier's in different nlg architectures
in overgenerate and range we can calculate it da is better
for every generated response and response with the shortest distance
to learn adaptation model
it statistical
our timit pair parra materially sorry error
you know right nlg we can use yes scores as probability based generation parameters
in your not we can encode
where an adaptation model into context catcher in control general generation behaviors
in the third experiment we explore fine grained adaptation models
first examine adaptation
i dialogue segment
i along the segment at feast one number of targets in the first three taskoriented
corpora
switchboard is simply segmented into five parts you can you yes for each segment in
taken averaged across all segments in the same position
in that you features on top of our work
corpora corpus scores for the lose features that slightly decrease of dialogues progress
well expiration two shows a distinct increasing trend
despite being a subset of look
this indicates that as dialogues progress
speaker six the bit more adaptation
behaviors one extra version features
however in switchboard both feature sets display the same decreasing trend
possibly due to the set of the switchboard
and the dialogues progress conversants have less discuss about the topic and are less interested
we then examine adaptation by speaker you compared yes for different speakers
in the figures red lines are da you have scores calculated using target
pupils director or color in switchboard
blue lines for target equals follower or receiver
in general red lines are higher than blue lines which means dialogue initiators
using the bit more adaptation
please refer to the paper for figures other corpora
the last experiment examined the trend of gas course is the window size increases the
window size and mean that a prime is pair and target after it
we begin with the window size of one in gradually increased it to five
is similar to the first experiment we calculate the average gas score for each window
size
in comparing here scores between dialogue in their original order pizzas dialogues with randomly scrambled
as
results show that original dialogues shown in red
lines
e f scores decrease its window size increases
randomized dialogs shown in blue lines yes scores stay relatively stable
it's result is consistent with previous work on repetition tk measures
do you conclude this paper shows that when the model for adaptation the and
on the features that used to different
conversational situations can have different adaptation models three the level of adaptation varies according to
which speakers
and the initiative and for the degree of adaptation varies over the course of a
dialogue
and decreases as the adaptation window size increases
this research was supported by the following grants
if you're time first author is on tonight will be happy to take any questions
you
although not shown
right speaking
and that'll
okay
no i
i
three
one question i wish i
i heard things about four
functions and
you know that
my
is that the first thing about the results reported question
i think a separate
in our case or
o
where r is
the way to extract
you are so that one
if you are information for example very well
you are
there are a
well i don't
mister well i
every part three with and
three
and not everyone that
in a in part so i think i want to use l o l o
or a reprint
adaptation paper we are looking into extract
where
like a whole
another problem of paper
you are
at this work this actually proposed work what working all to infinity the idea
language generation and was you have reached a whole
so
you need to find a match you change the order of it
do you have any
finally a match with a human rights respect people have done
one
your question
what i right there is a war
my question
you need a metric dress i rearrange order
we don't know
if this metric
three what humans would say about whether there's adaptation
so you have any ice metric right
sorry the problem is in most of my understanding is your question
so i rated differently for human evaluation
if a question of or
always useful like that
but not