okay thanks and but i'll and the buttons and then this a joint work with
the next five rate and small integration
so
some present detection on viral are any detection has recently become a popular and of
the problem
and the method of the these detection research is looking into the utterances of what
the point and of utterance and in isolation
however
as we know that sarcasm is a complex phenomena in natural language and speaker intent
is often unclear a list to provide additional context so this conversation context background knowledge
from of doctors
the topic of discussion in section
and that's why some researchers have argued that these are the sort of additional context
has to be provided to understand sarcasm bitter
and that actually triggers to our research here we look into basically to research questions
that far does conversation context held in predicting some present but conference in context is
that one type of context
and can identify what part of the conversation context actually triggering a sarcastic reply
so before going to the modeling let's look into some of the examples that we're
looking you know researcher
so this that we'd
and whether a user is saying that one more reason to feel really great of
what's line and then to put the has text are present which would be used
as a label
now is looking to that with this is
a hard to understand that why this to it is a sarcastic to eat right
but of the same time we see that the street was a reply to one
other tweet with the symbolic can see that at the rate
and
that was the reply to another to it from another user
where it says that plane window shades or open so that people can see
so we can understand that the user is sarcastic about the experience of line here
and that's why to put i would that has text are present
no let's look at a another domain
and this
these the discussion thread or stun it's taken from the internet argument corpus
but this user again
replying to some other user and reading about sarcastic about their reading have it
and think that all your reading too much into by
is to look into the context forcing we see that the point x this way
a lot longer than we because it's taken from a discussion forum posts and i'm
going to the detail of the post but what's the person is saying that you
think that well the what is not sixty million years old it's only two thousand
years old and they bring
definitely different arguments
and the user to is just saying that go you're eating too much into by
the and that's what they're
sarcastic about
so the outline of my job here two days like with fast discuss about though
i don't results are some of the state-of-the-art in sarcasm detection i'll also talk about
but data
and then i'll go over the first research question that can conversation context l in
detection some thousand and then they're seconds question that what the context actually triggering the
sarcastic reply and can identify that
and finally alicante would and a point to some future work
so as explained before there's a sergeant sarcasm environment detection in the recent years
and that of the research looking to see this as a typical binary classification problem
of not be which we we're doing like for an entity tagging or
relation extraction that sector
but there are some of the directions of the research also has come up in
started and the second recently and out just if you talk about that are for
instance right of it all the look into the context in can really characteristic of
sarcasm a then in our previous work in yemen we looked sarcasm detection as a
what sense disambiguation problem
i miss rate all they have looked into cognitive feature set as i tracking feature
for sarcasm data sent to find a lot they look into timbre posts where images
and the text can be combined for a second predictions and yesterday we saw one
poster from a horribly at all that the looked into rhetorical questions where
a lot of sarcastic utterances are actually rhetorical questions
from the respected of role of context
that's all the users have looked into orders context so that they actually model orders
previous x in a previous
post like into it
to understand that if the authors of sarcastic on that
are there is some work in the conversation context also especially balance but they look
into like n-gram characteristics and their performance we don't the context was not much different
and finally we think that combining wall knowledge how would be like really crucial to
understand sarcasm but
and the same time we think it's
much harder problem so for instance if we see this to it
that users thing that driveway that we are having a long time so that means
that
alignment has something to do with green and this sort of all knowledge is very
difficult
to include
in a system
your research we
particularly look into the conversational context here
so we use two sets of data the four cities like discussion four
and that was that it partially released in the lasagna conference eigenvalue rugby at all
and the structure is that you have a sarcastic response and which replies to a
context and vocalic discussion forum post
some of the characteristic of the data that the data was like allocated that comment
level using crowdsourcing and it's a balance set of training data close to five thousand
a post
from the context and response
we also collected data from twitter
and we look specifically into sarcastic with which are actually like previous tweet and when
we
we collected the previous fight using this parameter like at user
and then whenever possible we just collect the full set of the dialogue so sometimes
like one user is sarcastic about out of their tweet and that we use they
can reply to another with so we have collected the full trade
and labels are provided by the authors so that's one different from the discussion forum
where it's not just one delegating here the labels are provided and we use that
have sex sarcasm and has tech sarcastic to understand what to do sarcastic to each
other
again
it sort of on about in balance that but it's still close like two thousand
and thirteen thousand it for sarcastic and on circuits degrees respectively
and more than thirty bucks of the data we had a couple of sentences as
a context
also it should be given as like into it as max for the contextual characteristics
here
looking to the model can conversation context help in sarcasm detection
our baseline
features for svm linear kernel with discrete features
and do that the features with actually a part one very well in previous the
sarcasm detection research and that's where you're using
n-gram features we use like bigram and trigram and unigram we also used two sets
of lexicons like they are lexical for sentiment and
liwc louis for pragmatic features
absolutely like alteration of the sentiments between the point text and sorry a sarcastic post
and then we used a set of sarcasm markers which are basically indicators of sarcasm
a and that's been used in many research and linguistic and communication signs are for
instance like morpho syntactic features so use of various interjections questions examine some signs
and then when you're doing like speech recognition i'm sarcasm the voice into an snr
modulations is like they have like strong features
but one opinion to natural language and people like typing the put like this kind
of typographic feature set this quotation mark a different type of you multi phones
you move these are used also capitalisation such as this what the time showing like
never
and finally a list of intensify record is also shown to be a strong features
in sarcasm detection like the worst degrade based and beta next section
we also experience with the lstm network of which are able to a lower long
distance dependencies
and in our architecture we are using two lstm one read stuff context and one
reads the response
your six print with that consent base variations where first we used a lot and
sentence level attention and that's like a hierarchical model
and in the second we give the word embeddings static and we didn't do any
a bit attention on the words but we only put that consent on the sentence
of the context because we also one two
answer our second visits question that
what part of the context also help in identifying the sarcasm so that's why you're
doing the sentence level only
this is a schematic diagram of our utterancelevel for the sentence a sentence but seconds
and so you can see that the left hand side we have one lstm to
learn the content
and in right hand side we have one lstm to learn the response the real
there is a sentence embedding then you have the hidden layer then we have attention
and then you have the context vector
and what we mean here that we are concatenating the final vector representation from the
context and the response concatenating and then we're passing it to the softmax
a finally we also experimented with other variation the fella stimulus the conditional density and
that was introduced from the deep mine group for the a natural language inferencing tasks
like textual entailment
what we're doing here is that again we are using two lstm
but the response lstm is conditioned on they represent on the context of spam
so which means that
the sale state is for the response is initialized by the final state of the
context and
that's
been shown that it's like really a strong character structure architecture for
nestling in inferencing task
more some more details that we split the data into a dt in general like
training is like eighty percent data ten percent data we used for the parameter tuning
and we use two sets of word embeddings for discussion forums so we use that
standard go would be ten few hundred dimensional what effect model that for twitter we
used as basic model that was been trained on tweets and we have used you
know this is for
no so the results are this talk about some of the macro-average the that allowed
us point out something important how values and
the more i thought the number seven in their in more details in the paper
here first we look into the comparison of the is em results with the response
and context bus response we said that into a terry there is a some of
that some part of the improvement is they're using the context
but not in the discussion forum data and we suspect that
it's might be possible because a lot of and the context of soul all in
the discussion forums and
is v m which are mostly based on n-gram features for not every to learn
that
if you look into the lstm variations we only the we don't that instant model
we see that
the first observation is using context help
for both lstm models here
but at the same time they list in conditional model which was kind is the
weather response was conditioned on the context
it performs reasonably well so we see there is like
five to six percent improvement of if one for the discussion forum and for the
preacher results more
and finally that the results from using attention over the sentences with the response and
with the context and we see that sort the treated it to the context and
response performs past and so the discussion forum it just like comparable to the conditional
now we i'm not showing the results we when we do the attentional what in
sentence because that are phones a little more than that
and what we think there isn't could be that
are we don't have enough training data and when you're putting attention what on the
warden sentence level that are too many parameters to model so in the future we
hope that
well when you're we can put more data under discussion forum at which are especially
to
experiment in that direction and that might prove better
no
we go to the second research question that and the identify
what of the context figure out the sarcastic reply
on what we're doing here that's basically we're looking to that happens and wait and
if the indicate what part of the context
triggers the sarcastic reply here
so we evaluated i mention weight with a crowdsourcing experiment on amazon mechanical turk where
we asked i don't parse that we
poster sarcastic reply on the context from the discussion forum and we have that are
those that are can you identify one or more sentences from the context
but you think that we got the sarcastic reply
and we selected the
dark colours where like really
much higher qualifications so we thought that due to be able to do that starts
we're and with our own the eighty five heats here and we selected only the
post with the context length are between three to seven sentence because we think that
if we put context much longer which will be a hard task for that are
percent they might be interested to do that
and we use the major regenerating and interestingly when you compared to that can wait
we found that
ask for more than forty possible but i'm the sentence that got the highest
attention weight is also
comparable to what that are course the thinking and the put the majority voting for
that
so if you see into the same example that i should be for are regarding
the reading have it up because i will
left hand side of the sentences and the right hand side with show that heat
map
so for the heat map also in the left hand side that engine the weight
and the right hand side it just how we represent the majority voting
we see that the for sentence was selected from that can generate and also but
are cars
but at the same time that can send don't put much weight into the other
sentences for the target still think that
there is some wait for the second
and part sentence
we also looked into to analyze that consent weight in the light of like how
it's actually telling us about sarcasm
and we in the paper we actually discussed what various characteristic of sarcasm but in
the interest of time i'll us basically look into one issue which no less the
context in country
so you can think of the conquered context incongruity is characteristics of sarcasm and
with this very simple example i can say that okay say
i'll going to the
emergency room one every week so there is a sentiment like positive sentiment loss
but there is a negative situation like going to the abundance a room on every
weekend and so you can think of this early in the country
or inconsistency between the sentiment of the situation and they're a lot of research in
a linguistics and communication and also in philadelphia where people talk about this context in
one greedy and the think that
this one of the very strong characteristic of sarcasm
a here in our research we actually found out that a lot of time that
i've been sent especially for that which you're the one example am showing
has actually puts more weight into that includes features here so for instance in this
example of the context which was that
it talking about like advertisement saying right now it just totally do you press
about c of mediocre s
and the response was like all really got see what you're doing
and we see that depressed and mediocre actually got more weight
with the
lots one responses got see so you can see that there is something some sort
of incongruity and there are some more examples in our paper is that we
i showed that how this context incongruity is happening here
so at the same time we also looked into sarcasm markers are which is basically
indicators like explicit indicators of sarcasm
and we seems like especially here i can send a concluding much weights on the
markers such as various multi columns anymore jeez
you can think of and those actually actions and put more weight and
that means the model of thinking that these are like some features
a one interesting thing was we found out that sometime in the context you have
a like cost like anymore tickle like a set face
and then in the response maybe you have a like smiley face
and that attention models actually learning that these opposite characters of emotive content may be
small feature and that would be
a small weight over there so that so interesting observation
at the same data like interjections like are and those words actually have received a
lot of attention which so
but
one of constant here that
however
plus the classification task was not forced only to run on the attention the weight
and it's again
not does it say tall in two thousand fifteen to have discussed
on this topic that
the integration based on any attention rate has to be taken
care taken with here and so you have to be little courses here because
the classification was not only working on this with
so
we're interpreting here and we are mostly like observing the races
so to confuse
we found out that for sarcasm detection when you're using contextual information from the dialogs
it shows like better accuracy
also we i think to identify what portion of the context may trigger the sarcasm
and we ran some experience on the interpretation of the context to analyze a different
attitude of such as a like context incorporate the n
other characteristics or read discussed in the paper
a sort of future work we are interested in to like for large scale experiment
and we hope that there will be like more data released from the discussion forum
that we can use for more parameter tuning especially like into the award and second
sentence level able to attention models
also we saw yesterday on posted at how the comments which are reply to the
sarcastic was where you stand
in this kind of analysis so we also interesting
another thing here to observe you have that the three the data is a self-labeling
sarcastic data so people are
posting to its and they are posting the highest tax themselves
but in the discussion forum the sarcasm was perceived to because you have other annotators
coming and then they are not doing so
we are a very interesting to do this sort of analysis that what's the difference
between cells level and possible sarcasm
and finally we follow the in our experiment after doing a lot of error analysis
that
there are specific aspect of conversation like you more than on that users have used
many times and when we also downloaded or like the mechanical turk or they were
actually able to identify those here morrison ones which could actually trigger sarcasm
but i want model actually couldn't so that's one interesting direction that
we hope that will continue or research into
thank you
you questions
you like this slide on an attention experiment i guess for you had the one
example
a series of tweets and i think you where you compared against a crowdsource workers
this can you in that i think i
this is like for attention or no so this is like a we're showing the
discussion forums one so we experiment with that are course only on the discussion forum
and we
proposed like the sarcastic pose
and also the context for the context for like multiple sentences so we have that
doctors that
okay like
we already telling that this is a sarcastic or stand this the context that actually
has triggered the sub doesn't
so how about that can identify the tweet sentence is more important
and we just like five doctors and they put their voting on the sentences that
they think that okay it's one is more important than is to an effect your
and here the right hand side it's sort of the heat map but it's showing
the method of importing
and then we compare the majority voting with that mention that was and here for
this to be example yes one
actually has got a maximum attention the weight from or experiment at the same time
that are first thing that actually also like doctors think that this of the thing
is that actually
i might be triggering the sarcasm
in this
sponded then
again
so this the macro-average f
sorry the macro-average if once we are showing here between the sarcastic and on circuits
to capture
so we basically to the macro-average if one from the
both binary classes so in the paper we actually shortly
separately all categories but just for the interest of time taken that macro-average report
so c plus are means that we're doing the c is like the context
and the response and the all like all the response model
is that
right this normal questions then that's thanks to john again