Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. After spending the first 25 or so years of my life on the east coast of the United States, including about 10 years in Brooklyn, I moved west.
And while I liked California right away, I also quickly had a problem.
People out here just didn't get my sarcasm.
When I said things like, this phone lasts a whole three hours on a charge.
That's awesome. Or, yay, the Patriots are in the Super Bowl.
Again, my new neighbors took me literally.
They thought I was being serious and they thought I was a little odd and kind of mean.
But really, I was just being sarcastic. If only I'd had today's guest with me to translate.
Dr. Pushpak Bhattacharya is Director of the Indian Institute of Technology, IIT, Patna, and a professor at IIT Bombay.
He spent the past few years using GPU-powered deep learning to detect sarcasm online.
Along with a team of grad students, linguists, and psychologists, Dr. Bhattacharya has advanced state-of-the-art when it comes to sarcasm detection machines.
Pushback, thank you so much for joining the podcast.
My pleasure. So in preparing for this, you know, I thought, Growing up on the East Coast, being from a long line of snarky children. snarky Brooklyn Jews, if you will, I thought I knew all about sarcasm.
But in preparing and reading some of your work, I started to sort of question well, wait, when is it sarcasm and when am I just lying or being mean for no reason?
So can we start with you explaining sarcasm? what exactly sarcasm is and why people, and machines in particular, sometimes have such difficulty picking up on it.
I gave this talk on... sarcasm detection, a computational and cognitive approach.
We must, of course, first start with what is sarcasm after all.
The parent of sarcasm is actually irony.
Irony is a situation where two factors interact with each other and they're incongruous.
So first and foremost, sarcasm is a kind of irony.
What is irony? Irony is an incongruous situation.
So if we say the fire brigade's office was destroyed in fire.
The entity which is supposed to take care of fire becomes victim of fire is an ironical situation.
It is incongruity. which is at the heart of many similar situations, including sarcasm.
The genesis of humor again is incongruity.
I am reminded of this particular exchange between a man who parked a car at the wrong place and got a cheat on the car where the policeman had written parking fine.
So he saw that small note on his car and went and thanked the policeman.
So what is the incongruity here? The incongruity is in the worldview of the car owner and the worldview of the policeman.
The understanding of the situation is incongruous.
The incongruity arises from the ambiguity of the word fine, which has two meanings.
So there is linguistic incongruity, there is visual incongruity, and they give rise to irony, sarcasm, humor, and so on.
Now, coming to sarcasm in particular, sarcasm originates from the Greek word sarcasmos.
So sarcasmos the Greek word means to tear flesh with teeth.
I mean, this conjures up an image of fearful nature.
So sarcasm is always meant to hurt and injure.
So it's an incongruous situation which is meant to hurt or Indian.
So really, I was being mean when I moved to California.
It wasn't just the way people from New York talk.
We're actually mean at heart. I don't know that.
But yes, so it is meant to hurt. But before I proceed further, I must put on record my heartfelt thanks to the team of graduate students who made this deep research on sarcasm possible.
It came in the long line of researchers who worked on sentiment analysis.
I must mention Balmurli, Raksha, Aditya Joshi, Aditya Joshi's thesis. he's on sarcasm, Abhijit Mishra, Kevin Patel.
So huge effort from this long line of PhD and graduate students, Vaibhav.
Absolutely. And many of them. You mentioned sentiment analysis, which is where perhaps there are commercial and other applications for the work you're doing.
Maybe you can kind of step back to that. how you, what you were working on at the time and how that led you into researching sarcasm.
Yes, I think researchers always enjoy going back to the past times and try to discover the genesis of their current work.
We started working on sentiment analysis, I think in the year 2000.
And we were given a problem by one of the companies where they say that people express a lot of opinion about their production services.
And they're interested in mining the opinion of the people, understanding their sentiment towards their work.
So we got a lot of textual data from them and wanted to detect if they had a positive or negative polarity.
And at the time when I went to ACL, EMNLP, AI conferences, People were already talking a lot about sentiment detection and opinion mining.
So while doing sentiment analysis, we found that there were two problems in sentiment analysis. which defied the ordinary algorithms.
The first problem is thwarting. The second problem is Thwarting is the problem where there are lots of positive things sentiments statements about, let's say movie, the movie has beautiful dance sequence, great song. great stage management and so on.
Sure. However, at the end, there's a statement.
However, the movie doesn't agree. So this is a case of thwarting because the overall opinion is negative, even though there are lots of positives.
So this is one problem where we found that the algorithms were showing poor accuracy.
And the other case was sarcasm, where there is no surface level indication of negativity.
However, the opinion actually is negative.
So this is misleading the algorithm. And then we took up the certain sarcasm detection.
And so how did you attack the problem? And in particular, I mean, speak to both, but I'm thinking in particular about this, how you described it.
This problem where the words, the text on the page, the tweet in the tweet stream, taken at face value is positive and sentiment.
There's no way around it. This phone is awesome. but really with context and with the kind of context that a human just sort of naturally develops over time or, If they're immersed in Twitter all day, like certain podcast hosts are, we just sort of learn to pick up on it.
How do you teach an algorithm to detect that?
We started investigating the linguistics and cognitive science aspects of sarcasm.
And we did find quite a lot of work on the properties of sarcastic text from traditional literature on linguistics and cognitive science.
And we found that at the heart of sarcasm is incongruity.
As I said before, Though there were works on, uh, sarcasm detection, there were research groups who investigated sarcasm, but they did not, uh, take linguistic inputs and the text markers which come from cognitive viewpoint and linguistic viewpoint quite seriously or They did not undertake a systematic study of this.
Now, our starting point was this particular sentence, which I'm very fond of.
Uttering in all my talks, suppose a person goes to a party and nobody pays any attention to him.
So while leaving the host asks him, how did you like the party?
And he says, oh, yes, I love being ignored.
Right. The same sentence contains a positive sentiment bearing word love. and a negative sentiment-bearing word, ignore.
So how come these two words of opposite polarity are coming together So that should indicate something interesting or abnormal about the text.
So we dug a bit further. And we understood that sarcasm always originates in incongruity, either in the same sentence or in the neighboring sentences when it is a dialogue.
So now, Investigating further, we could say that incongruity can be of two kinds, implicit or explicit.
Implicit incongruity is a little easier. where the positive sentiment bearing and negative sentiment bearing words appear together and they're on the surface, on the text itself.
I mean, this is the case of explicit incongruity.
That's easier. But implicit incongruity is far more difficult.
And I have. more interesting things to say going forward in the context of deep learning, word embedding, and so on.
So the next implicit incongruity and example would be, I like this research paper so much that I made a teacup outfit.
I think one of my undergraduate professors wrote that on a science paper of mine, but that's another story.
Okay. Okay. So now here, the incongruity is that made a teacup out of that research paper is an implicitly negative statement.
And I like the research paper is a positive sentiment bearing phrase.
So again, the coexistence is incongruous and that is an indicator of sarcasm.
Now, next question of course would be, how do you detect this implicit incongruity?
Here we thought of something It's quite novel, which got a lot of attention from the community.
MIT Tech Review carried an article on this particular work of ours.
ACM Tech Review also covered this work. So what we did was we reduced incongruity to the problem of similarity.
Similarity in the following sense, I'll again take another example, which is there often in my talks.
Most of the examples are due to Aditya Joshi, whose PhD thesis has been on sarcasm detection, he has a lot of insight into this problem.
So this example is a man needs woman like a fish needs a bicycle.
So this was also from a sarcastic tweet.
And this is clearly from a man who is somehow disappointed with some situation.
The clue to sarcasm is the following. The similarity of man and woman is much, much more than the similarity between fish and bicycle.
Here you see the incongruity is now being captured, from using the instrument of similarity detection.
And the similarity of fish and bicycle is low That is detected by computing what is called cosine similarity between the word embedding of fish and word embedding of bicycle. which is the modern-day contribution to semantics or distributional similarity.
So to put in simple terms, we saw an incongruity between the first phrase, man needs woman.
And the second phrase, fish needs bicycle.
And they are linked by a preposition, which is like.
And like, the property of the like as a preposition. is that the two phrases, the two phrases which are to the left and right of like, They must have similar properties.
Now, man and woman have high similarity value, fish and bicycle has Low similarity value, so how come?
I mean, this is a situation of incongruity.
And that gave rise to a very simple algorithm, which again gave rise to very reasonable and nice accuracy for detecting sarcasm.
So this caught a lot of attention of any machine learning community.
That's fascinating. It's breaking everything down into its ones and zeros, so to speak. and analyzing the relationships and how you, you said something earlier about I think you used the words implicit and explicit to me.
I was thinking of levels of sophistication of sarcasm and And being mean-spirited, I suppose.
But how far away can the two... I'm thinking of them as two phrases.
How far away can they be separated for your algorithm before your algorithm can't detect them anymore?
In other words, and this leads to a sort of larger question about how you constructed and trained your models and, you know, you... said that this work on sentiment analysis dates back, or your work, I believe, dates back to 2000.
And I was trying to remember... I was online at that point.
Social media wasn't really a thing that I was aware of.
And so nowadays... When I'm in the mood for snark and sarcasm, I go to Twitter.
But I'm just wondering what types of data and what types of conversations, what types of texts can these algorithms handle?
And in other words, if we're looking at something where there's a long work of text by one author, I would imagine it might be a little bit easier to detect versus... transcript of real-time social media might be a little more difficult.
There are so many important points this question touches.
So I was Just jotting down the important points.
I apologize for I didn't construct the question more concisely.
It's just there's so many factors to this to think about.
Yes, there are a few crucial points which I definitely should not miss.
One is that you had a question about how do you decide the threshold of what is sarcastic and what is not?
Non-sarcastic. especially in case of implicit sarcasm.
Turning a research paper into a teacup is not the... intended use of a research paper okay this is what is called world knowledge which ai has always grappled with the problem of world knowledge and capturing world knowledge and representing world knowledge.
So how is the machine expected to know this, that A research paper is not supposed to be turned into a teacup.
A more difficult case, but very, very ubiquitous case is that of numerical sarcasm.
So you gave an example of numerical sarcasm right in the beginning. where you say that I'm enamored with this mobile phone of mine, which has a battery backup of half an hour.
Half an hour, right. This kind of sarcasm is called numerical sarcasm.
And this is, again, a very pioneering kind of work from us. and its genesis is again in the chain of sentiment analysis, sarcasm detection, and numerical sarcasm detection.
Ordinary sarcasm detection algorithms fail quite badly for numerical sarcasm detection problems.
So how does one know that half an hour of battery backup is bad?
The same half an hour may be good in another situation.
We started working on numerical sarcasm.
Things like waiting for three hours in freezing cold is so much fun.
So, so much fun is again positive, but waiting for three hours in freezing cold is negative.
So what we did was we tackled this problem in what is called first a rule-based manner.
And then we resorted to machine learning and next to deep learning.
So here, I mean, whatever I say now, actually has an important insight and probably an important message also for our community.
We turned all these sarcastic sentences into a small data structure.
So my phone has an awesome battery backup of half an hour.
So here we take the sentence, pass it through basic natural language processing tasks like dependency parsing and a part of speech tagging and number detection.
And from that, we create a data structure which captures all the noun phrases inside the sentence.
And then we have a training, we have a repository of training data where these kinds of sentences are marked sarcastic and non-sarcastic by annotators.
We again collect the noun phrases from these kind of sarcastic and non-sarcastic tweets.
So now, suppose I have a sentence My phone has an awesome battery backup of 30 hours.
In my training data, we have sentences. which say awesome battery backup of let's say 30 hours, 24 hours, 19 hours and so on.
That training data gives rise to a data structure and tells me that there is a threshold above which the value of battery backup time is non-sarcastic.
Right. And below which it is sarcastic. So this is told to me by the training data.
And then I look at the training data, I apply my own mind as a human being, and then I say, okay, the threshold of maybe five hours is what I can take for sarcasm and non-sarcasm.
So we got this kind of rule-based system.
We gave us a tremendous amount of accuracy.
He gave us about 80%, but when we sent research papers based on this work, they always came back rejected.
And why were they rejected? The reviewers always said this is so ad hoc.
How do I know that five hours is the good threshold and not 10 hours and so on?
So we then decided that we'll Let the data do the talking.
Okay, not we, but the data. I was hoping that you responded by shipping them a case of five-year-old phones and let them, you know, experiencing for themselves, but your method was probably a little more on task.
Yes, and if it was not anonymous, we wouldn't be doing that.
Right, right. Keeping the big picture in mind.
So to our surprise, you know, we found that when we used support vector machine, then K nearest neighbor and all these kinds of machine learning algorithms, and extracted features from this text, pass it through a classifier. then this kind of threshold emerged from the data.
And when it emerged from the data, the reviewers sort of said, okay, what can we do?
The data says so, so we are happy. And then maybe some reviewer again grumbled that no, no, no, after all, you are extracting the features and there goes your bias of which feature is important, which is not.
So we said, fine, we'll not do that also.
We'll relegate everything to deep neural network. deep neural networks to automatic feature discovery.
So they took the tweets and convolutional neural network, recurrent neural network, all those usual Techniques of deep neural network, they again process the data and on the test corpus, which is standard, we got about 90% accuracy.
So 80% in traditional machine learning, 90% in deep neural network learning, and rule-based system gave us about 80% accuracy. deep learning based system got more trust from the community saying that, okay, nothing is biased by human judgment and so on everything.
He's from data. We trust the work, like the work, which is fine.
Great. So the message that I'm trying to sort of convey is that These thresholding decisions also have to be learned from the data and the features which read which lead to these thresholding decisions, even that also, if they come from the data, that is all the better.
Of course, right. So during the course of your work, what, if anything, surprised you, whether about the technical aspects of building and training your models or kind of the more psychological and linguistic aspects of sarcasm and irony?
What surprised you? Right from the beginning, everything has been sort of surprising or eye-openers.
We didn't know that people... express so much of sentiment electronically when we started the work.
And then we were surprised to see 11% of tweets are sarcastic.
There is so much of sarcasm. And we always found that political tweets almost always contain sarcasm.
Then heads of the institute, leaders of parties, they're always butts of sarcasm.
I'm biting my tongue very hard and that's all I'm going to say.
And is this global? The languages that we have seen, Indian languages, English, and to some extent, German also.
It's quite a universal phenomenon. By the way, there's an interesting opposite phenomenon, which we call as humble bragging, you know.
So if somebody says, oh, you know, my life is miserable.
I have to sign 500 autographs every day.
And my life is miserable. That we call is humble bragging.
That is opposite of sarcasm. Has that at least that example has that same structure, right?
The two, the incongruity. Yes. Why would not you call it sarcastic, right?
Or I'm wondering if that's also a feature of humblebrags.
Yes, that is. Even humble bragging has exactly that problem of incongruity.
So an interesting question, as you say, is to separate sarcasm from humble bragging.
Here we have to take help of dependency parsing.
We have to identify which is the main clause and which is the subordinate clause, syntactically speaking.
And we also have to go a little further and detect which part of the sentence is main sentiment bearing and which part is additional sentiment bearing.
So that's an important question. And then the other thing that surprised us is the amount of numerical sarcasm.
People use numbers all the time. and talk about sarcastic situations.
So about 18% of sarcastic tweets are originating in number. that is again something which sort of surprised us.
So 11% of all tweets, roughly, are sarcastic.
And 18% of sarcastic tweets are numerical sarcasm.
You know, I'll admit when you said the 11% number, my initial reaction was, that's all?
Really? It's not higher? which I think has more to do with the echo chamber that I live in on Twitter.
I follow a lot of technology journalists and they are a sarcastic lot, let me tell you.
I must also mention this the other element of surprise to other elements of surprise where there is no what is called inter annotated agreement when it comes to sarcasm.
So two annotators when they're marking pieces of text as sarcastic and non-sarcastic. we do see that the amount of agreement between them is typically quite low.
It comes in the range of what is called kappa value.
Kappa value of 0.3, 0.2, which is low. I like this research paper so much that I made a paper bag out of it or teacup out of it.
You know, in this kind of situations, many times people fail to catch the sarcasm.
Maybe human brain is not naturally wired for detecting sarcasm or probably sarcasm requires high level of intelligence.
I'd say so. Yes. There are quite a few cultures who do not approve of sarcasm, by the way.
There are quite a few cultures, there are quite a few economic groups quite a few groups with particular level of education and upbringing who do not approve of sarcasm.
So, sarcasm has a lot of cultural implications and Apart from there being psychological and intelligence level issues with sarcasm.
Now, the very interesting part, which I should not miss is this question of why people use sarcasm.
And here I have a theory which may be interesting at times, could be considered seriously.
So sarcasm is actually lowering a person's defense and then attacking.
Can you explain a little bit? Our audience isn't so smart.
They might not get it. See, that was I was being sarcastic, but kind of the opposite because it was being nice, not mean.
Nevermind, you explain doctor. So first of all, immediate answer to why people use sarcasm.
The answer would be human beings like novelty, dramatization, intensity, and so on.
But a more intriguing possibility is that The first part of the sarcasm, which is positive, it lowers the defense of the listener.
I like being, and now the host is expecting a kind of, praise or some positive word.
I like being attended to. I like being respected.
But Now, bang comes the word, which is negative.
And if you had started with, oh, I hate being ignored, then the person's defense is already up.
And he knows how to cope with the next part of the text, which is now coming.
But I like Defense is low, and being ignored, that hits you very hard.
And that seems like it's a fairly common... tactic or technique in making any sort of an impact.
But, you know, what came to mind immediately was a comedian's Reaching that level of sort of comfort and, as you said, kind of lowering the audience's defenses a little bit.
But then there's a twist. It's like setting them up for... the verbal twist that they're not expecting that brings that drama and excitement and, and, you know, the sort of wow, wow moment.
Yes. And human beings like, I think, twists in the story.
And then you also were asking about techniques when we looked at the techniques.
Did any techniques surprise us? Yes, I think we shared the universal experience of being surprised at the performance of deep learning.
Layers and layers of neurons, they're delivering high accuracies and we are not able to explain why they're working so well.
So we share our surprise with the rest of the NLP community on the way deep learning is performing well.
Did you find over the course of, were you working with the same team or same basic core people on your team for a number of years?
Yes. I mentioned the graduate students.
Aditya Joshi did his PhD with me. It took about five years.
Okay. Investigating primarily sarcasm apart from other things.
There was Raksha Sharma who worked on intensity of sentiment.
Sarcasm, by the way, is very intense sentiment.
So she worked on intensity of sentiment and then, uh, the use of eye tracking is very important, you know?
Hmm. Sarcasm is often detected by physical behavior.
The eye example is a very good indicator of whether the piece is sarcastic or not.
So we always found that people first go forward in the sarcastic tweet and then come back and aggression many times. over to finally understand the sarcasm.
If only this was a video podcast. Everybody could see the sarcastic faces you and I are making at one another from across the world.
But unfortunately, the reason I ask about that. how long your group was together, I'm wondering, as a group or even individually, did you notice yourselves becoming more sarcastic, less sarcastic?
As you went along, did you develop a tolerance or a disdain?
Did you reach new levels of sarcasm sophistication?
What was the effect on you and those around you as you spent so long on this work?
I work on almost all aspects of natural language processing, you know, versus it. very big research group, about 30, 40 people, and machine translation, information retrieval, extraction, mining, everything we work on.
So after, let's say, discussing sarcasm, I moved to a discussion on, let's say, information retrieval, which is a hardcore engineering subject.
At that point of time, there is a huge context switch.
But maybe I should really ask my graduate students who worked on sarcasm, Was there any effect of that on their personality?
So Aditya Joshi worked on sarcasm. No, I did not see him becoming sarcastic.
So he had an objective and, you know, scientist like approach to sarcasm.
It's serious work. I mean, and no sarcasm intended there.
It's obviously very serious, very high-level work.
Absolutely. ironically on a very um what i think of anyway is a very playful and humorous topic which which again probably speaks to the evil lurking in my heart but But we very, very definitely enjoyed working on these kind of problems.
There's a lot of intellectual kick. Did you form a sort of sarcasm network, if you will, with...
Others researching sarcasm outside of your group?
Is there kind of a global who's who of the nuts and bolts of sarcasm?
Not really. We did not find any group taking sarcasm so seriously and investigating this so systematically.
So the paper which we compared ourselves against and from whom we also took some resources, is the group by Rilof, Professor Rilof.
And that is the only group we interacted with.
But we have not seen anybody working on sarcasm very seriously around the world.
We mentioned earlier, I believe, this idea of sentiment analysis.
And it's something that as social media has proliferated and found its way into everything from advertising and marketing and business to politics to what have you.
This notion in this industry, or this industry, I should say, is built up around this notion of being able to use a bot, use software to keep track of what the world thinks of you or your brand or your company or what have you.
Is sarcasm mucking up the works when it comes to sentiment analysis?
Is it... throwing, you know, people are coming away with a skewed idea of their reputation because their tools can't account for sarcasm?
Yeah, this is one element of surprise for us In the sense that here is a very difficult problem, which is staring us in the face.
And which is sort of derailing the conventional algorithms and how come people are still not taking this problem seriously?
So this was a bit surprising for us. One reason why people did not tackle sarcasm in its full extent is probably because it requires a pretty reasonable amount of annotated data.
So you have to take tweets and annotate them, mark them as sarcastic, non-sarcastic.
This is some human labor. And we were fortunate to have a very good team of annotators, experts in language and Their detection of sentiment and sarcasm is really very good.
So they annotated a lot of data for us. Then another advantage we had was that unlike many NLP groups, we give tremendous importance to linguistics.
In fact, we have And at Dej in our lab, we say, Linguistics is the eye, computation is the body.
We are guided by linguistics, cognitive science all the time.
So that orientation is important. So that helped us.
Home on to incongruity and capturing incongruity, representing incongruity in the algorithm as the handle to tackle sarcasm.
So this kind of systematic approach we have not seen.
But people should be worried that their algorithms are not working because of sarcasm.
And maybe you were right in one way that they have probably kept this 11% of tweets to be tackled later.
Which is sarcastic. Yeah, we're always being pushed off to the back of the queue.
We'll deal with you sarcastic guys later.
And it just fuels the fire. It makes us more sarcastic in the end.
Our guest today has a serious approach to sarcasm and not sarcastically.
I agree. fascinated by it and think it's wonderful.
You're doing the work, Dr. Pushback. Bhattacharya, if I didn't mispronounce that too badly, director of the Indian Institute of Technology, IIT Patna, and a professor at IIT Bombay.
And amongst the output of your work to date, and I say you, you and the team who you so graciously have called out during the podcast. is a browser-based engine called Sarcasm Suite.
And also, or as part of that, there's code for Sarcasm Bot. a chat bot that generates sarcasm on its own in response to a question.
So what's next for sarcasm? Are you continuing the work?
Are you moving on to focus on a different aspect of NLP?
What's in the future for you? Yeah, we are going deeper into sarcasm research.
We want to tackle numerical sarcasm. You know, there are many loose ends which need to be tied up.
Target of sarcasm identification is very important.
So when you say, I love being ignored, John said, the target of sarcasm is the host, but the person who is being sarcastic is not me, the narrator, but John.
Identifying from the text which element of the entity is a target of sarcasm is an important problem.
So this is similar to what is called the aspect-based sentiment analysis, which part of the camera you are praising, which part of the camera you are not very happy about.
This is what is called aspect-based sentiment analysis.
Then in the dialogue situation, the sarcasm detection is much more challenging.
The incongruity may be quite some distance apart.
So normally they're nearby, but it could be some distance apart.
Then catching this by an algorithm is non-trivial.
And then we are also very serious about intensity of sarcasm.
There can be sarcasm which is mild, going up to intense sarcasm.
And the problem of humble bragging also intrigues us.
May not be so much in quantity, as it exists in the internet.
But humble bragging is also a phenomenon.
Separating humble bragging from sarcasm.
Because If you just use incongruity, then humble bragging and sarcasm are non-distinguishable from each other.
You need additional machinery. So these are the problems, you know, dialogue sarcasm, intensity of sarcasm, targeted sarcasm, and the other part is multilingual sarcasm.
Sarcasm is very culture and language specific.
People use different image arrays. They use metaphor also for sarcasm.
So these are very nice and intriguing problems.
I, for one, hope you weed out the humble braggers because I think they're dragging the good name of sarcasm down.
Sarcasm, as you said, is... highly intelligent and nuanced form of often colloquial communication.
I'm just trying to make myself feel better.
We could talk about this all night, or at least I could listen to you talk about it all night.
It's fascinating stuff. But for now, unfortunately, we need to put a wrap on the podcast.
If people want to follow your work online, I know you have a number of of papers published and you're giving talks and there have been articles.
There's a great article on the NVIDIA blog, so people should for sure Check that out from just a few weeks ago.
Where else can people go to find out more about you and your team's work?
If you go to Google and give Pushpak Bhattacharya, it takes you to the homepage and under that there are publications.
So there is publication link, and if you follow the link, there are these research papers.
And the Sarcasm Suite is in GitHub. Is that correct?
It's in GitHub. It is also from our Link from our lab, which is Center for Indian Language Technology.
Perfect. well center for indiana googling this will take you to that lab center for indian language technology excellent pushback thank you so much again for making the time and um all All snark and kidding aside, it's really fascinating to think about. the origins of how we communicate with one another and breaking that down into Something that you can use to train a machine to pick up on it.
And it's fascinating stuff. I, for one, I'm going to be. following your work.
It's great. So thank you so much again for taking the time to tell us about it.
Thank you very much for the opportunity.
