How to detect sarcasm in texts
I received an email asking how I would go about the detection of sarcasm in texts. I have a long standing interest in this topic, but only from the side lines, as I contribute to other NLP tasks (sentiment analysis, topic detection and recently, identification of emotions in text). A lot of papers on sarcasm have appeared but I didn’t review them. I’d be interested to see what approaches are explored to date. The following is how I would approach the task:
I’d define sarcasm as the voluntary act for a locutor to:
- convey an implicit meaning which is different (and often, opposite) to the meaning which is explicit in the literal message
- with the effect to create a kind of “dark humor”, at the expense of the event or entity which is the topic of the literal message.
See below for an example involving a response by Elon Musk to a tweet by @RocketLab360.
Sarcasm is, in my opinion, a case that is difficult to approach by machine learning. Indeed, a sarcastic connotation is revealed by very subtle clues, which would not be easily picked up by a model.
On the other hand, these subtle clues are detectable if we examine carefully:
- the different semantic aspects of the text (punctuation, vocabulary, grammatical structure …)
- the context: what comes before the text under examination, what comes after it, or even the speaker’s profile (I expect the Twitter account of the United Nations to be less susceptible to sarcasm than an account coming from entertainment).
On the semantic aspects: I am a priori confident of the fact that many sarcastic sentences leave semantic traces of their connotation - it is rarely 100% dependent on the context. Indeed, it is the subjective experience I have of it when I read sarcastic tweets. By re-reading them carefully, we can distinguish objectivable characteristics:
- sentence length (strongly positive sentences which are really terse can be a hint of sarcasm.)
- use of punctuation (“…” can be a hint of sarcasm)
- some vocabulary markers (“really” is a reinforcer, but “reaaaally” can be a hint of sarcasm)
- excessivity / exageration: when the markers of positive sentiment or emotion tend to pile up in a short sentence, the explicit meaning can be a strong praise, but it can also be one of the mechanisms that points to sarcasm.
An example of a sentence which combines all these traits would be:
Clearly, the guy is reaaally a genius…
An algorithm detecting sarcasm should score a 100% certainty on this, and it is not difficult to build such an algorithm.
So my approach would be labor intensive: read a lot of sarcastic-type sentences (difficulty: where to find them easily?), and reflect on each of them. What aspects of the sentence contribute to my subjective opinion that the author is being sarcastic? And then I’d try to codify those semantic traits as I did for Umigon: creating lexicons and their heuristics + sentence-level heuristics.
I understand that one can be uncertain about the effectiveness of this approach: is it really possible to arrive at a list of semantic markers for something as subtle as sarcasm? My response is to turn the table on this argument. When or if sarcasm is impossible to pinpoint semantically, then a computational approach cannot get a grasp on it… but neither can humans.
To put it differently: when sarcasm is entirely context dependent (no semantic clue available), then machines and humans alike tend to be unsure about the sarcastic character of the utterance. This shows that the computational approach has not hit a limit, but instead that the person who made the sarcastic comment used too few markers to make its intentions (sarcasm or not?) entirely knowable.
An excellent example of this is an exchange between Elon Musk and an unofficial account sharing news about Rocket Lab on Twitter, dating from Oct 1, 2021. The account “Everything Rocket Lab” tweeted:
A small Neutron update before the official big one:— Everything Rocket Lab (@RocketLab360) October 1, 2021
- Rocket Lab is still in the final downselect between the ultimate launch site according to Peter Beck.
- Neutron has become a bigger rocket! The rocket is now 46 meters high with a 5-meter diameter fairing.
📸: Rocket Lab pic.twitter.com/S51bx9x4J4
To which Elon Musk replied:
Will be Falcon 9 size sooner or later— Elon Musk (@elonmusk) October 2, 2021
In Elon Musk’s response, there is no clear semantic hint indicating sarcasm (I guess the “sooner or later” + terse comment convey it, but this is not deterministic). If there is sarcasm, it is all context dependent, which made it hard for the observers of the exchange to be sure.
So, sarcasm or not? (I personally think this is sarcasm but could be wrong). Here, a computational approach wouldn’t detect sarcasm, but neither would humans with any degree of certainty, because the context is pretty specialized (difference in development between SpaceX and Rocket Lab). So in terms of accuracy, the detection of sarcasm has a bar which is the level which humans are capable of detecting - and this is lower than in other NLP tasks (sentiment analysis, emotion detection, etc).
Still, how could the context be appraised in a computational manner? We would have to develop a tool which would be “platform specific” I think. Typically, a platform that allows reactions in the form of emojis would help, as one could rely on commenter’s reactions such as “😆” or “😬” or “🤣” or “🤦” to signal that the content may be sarcastic. Apart from this, we could fancy more complex approaches (characterizing the profile of the locutor as “prone to sarcasm or not”, but that would be a refinement, not a first step in my opinion).
So that would be my starting point! I’d be happy to exchange with researchers and students interested in the subject. If you are interested in my projects in text mining and graph mining, have a look at Nocodefunctions, a free web application that makes it easy to test and use the tools I have developed in the past years.