Reading Time: 7 minutes

Thanks in due part to new channels of communication. Over the last 20 or so years, texting and instant messaging have changed how we talk. The format is so different, it has sparked new conventions in writing. Teens pioneered new kinds of shorthand, while languages that did not use Latin-based alphabets but did not have their own chat formats invented new forms of transliteration (the Arabic Chat Alphabet is the best example).

But the biggest issue with moving conversations online might be the loss of face-to-face and ‘ear-to-ear’ contact. Catching the intonations, body language and speech speeds of a conversation partner is critical to understanding a speaker’s explicit and implicit intent when talking.

One of the biggest gaps in online conversation is understanding sarcasm. Failing to identify sarcastic comments might actually, in a worst-case scenario, force someone to interpret the exact opposite of a speaker’s intent.

There is actually an axiom addressing this: Poe’s Law, named for Nathan Poe his 2005 comment during one conversation has resonated across the web:

“Without a winking smiley or other blatant display of humor, it is utterly impossible to parody a Creationist in such a way that someone won’t mistake for the genuine article.”

Some have even called for new forms of punctuation like the (SarcMark or Snark Mark) to better identify sarcasm and other kinds of intonation online to prevent such misunderstandings. One could argue that in an era where extreme left-wing and extreme right-wing views are finding new support, this kind of punctual distinction is more important than ever.

But if people are having trouble reading sarcastic intent behind someone’s typed words, imagine the problems for computers.

Sarcasm detection is a complex challenge in natural language processing, specifically in sentiment analysis. There are a few approaches to help machines get this kind of nuance, but none are perfect – yet. In “Automatic Sarcasm Detection: A Survey”, a trio of researchers summarize the major tactics (Joshi, Bhattacharyya, Carman).

Can You Make a Sarcasm Formula?

There are a few ways to define sarcasm. Variations might say it is a type of bitter or sharp kind of satire or irony intended for criticism that is directed at an object or individual in order to 1) mock, 2) inflict pain or 3) convey contempt. It often conveys the opposite of its literal meaning and often is employed humorously. But not all forms of sarcasm are ironic and not all forms are meant to express the opposite of their implicit meaning.

It might have an element of understatement or overstatement rather than one of negation. Inflection of the language and the context of the statement is also important. Sometimes sarcasm and irony are used interchangeably and has to include subtypes like rhetorical questions, hyperbole/exaggeration, jocularity (something done as a prank or joke), and cynicism.

This is just the linguistic part. Calculating this in a way your Python-based algorithms can recognize it for NLP (or sentiment analysis specifically) is even trickier (in INFI’s case, employing sarcasm might also convey something about the user!).

One of the more technical definitions is this from Ivanko and Pexman (Context Incongruity and Irony Processing):

‘Speaker S generates an utterance u in Context C meaning proposition p but intending that hearer H understands p’.’

S = Speaker, H = Hearer/Listener, C = Context, u = Utterance, p = Literal Proposition, p’ = Intended Proposition

Turning this into code is even harder.

Methods for Identifying Sarcasm

Some strategies use statistics approaches to mark sarcasm (or irony in general). The list of features researchers have already tried to study to better detect sarcasm is pretty long (Joshi, Bhattacharyya, Carman). It includes punctuation, emoticons, skip-grams, part-of-speech n-grams, unigrams, semantic imbalance, implicit and explicit congruity, ellipses, hyperbole, and interjection.

And that’s just within 18 studies in a seven-year period.

If anything, it demonstrates that programming machines for intuition like this is still a major work in progress.

Rule-based approaches employ a few tactics themselves: parsing to identify negative expressions within generally positive text or positive-connoting verbs in negative contexts, tracking down hyperboles by using interjection, and simile detection.

Deep learning is popular in AI and NLP in general but has not broken out in sarcasm detection. Identifying word embeddings is one method, while finding a reliable way to define context is still a massive challenge.

Sarcasm Research Twitter-Dependent? Yeah, Right

The field is extremely popular, but there are some limits to where researchers have gone. That being said, there is still plenty of ground to cover for sarcasm detection. The most popular kind of data set to work off of is Twitter, but drawbacks definitely exist.

Most focus on how hashtags are used. There also isn’t much discussion about the effect fake accounts have had on those data sets in light of their proliferation during political campaigns and thanks to private PR/influencer companies.

Another issue throwing off research is that the Twitter API is considered easy to implement, making data collection easy. Thus, it is used often primarily because of availability, in the same way many collegiate psychology studies might be flawed for dependency on using a specific demographic – university students – as survey and test subjects.

Perhaps worse, hashtag-focused studies sometimes depend on tweets being explicitly marked as sarcastic (e.g., #sarcasm, #irony). That precludes text where authors wanted readers to figure out on their own the tweets’ sarcastic nature.

The Future of Sarcasm Detection – So Many Untapped Opportunities

Additionally, most of the research is in English, bringing up issues of transferability to other languages, as well as cultural influence. The lack of research on other languages might be more problematic than usual in NLP.

In speech, languages diverge wildly to express sarcasm and various studies show different sarcasm styles: Cantonese-speakers raise their pitch while English-speakers lower it. Francophones raise their voice and draw out pronunciation. Mexican Spanish-speakers slow down. These are vocal features that might also be cultural and are not going to be transferable language to language, perhaps not even dialect to dialect.

There also is not much sub-classification in the research. In other words, different kinds of sarcasm aren’t pointed out by studies – illocutionary, like-prefixed, lexical, and propositional sarcasm. There seem to be very few deep-learning applications to detect sarcastic language.

A high number of studies are dependent on manually labeling sarcasm or at most semi-supervised machine learning. In a time when deep learning is everywhere, it is kind of surprising.

While some might argue that NLP is coalescing and many basic problems have been resolved, one could argue we have only just begun to scratch the surface.

As research continues, further investigations into linguistic nuance will probably product new ways to break it down. That will drive the complexity of artificial intelligence – not only to identify sarcasm and nuance, but to actually generate or reproduce it. The same tactics will be important to identify rhetorical questions , figurative language, cultural references and more.

References:

(In order of appearance)

AbuSa’aleek, AO. “Latinization Of Arabic Language In The Electronic Communication: Concept And Practice.” Aligarh Journal of Linguistics. Volume 4, 2014. https://www.researchgate.net/publication/286397642_LATINIZATION_OF_ARABIC_LANGUAGE_IN_THE_ELECTRONIC_COMMUNICATION_CONCEPT_AND_PRACTICE

Aikin, Scott F. “Poe’s Law, Group Polarization, and the Epistemology of Online Religious Discourse.” https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1332169

Joshi, Aditya, Pushpak Bhattacharyya and Mark J Carman. “Automatic Sarcasm Detection: A Survey.” arXiv:1602.03426v2 [cs.CL] 20 Sep 2016. https://arxiv.org/pdf/1602.03426.pdf

Ivanko, Stacey L and Penny M Pexman. 2003. Context incongruity and irony processing. Discourse Processes 35, 3 (2003), 241–279. https://www.researchgate.net/publication/247807734_Context_Incongruity_and_Irony_Processing

Cheang HS, Pell MD. “Acoustic Markers of Sarcasm in Cantonese and English.” J Acoust Soc Am. 2009 Sep;126(3):1394-405. https://www.ncbi.nlm.nih.gov/pubmed/19739753

Loevenbruck, Hélène, Mohamed Ameur Ben Jannet, Mariapaola D’Imperio,
Mathilde Spini, Maud Champagne-Lavau. “Prosodic cues of sarcastic speech in French: slower, higher, wider.” 14th Annual Conference of the International Speech Communication Association (Interspeech 2013), Aug 2013, Lyon, France. Proceedings of Interspeech, pp.3537-3541, 2013. https://hal.archives-ouvertes.fr/file/index/docid/864346/filename/Loevenbruck_etal_Interspeech_2013.pdf

Rao, Rajiv. “Prosodic Consequences of Sarcasm Versus Sincerity in Mexican Spanish.” n Concentric: Studies in Linguistics 39(2), November 2013. http://rajivrao.weebly.com/uploads/9/3/8/5/9385302/rao_forthcoming_sarcasm_sincerity.pdf

Orbay, Shereen, Vrindavan Harrison, Amita Misra, Ellen Riloff and Marilyn Walker. “Are you serious?: Rhetorical Questions and Sarcasm in Social Media Dialog.” Proceedings of the SIGDIAL 2017 Conference, pages 310–319, Saarbrucken, Germany, 15-17 August 2017. http://www.aclweb.org/anthology/W17-5537

Pérez, Antonio Reyes. “Linguistic-based Patterns for Figurative Language Processing: The Case of Humor Recognition and Irony Detection.” Doctoral dissertation. http://users.dsic.upv.es/~prosso/resources/ReyesPhD.pdf

CONTRIBUTORS

ABOUT US