Why Is Sarcasm so Difficult to Detect in Texts and Emails?

This sentence begins the best article you will ever read.

Chances are you thought that last statement might be sarcasm. Sarcasm, as linguist Robert Gibbs noted, includes “words used to express something other than and especially the opposite of the literal meaning of a sentence.” A form of irony, it also tends to be directed toward a specific individual.

However, it’s not always easy to figure out if a writer is being sarcastic – particularly as we march ahead in a digital age that has transformed the way we communicate, with texting, emailing and online commentary replacing face-to-face chats or phone conversations.

In writing, the signal of sarcasm can be muddied. For example, say you’re texting with a friend about meeting at the movies:

Friend: I’m waiting at the front. Movie starts in 5.

You: I’m on my way now. Should be there in 10.

Friend: I’m glad you were watching the clock today.

Was the friend being sarcastic or sincere? The later you are, the more upset they’ll likely be, and the higher the probability their response is a sarcastic jab. But if your friend knows you’re usually much later, they could be sincere.

So there’s one thing to look for: How well does the attitude the writer is conveying agree with the situation and the person?

Nonetheless, the struggle to interpret written sarcasm is real.

Studies have shown that people realize that they have a tough time interpreting sarcasm in writing. Studying the use of email, researchers found writers who think they’re being obviously sarcastic still confuse readers.

Sarcasm thrives in ambiguous situations – and that’s the main issue.

When delivered in person, sarcasm tends to assume a cutting, bitter tone. But written messages don’t always get that attitude across or give you much else to go on. We still need more information.

Signals that go missing in texts

Studies have examined the use of sarcasm in a variety of everyday situations, whether it’s at work to give criticism or praise, or in situations where social norms get violated. (Be on time to movies, people!)

The problem is that a lot of previous studies of sarcasm have been done on spoken sarcasm, which tends to give listeners cues.

When you have a conversation with someone face-to-face (or FaceTime-to-FaceTime) and they say something sarcastic, you’ll see their facial expression, and they may look slightly bemused or tense. Equally or more helpful, the tone of their voice will likely change, too – they may sound more intense or draw out certain phrases.

You’ll also be firmly grounded in the real-time context of the situation, so when they say, “Man, nice job ironing your clothes,” you can look down – and see your wrinkled shirt.

All of these cues have been researched, and we know enough about them that we have the ability to artificially make a sincerely spoken statement sound sarcastic.

And yet when we text, a lot of that information goes missing.

There are no facial cues, no vocal tones and maybe even a delayed response if a person can’t text you back immediately. And if you don’t know the person all that well, there goes your last potential cue: history.

Emojis to the rescue?

So after what you thought was an unexceptional first date – exactly how do you interpret the following flurry of texts?

Date: I had a great time. (12:03 a.m.)

Date: That was the most fun I’ve had in years. (12:05 a.m.)

Date: Really, it could not have gone better. (12:30 a.m.)

Was the date really that good? Did they really seem like they had that much fun? Or are they just a jerk lamenting the wasted time? All valid questions. And the recipient could come to a lot of conclusions.

Fear not. The digital age has developed some ways to mitigate some of the tortuous ambiguity. You can probably include an emoji to make it clearer to a reader something was meant sarcastically.

Date: I had a great time. (12:03 a.m.)

Date: That was the most fun I’ve had in years. (12:05 a.m.)

Date: It really, could not have gone better. (12:30 a.m.)

Ambiguity reduced, and facial expression taken care of. Probably not headed for date #2.

If we’re talking about email, we also have modifications that that can be made to text. We can italicize or bold words to change the way that a reader interprets the message.

Lastly, social media platforms like Twitter have given writers even more tools to allow people to communicate their intent. A study that included sarcastic tweets found that tweeters who include the hashtag #sarcasm tend to use more interjections (wow!) and positive wording for negative situations in their sarcastic tweets.

Algorithms have actually been built to determine the presence of sarcasm and rudeness in tweets, user reviews and online conversations. The formulas were able to identify language that’s outright rude pretty easily. But in order to correctly detect sarcasm, researchers found that algorithms need both linguistic (language) and semantic (meaning) information built in.

In other words, sarcasm’s subtlety means that the algorithms require more specification in their coding – unless you #sarcasm, of course.

The ConversationWith so many options to choose from, it’s time to make sure that text you send at 2:30 a.m. really gets your point across.


Sara Peters, Assistant Professor of Psychology, Newberry College

This article was originally published on The Conversation. Read the original article.

The Voice of Confidence – How Listeners Decode a “Feeling of Knowing”

By Xiaoming Jiang

As social animals, human beings possess a unique ability to communicate social intentions—such as a feeling of knowing—in their tone of voice. Vocal confidence can be used as a signal of one’s persuasiveness, expertise, and trustworthiness. However, scientific investigations of vocal confidence are relatively scarce. As a postdoctoral researcher in the Neuropragmatics and Emotion Lab led by Dr. Marc Pell at McGill University, I am working on an ambitious project investigating how the human voice can encode a speaker’s feeling of knowing, and how the human brain has evolved the ability to decode this complex mental state.

Perceptual and vocal attributes of voice of confidence

In a recent communication on social prosody (Jiang & Pell, 2014), we reported on a newly validated vocal database of over 3,600 expressions produced in North American English. We invited 6 professional actors or public speakers to record statements intended to convey that they were very confident, close to confident, or non-confident (in addition to a neutral comparison). These expressions were led by verbal probability phrases such as I’m confident, Mostly likely, or Perhaps which were congruent with the tone of voice in the expression. These cues were used to help speakers produce the appropriate confidence level, but the recordings were edited so that listeners could hear versions of the statements both with and without the explicit linguistic cue. A group of 60 native speakers ratings of the intended levels of confidence tracked the speakers’ intentions with high accuracy. The average confidence rating was lowest for non-confident voices (2.3), followed by close-to-confident (3.5), and neutral (4.0), and was highest for confident voices (4.5). We found that the confident voice, as compared with the non-confident one, was characterized by specific acoustic signals: a lower pitch, reduced loudness, more flat intonation, faster speech rate, and more restricted change in loudness with the unfolding of the vocal expression.

How fast can we decode the voice of confidence?

Our lab followed up on these findings by investigating how these differences in perceiving confidence relate to neural signals. We recorded electroencephalograms (EEG) while listeners decoded vocal signals of varying levels of confidence. When only vocal cues were available (the linguistic cues like “perhaps” were not included), there were neural differences in the confident and non-confident voices after only 200ms. These neural signals appear to reflect increased attentional processing for the confident voice. The close-to-confident voice received more continuous attention over the course of the statement, as reflected by a larger positive wave evoked at 370ms when processing this voice compared to the more definitive confident or non-confident voices. Neutral-intending voices tended to be rated as reasonably confident, and neutrally they were best distinguished from other voices in a late positive wave occurring at around 900ms after the speech act. This may reflect a mismatch between the task goal of rating confidence and the speaker intention—not to include information about intention (Jiang & Pell, 2015).

We were also interested in how vocal and verbal speech signals interact. We found that, when vocal cues are combined with verbal cues, the differentiation of confident and non-confident voices occurred earlier—at 100ms. Compared with vocal-cue only expressions, neural responses to expressions with combined verbal and vocal cues were greatly reduced. Verbal information thus appears to facilitate the processing of vocal information for inferring speaker confidence, leading to less heightened neural activity but also to more rapid neural differentiation between varying levels of confidence (Jiang & Pell, 2016a).

How can we resolve conflicting message in speech signaling feeling of knowing?

We examined how lexical and vocal cues can facilitate each other, but we were also interested in how people decode mixed messages—those where verbal and vocal cues conflict in conveying feelings of (un)knowing. We reported two different cases, where a listener’s brain appears to decode with distinct strategies (Jiang & Pell, 2016b). Hearing I’m confident followed by unconfident voice places a heavy burden on the integration and update of vocal information into the verbal context, using brain structures dedicated to conflict resolution. Hearing Maybe followed by confident voice induces delayed inferential mechanisms, using the brain structure focusing on perception of other’s hidden intention. Thus resolving conflicting messages regarding confidence depends on what type of conflict is being resolved, with distinct brain regions playing different roles in processing depending on whether the confident signal is preceded or followed by the unconfident signal.

Females are better at confidence decoding?

We also found a sex difference in decoding feelings of knowing from the voice. Female listeners in our study had stronger sensitivity to vocal cues of confidence, rating confident-intending expressions as more confident and non-confident-intending expressions as less confident than male listeners. We believe that these sex differences are mediated by individual differences in the neural responses to vocal expressions or by differences in personality, such as trait anxiety and trait empathy. It appears that female listeners engaged in very early acoustic analysis and made delayed social inferences for complex messages. In contrast, male listeners detected the relevant social information in the voice and immediately changed their levels of attention. A statistical mediation analysis we performed also found that females tended to have higher levels of trait anxiety, and that this, in turn, influenced early stage neural processing.

Overall, our research project has found that people appear to be good at evaluating confidence in another person’s statements, and this evaluation involves the interaction of both vocal cues—such as speed and intonation of speech—and linguistic cues—such as preceding a statement with the probability phrase “I’m confident” or “perhaps.” These cues also appear to be processed rapidly at the neural level—in some cases being differentiated as rapidly as 1/10th of a second into a statement. However, the way that mixed signals of confidence generate brain activity point to different underlying integrative processes: some related to conflict resolution and others related to detecting concealed intentions. These reports have also raised intriguing new questions currently under investigation regarding how personality traits can be inferred from how “feelings of knowing” affect the voice. We hope this research project will continue to shed light on how we make inferences in mental states of speakers—a common process used to evaluate teachers, business negotiators, politicians, and many other influential figures.


Xiaoming Jiang is a postdoctoral researcher in the Neuropragmatics and Emotion Lab led by Dr. Marc Pell at McGill University. Jiang investigates how the brain makes social inference through vocal cues, more specifically on how inferential processes takes place in a cross-cultural communication setting.

Reference

Jiang, X. & Pell, D. M. (2016b). Feeling of another knowing: how “mixed messages” in speech are reconciled. Journal of Experimental Psychology: Human Perception and Performance. In Press.

Jiang, X. & Pell, D. M. (2016a). Neural responses towards a speaker’s feeling of (un)knowing. Neuropsychologia, 81, 79-93.

Jiang, X. & Pell, D. M. (2015). On how the brain decodes speaker’s confidence. Cortex, 66, 9-34. 

Jiang, X., & Pell, M. D. (2014). Encoding and decoding confidence information in speech. Proceedings of the 7th International Conference in Speech Prosody (Social and Linguistic Speech Prosody), 576–579.

 

Learning a New Language as a Refugee: A Bitter Pill? Or Easy as Pie?

By mid-December of 2019, the number of international migrants across the globe had reached at least 272 million. For all of these migrants, like the millions who migrated before them, learning the local language is usually an essential first step toward self-sufficiency and economic success. Becoming fluent in the most commonly spoken language of one’s new home increases migrants’ chances of finding a job and helps them become integrated into their new community. This may be especially true for refugees who fled their nations under harsh conditions.

As it turns out, migrants differ greatly in how effectively they learn the dominant language. This means that insights into how and why migrants pick up new languages are crucial for helping migrants acquire a new language more easily. Our recent research in the Netherlands addressed this issue among a large sample of adult Syrian and Eritrean refugees.

In keeping with past research, we found that intelligence was the strongest predictor of Dutch language skills. This effect was not surprising as many earlier studies have shown that intelligence is indeed an important trait for learning knowledge and skills in work and educational settings.

But we also examined social, emotional, and motivational predictors of people’s success at picking up a new language. For example, we assessed mental health by asking participants how often they had experienced psychological distress in the previous month. We found that, the lower the psychological distress, the better the local language skills. Good mental health might offer a source of confidence and motivation to learn a new language. In addition, refugees who reported being more motivated to work had better Dutch language skills. Given that local language skills help people get and keep better jobs, having a strong desire to work may boost motivation to learn a local language.

Finally, age and educational attainment were also related to Dutch language learning. Younger adult refugees learned the local language more easily than older adults. The well-known advantage of children in second language learning thus appears to continue into adulthood. However, the degree to which this advantage is due to differences in raw ability or motivation in not clear. Older refugees might be less capable of learning a new language, or instead they might (sub)consciously question the relative cost-effectiveness of learning the local language. Consider the extreme case of a person who immigrates to a new nation at age 80. Is it really worth his or her time and effort to become fluent in a new language? Finally, a higher level of education in the person’s country of origin was an advantage. More years of formal schooling in Syria or Eritrea prior to immigration usually meant more success learning Dutch.

Although laws and policies in one’s new nation might facilitate or hinder how quickly refugees learn a new language, psychological traits may also play a role. Based on the current findings, we encourage communities to meet immigrants where they are and try to address the variables we’ve identified here. First, more intensive and user-friendly language courses should be offered to refugees who are older, who are less educated, and who score lower on intelligence tests. Second, providing professional mental health support is essential for refugees with psychological distress. Third, refugees need support and guidance in searching for jobs. These interventions could mean the difference between a bright future and years of struggle in a new home.


For Further Reading

Asfar, D., Born, M. P., Oostrom, J. K., & van Vugt, M. (2019). Psychological individual differences as predictors of refugees’ local language proficiency. European Journal of Social Psychology, 49(7), 1385-1400.

 

Dan Asfar is Ph.D. student at the Vrije Universiteit Amsterdam who studies the assessment of refugees.

Janneke K. Oostrom is an associate professor at the Vrije Universiteit Amsterdam who studies employee selection and assessment.