The following is an excerpt from The Language of Food: A Linguist Reads the Menu, in which author Dan Jurafsky analyzes user-generated restaurant reviews to draw conclusions about the language -- positive or negative -- that we use to talk about food:
You can always get a good argument going in San Francisco by asking people for their favorite taqueria. I lean toward the carnitas at La Taqueria on Mission, but our friend Calvin can be pretty eloquent on the subject of the al pastor at Taqueria Vallarta on Twenty-Fourth. San Franciscans are similarly contentious about the best dim sum, and have been politely disagreeing about tamales since the 1880s, when the city was famous for the vendors plying the streets every evening with pails of hot chicken tamales. (Some things, of course, are simply not a matter of opinion, like the best place for roast duck -- it’s Cheung Hing out in the Sunset, but don’t tell anybody else, the line is already too long.)
It’s not just San Francisco. You can’t go on the Internet these days without stumbling over someone’s lengthy review of a restaurant, wine, beer, book, movie, or brand of dental floss. We are a nation of opinion-holders. Perhaps we always have been: in De Tocqueville's prophetic study of the American character, the 1835 Democracy in America, he noted that in the United States “public opinion is divided into a thousand minute shades of difference upon questions of very little moment."
Consider online restaurant reviews, those summaries of the wisdom of the crowd that have become a familiar way to discover new places to eat. Take a look at this sample from a positive restaurant review (a rating of 5 out of 5) on Yelp (modified slightly for anonymity):
I LOVE this place!!!!! Fresh, straightforward, very high quality, very traditional little neighborhood sushi place... takes such great care in making each dish... You can tell the chef really takes pride in his work... everything I’ve tried so far is DELICIOUS!!!!
And here are bits of one negative review (a rating of 1 out of 5):
The bartender was either new or just absolutely horrible... we waited 10 min before we even got her attention to order... and then we had to wait 45 -- FORTY FIVE! -- minutes for our entrees... Dessert was another 45 min. wait, followed by us having to stalk the waitress to get the check... he didn’t make eye contact or even break his stride to wait for a response... the chocolate soufflé was disappointing... I will not return.
As eaters we use reviews to help decide where to eat (maybe give that second restaurant a miss), whether to buy a new book or see a movie. But as linguists we use these reviews for something altogether different: to help understand human nature. Reviews show humans at their most opinionated and honest, and the metaphors, emotions, and sentiment displayed in reviews are an important cue to human psychology.
In a series of studies, my colleagues and I have employed the techniques of computational linguistics to examine these reviews. With Victor Chahuneau, Noah Smith, and Bryan Routledge from Carnegie Mellon University, my colleagues on the menu study of Chapter 1, I’ve investigated a million online restaurant reviews on Yelp, from seven cities (San Francisco, New York, Chicago, Boston, LA, Philadelphia, Washington), covering people’s impressions between about 2005 and 2011, the same cities and restaurants from our study of menus. With computer scientists Julian McAuley and Jure Leskovec, I looked at 5 million reviews written by thousands of reviewers on websites like BeerAdvocate for beers they drank from 2003 to 2011.
As we'll see, the way people talk about skunky beer, disappointing service, or amazing meals is a covert clue to universals of human language (like the human propensity for optimism and positive emotions and the difficulty of finding words to characterize smells), the metaphors we use in daily life (why drugs are a metaphor for some foods but sex is a metaphor for others), and the aspects of daily life that people find especially traumatizing.
Let's start with a simple question. What words are most associated with good reviews, or with bad reviews? To find out, we count how much more often a word occurs in good reviews than bad reviews (or conversely, more often in bad reviews than good reviews).
Not surprisingly, good reviews (whether for restaurants or beer) are most associated with what are called positive emotional words or positive sentiment words. Here are some:
Bad reviews use negative emotional words or negative sentiment words:
Words like horrible or terrible used to mean “inducing horror” or “inducing terror,” and awesome or wonderful meant “inducing awe” or “full of wonder.” But humans naturally exaggerate, and so over time people used these words in cases where there wasn’t actual terror or true wonder.
The result is what we call semantic bleaching: the “awe” has been bleached out of the meaning of awesome. Semantic bleaching is pervasive with these emotional or affective words, even applying to verbs like “love.” Linguist and lexicographer Erin McKean notes that it was only recently, in the late 1800s, that young women began to generalize the word love from its romantic core sense to talk about their relation- ship to inanimate objects like food. As late as 1915 an older woman in L. M. Montgomery’s Anne of the Island complains about how exaggerated it was that young women applied the word to food:
The girls nowadays indulge in such exaggerated statements that one never can tell what they do mean. It wasn’t so in my young days. Then a girl did not say she loved turnips, in just the same tone as she might have said she loved her mother or her Saviour.
Semantic bleaching is also responsible for meaning changes in words like sauce or salsa from their original meaning of “salted,” but I am getting ahead of myself. For now there’s much more to learn from reviews.
Let’s start with the negative reviews. Consider the very specific and creative words used to express dislike (sodalike, metallic, wet dog water, force-carbonated, razor thin) in this strongly phrased negative beer review from BeerAdvocate:
Clear light amber with a sodalike head of white that immediately fizzles to nothing. Very sodalike appearance. Aroma is sweet candy apricot with slight metallic wheat notes. Flavor is wet dog water infused with artificial apricot. Bad, bad, bad. Mouthfeel is razor thin, watery, and highly force-carbonated. Drinkability? Ask my kitchen sink!
My colleagues and I automatically extracted the positive and negative words. While reviewers generally called beers they disliked “watery” or “bland,” they tended to describe the way they were “bad” by using different negative words for different senses, distinguishing whether the beer smelled or tasted bad (corny, skunky, metallic, stale, chemical), looked bad (piss, yellow, disgusting, colorless, skanky), or felt bad in the mouth (thin, flat, fizzy, overcarbonated).
By contrast, when people liked a beer, they used the same few vague positive words we saw at the beginning of the chapter -- amazing, perfect, wonderful, fantastic, awesome, incredible, great -- regardless of whether they were rating taste, smell, feel, or look.
The existence of more types of words, with more differentiated meanings, for describing negative opinions than positive ones occurs across many languages and for many kinds of words, and is called negative differentiation. Humans seem to feel that negative feelings or situations are very different from each other, requiring distinct words. Happy feelings or good situations, by contrast, seem more similar to each other, and a smaller set of words will do.
Negative differentiation comes up in all sorts of domains. For example, across languages there seem to be more adjectives to describe pain than pleasure. We use more varied vocabulary to describe people we dislike than people we like. People even describe attractive faces as more similar to each other while unattractive faces differ more from each other. This generalization that there are more different ways to be negative than to be positive was most famously stated by Tolstoy at the beginning of Anna Karenina: “Happy families are all alike; every unhappy family is unhappy in its own way.”
Words for smell seem particularly disposed to the negative trend. English, for example, has no commonly used positive word meaning “smells good” that corresponds to delicious for taste or beautiful for sight. Languages generally seem to have a smaller vocabulary for smell than for other senses, relying on words for tastes (like sweet or salty) or names of objects (like gamy, musky, skunky, or metallic).
Some languages do have somewhat richer olfactory vocabularies, like Janet’s native language, Cantonese. Unlike English, Cantonese has a common word that means “smells good,” heung 香, often translated as “fragrant.” Fragrant in English is rare and poetic, but the everyday Cantonese heung (and its Mandarin cognate xiang) is just how you say you like the smell of what’s cooking. It’s such a frequent word that you’ve all seen it: heung is the first part of the name Heung Gong (Hong Kong; “smells-good harbor”).
Cantonese is particularly rich in words for negative smells. Here are some:
suk1 the bacterial smell of spoiled rice or tofu
ngaat3 the ammoniacal smell of urine, ammonia
yik1 the smell of rancid or oxidized oil or peanuts
hong2 the stale, rancid smell of old grain (uncooked rice, flour,
seng1 fishy, bloody smell
sou1 musky, muttony, gamy, body odor smell lou3 the smell of overheated tires or burnt hair
Note the numbers following each word. Cantonese has six tones, characteristic rising or falling pitches, and the meaning of a word varies depending on the tone used. The richness of this language is not limited to ways to say stinky.
Many of the words listed above exist in other Chinese dialects as well, and some are very ancient. An essay on cuisine in a third-century bce Chinese encyclopedia (which Chinese cuisine scholar Fuchsia Dunlop calls "perhaps the world’s oldest extant gastronomic treatise”) records the ancient advice of the sixteenth-century bce cook Yi Yin on how to eliminate fishy (seng1 腥) and gamy (sou1 臊) smells.
Sadly, use of this ancient and rich negative smell vocabulary seems to be dying out of Cantonese. Studies show that younger Hong Kong speakers know fewer of these words than their elders, as sanitization and plastic wrap eliminate opportunities to experience what linguist Hilario de Sousa delicately calls “the variety of olfactory sensations experienced by their ancestors.”
The minimal smell vocabularies of many languages may be recent and due to urbanization (languages retaining the vocabulary are often spoken outside the cities), ancient and genetic (many genes coding for the detection of specific odors are turned off in humans, perhaps dating back to the development of tricolor vision in primates), or related to human variation in smell perception. For example, genetic variations lead to differences in detecting the grassy smell of sauvignon blanc, partially caused by the flavor compound cis-3-hexen-1-ol. The ability to detect the sulfurous smell of asparagus in urine has similar genetic links; according to one recent experiment, about 8 percent of people don’t produce it, and about 6 percent can’t smell it. (My biologist wife, upon reading that paper, immediately conducted an impromptu experiment on yours truly by cooking up a big batch of asparagus.) The vast variation over the many different abilities of smell might have made it harder for a language to develop a stable shared olfactory vocabulary.
The greater differentiation of negative smells is but one aspect of negativity bias, the idea that humans are biased to be especially aware of negative situations. Bad reviews like the one at the start of the chapter display another. To understand, we need to look beyond the negative emotional words like horrible, terrible, awful, and nasty and focus instead on the story being told. Yes, story.
Linguist Douglas Biber has shown that we use past tense, communication verbs (said, told), and event words (then, after) much more frequently when telling stories, and the negative reviews are filled with these features. Let’s also look at the common nouns most strongly associated with them:
Not a one of these words refers to food! Instead, bad reviews are stories about bad things done by other people. The waiter or waitress made some mistake, messed up the order or the bill, or had a bad attitude, the manager didn’t help, the hostess caused a long wait, and so on.
In addition, bad reviews overwhelmingly use the pronouns we or us (“We waited,” “our entrées,” “us having to”). While other reviews use those pronouns too, “we” and “us” are vastly overrepresented in negative ones. What is the common denominator of these three features: negative emotional words like terrible and horrible, narrative stories about other people, and a vast increase in we and us, all strongly linked to 1-star reviews?
The answer comes from the pioneering work of Texas psychology professor James Pennebaker, who for decades has studied how words like function words are veiled cues to people’s personalities, attitudes, and feelings. Pennebaker has particularly studied the aftereffects of trauma. His “social stage model of coping” suggests that immediately after a traumatic event, people feel a need to tell stories about the event, stories expressing their negative emotion, and suggests that traumatized people seek comfort in groups by emphasizing their belonging, using the words we or us with high frequency.
Pennebaker and his colleagues identified these tendencies in bloggers talking about their feelings after September 11, 2001, in fans writing about the death of Princess Diana, and in student newspaper articles after campus tragedies. In each case, what people write is just like terrible reviews of restaurants: narratives, stories about the negative things that happened to them, bulwarked against these negative emotions by the solidarity of us and we. In other words, bad reviews display all the linguistic symptoms of minor trauma.
We always confirm our automated methods by carefully reading selected samples of the reviews. And the tendency toward negative bias is clear, from the negative differentiation in describing skunky beers to the trauma narratives of bad restaurants.
Why do we find negative things more intense and more differentiated than positive things? One possibility is that negative things in the world really are more different from each other than positive things. Perhaps there really is more difference between being evil, brutal, sad, sick, or skunky than there is between being good, gentle, happy, well, or nice. Another possibility is that negative things aren’t actually more different or more potent than positive things, but it's evolutionarily useful for us to treat them as if they were so. Humans need to worry about and be exceptionally good at distinguishing among negative events. The intuition of this theory is that there are a lot of ways for things to go wrong in life, and even though they may be very rare (like tiger attacks and earthquakes and bee stings), they require very different responses. Having different words to talk about how to avoid them helped our ancestors outlive the tiger and the earthquake.
Of course, reviews aren’t all negative. What are the metaphors and other linguistic structures that reviewers use in positive reviews of food or wine?
Let’s start by talking about sex.
Adrienne Lehrer, a linguistics professor at the University of Arizona, studied how wine reviews changed over time from 1975 to 2000. She noticed that in the 1980s wine reviewers began to increase their use of the body as a metaphor, starting to use words like fleshy, muscular, sinewy, big-boned, or broad-shouldered. At the same time, influential wine writers like Robert Parker began to emphasize the sensual pleasure of wine, repeating words like “sexy” and “sensual,” describing wines as “supple and seductive,” “offering voluptuously textured, hedonistic drinking,” or even “liquid Viagra.” Literature professor Sean Shesgreen says that all this erotic talk about wine as “pretty and caressing,” “rav- ishing,” “pillowy,” and “overendowed” affirms that “in the kaleidoscope of Americans’ fixations, gastronomy has eclipsed sex.”
This metaphor of sex seems especially associated with expensive foods as well. We examined this in the million restaurant reviews by extracting every mention of sex (or related words like sexy, seductive, orgasms, or lust) in the reviews. We then used regression, a statistical technique that allowed us to ask how these mentions of sex were associated with people’s ratings of a restaurant, after controlling for factors like the type of cuisine and the city.
Reviewers who liked a restaurant were indeed more likely to use sexual metaphors. But we also discovered an economic interaction; mentions of sex like these are especially frequent for expensive restaurants:
The apple tarty ice cream pastry caramely thing was just orgasmic sumptuous flavors, jaw-droppingly good, sexy food
succulent pork belly paired with seductively seared foie gras
The association is quite strong: the more mentions of sex in a restaurant review, the higher the price of the restaurant. People use a very different metaphor when they like the food at cheap restaurants. In reviewing inexpensive restaurants, they use the language of addiction or drugs instead of sex to talk about their fries or garlic noodles:
- garlic noodles... are now my drug of choice
- these cupcakes are like crack
- be warned the wings are addicting
- ...every time I need a fix. That fried chicken is so damn good!
- I swear the fries have crack or some sort of addicting drugs in them
The examples above show what we “crave” or are “addicted to”: chicken wings and fried chicken, cupcakes, garlic noodles, French fries, and burgers. It’s the snack foods and bar foods, guilty pleasures because of their fat, sugar, and deep-fried goodness that invite the comparison to drugs. Researchers still aren’t sure of the biochemical link between junk-food cravings and drug addiction, but in any case the cravings for fat and also sugar are quite strong. A study that varied the fat and sugar in chocolate milkshakes suggests that sugar may light up the reward center of the brain even more effectively than fat. Writer Adam Gopnik describes nights during his experiment in giving up dessert when he would wake up and -- like a golem controlled by external command -- sleepily wander toward the freezer and the ice cream.
In any case, the linguistic ubiquity of this metaphor of drugs demonstrates how deep this addictive understanding of junk food and desserts is embedded in our culture. By placing the blame on the food, we’re distancing ourselves from our own “sin” of eating fried or sugary snacks: “It’s not my fault: the cupcake made me do it.” Our research also found that women are more likely than men to use drug metaphors in reviews, suggesting that they are especially pressured to conform to healthy or low-calorie eating.
What are people eating when they talk about sex in reviews? We can study this by looking at food words that occur more frequently near sexual words. Two kinds of foods are associated with sex. One is sushi, because of the modern trend of giving sexy names to sushi like these:
sex on the beach roll foreplay roll
sweet temptation roll orgasmic spicy tuna roll
sexy mama roll sexy lady roll
hot sexy shrimp roll sexy lizzy roll
The other food most frequently associated with sex is dessert:
molten chocolate cake... honestly an orgasm on a plate
I still lust for the silky panna cotta and tantalizing sorbet marshmallows... so... sticky and sweet, they’re nearly pornographic warm chestnut mochi chocolate cake... seductively gooey on the inside
The examples above also exhibit another class of words associated with both dessert and sex: texture words like sticky, silky, gooey. Here are the sensory words most commonly used to describe desserts in the million reviews:
All of these are from the sensory domain of “feel,” of textures and tem- peratures. When we talk about desserts, we talk about their feel in the mouth, not their appearance, smell, taste, or sound. Americans usually describe desserts as soft or dripping wet, a tendency that linguist Susan Strauss, in her comparison of TV advertising in the US, Japan, and Korea, found to be a general property of food advertising in American English. US commercials emphasize tender, gooey, rich, creamy food, and associ- ate softness and dripping sweetness with sensual hedonism and pleasure.
This association between soft, sticky things and pleasure isn’t a necessary connection. For example, Strauss found that Korean food commercials emphasize hard, texturally stimulating food, using words like wulthung pwulthung hata (solid and bumpy), ccalis hata (stinging, stimulating), thok ssota (stinging), and elelhata (spicy to the extent that one’s nerves are numbed).
The link between dessert and sex is visible in many aspects of our culture, from the sensual advertising of chocolate to women (like Ghirardelli’s slogan, “Moments of Timeless Pleasure”) to modern music, where my students Debra Pacio and Linda Yu found that recent songs like Kelis’s “Milkshake” or Li’l Wayne’s “Lollipop” use dessert and especially candy as a metaphor for sex. There is a gender effect with dessert too. Our study shows that women are more likely than men to mention desserts in their reviews.
Dessert is also so prized that people find it very difficult to say any- thing bad about it. Notice the overwhelmingly positive sentiment of the 20 most frequent sentiment words associated with dessert:
In fact, the more Yelp reviewers mention dessert, the more they like the restaurant. Reviewers who don't mention a dessert give the restaurants an average review score of 3.6 (out of 5). But reviewers who mention a dessert in their review give a higher average review score, 3.9 out of 5. And when people do talk about dessert, the more times they mention dessert in the review, the higher the rating they give to the restaurant.
This positivity exhibited by reviews, filled with metaphors of sex and dessert, turns out to be astonishingly strong. Despite the negativity bias that makes us especially sensitive to negative situations, people are actually much more positive than they are negative.
One sign of our positive nature is word frequency. Positive words, though weak in variety, occur much more often in reviews than negative words. Restaurant reviewers use words like great, delicious, and amazing 3 to 10 times more often than words like bland, bad, or terrible.
Review scores themselves are also skewed toward the positive. Reviewing scores on most sites go from 1 to 5, so the median score should be 3. Instead the median score, whether for restaurants or beers, is about 4 out of 5. My colleague down the hall Chris Potts has shown that this skew is true wherever people review things on the web -- books, movies, cameras, you name it.
This tendency toward the positive is not a recent trend caused by the Internet, but has been shaping our language for millennia. Linguists are deeply interested in linguistic phenomena that hold across all languages, key to our goal of discovering true human universals. A bias toward positivity in vocabulary is one of the strongest universals we have found. This idea that people are positive is called the Pollyanna effect, after the heroine of Eleanor Porter’s 1909 book for children, Pollyanna, an orphan who always looked on the bright side. In common usage “Pollyanna-ish” describes a naïve or foolish optimism, but the Pollyanna effect is a more neutral observation of humans’ remarkable tendency toward optimism.
The Pollyanna effect is not just specific to reviews. If you ask Google how frequent a word is (or check the frequency in a carefully con- structed academic database of texts), positive words are (on average) more frequent than negative words. English good is more frequent than bad, happy than sad; Chinese kaixin 开心 (happy) is more frequent than nanguo 难过 (sad); Spanish feliz is more frequent than triste.
More subtly, positive words have a special linguistic status called unmarked. Markedness has to do with oppositions: in pairs of words like happy/unhappy, good/bad, capable/incapable, or honest/dishonest, the first of each pair is unmarked or neutral and the second is marked. There are many linguistic cues to which member of a pair is unmarked. The unmarked form is shorter (marked unhappy and dishonest have an extra un- and dis- than unmarked happy and honest). Unmarked words tend to come first in “X and Y” phrases like “good and evil” or “right and wrong.”
Unmarked words are neutral in questions. Ask- ing “Is your accountant honest?” is the neutral way to find out about the honesty of your accountant. If I instead ask, “Is your accountant dishonest?” that suggests that I already have some reason to believe you have a cheating accountant. Sure enough, across languages, the unmarked form is much more likely to be positive (happy, honest) rather than negative (unhappy, dishonest); it’s very rare across languages for a negative word like sad to be the basic form and unsad to be the way to say “happy.” Thus we have English words unhappy, incapable, uncomfortable, but not unsad, un-itchy, unklutzy.
The Pollyanna effect has been confirmed in dozens of languages and cultures, and comes up in all sorts of nonlinguistic ways as well. When psychologists ask people to think of items or remember them from a list, they name more positive things than negative things. When people forward news stories, they are more likely to forward the positive stories than the negative ones.
In other words, although humans have a lot of ways of talking about negative events, and are especially traumatized when other people are rude or mean to them, although people differ in all sorts of ways, perceive different tastes and smells, and range hugely in their personalities, these differences only serve to highlight a fundamental similarity as humans: we are a positive, optimistic race. We tend to notice and talk about the good things in life. Like dessert. And sex.
And all of this, joy and trauma, is visible in those reviews on the web, offering a little insight into the human psyche along with advice on where to go for dinner. Just don’t forget to order dessert.
Excerpted from The Language of Food: A Linguist Reads the Menu by Dan Jurafsky. Copyright © 2014 by Dan Jurafsky. With permission of the publisher, W. W. Norton & Company, Inc. All rights reserved.