What Makes Some Words More Memorable Than Others?
Machine learning brings new firepower to help answer a persistent question.
There appears to be a consistent, if complex, psychology behind the words we remember. | iStock/SvetaZi/Cory Hall
Why some words are more memorable than others has long been something of a mystery. Many people can instantly recognize a brand they have purchased before or remember who they talked to at the store but also struggle to recall some of the items on their grocery list.
Ada Aka, an assistant professor of marketing at Stanford Graduate School of Business, has been fascinated by this question since her first undergraduate cognitive psychology class. “Advertising slogans, movie quotes, conversations — some things just stick in your mind,” she says.
For more than a century, psychologists have tried to understand the memorability of words. Do certain sounds stick in our minds? Do we latch onto words related to particular themes? Aka found the existing literature provided few answers. So, she set out with new tools to tackle this nagging question.
In a recent paper published in the journal Cognition, Aka and her coauthors, Sudeep Bhatia and John McCoy of the University of Pennsylvania, paired “mega lab studies” with a machine learning model to test a large number of theories about word memorability.
First, they identified several patterns suggesting a consistent, if complex, psychology behind the words we remember. “There were certain words that many, many people across different conditions remembered. And there are other words that people tend to forget,” Aka says.
Overall, Aka and her colleagues found that memorability was linked to semantic categories. Informal and slang words like “hm” and “damn” were more memorable. So were words associated with death, like “bury” and “kill,” as well as words associated with religion, like “altar.” On the other hand, people were more likely to forget words connected to cognitive processes and time. Negative words tended to be more memorable than positive words.
These findings were consistent across the dozens of participants who took part in the study. “Looking across people at different time points and different lists of individual words — the memorability was quite similar.” Aka surmises there may be an evolutionary aspect to this common selective memory: The more closely tied a word is to survival, the more likely we are to remember it.
Unlike previous studies, which could only test one or a few theories at a time, this project used a computational model to analyze hundreds of variables that might affect memorability. That scope means Aka and her coauthors were able to capture a deeper, more complete picture of which kinds of words we’re most likely to remember. In doing so, they’re uncovering some fundamental information about memory as a whole.
Aka laid the groundwork for this study as a graduate student at the University of Pennsylvania. Working alongside her mentors and colleagues, she helped assemble a uniquely rich and extensive dataset of word memorability. The researchers asked volunteers to come into the lab for many sessions, where they would sit in front of computers and study lists of words pulled from a pool of hundreds of words. When they reached the end of each list, they would be tested to see if they could recognize words they’d seen or recall them from memory.
Aka draws an analogy between these two cognitive tasks — word recognition and recall — likening them to encountering someone on the street. Recognizing their face is akin to identifying a word you’ve previously encountered. Conversely, recalling their name is similar to word recall: It requires actively retrieving the information from memory. In this study, the researchers explored these two memory processes concurrently. “While most studies have focused on either task, there was a lack of a comprehensive framework that juxtaposed and delineated the unique aspects of what renders each task memorable,” Aka explains.
The volunteers were required to return to the lab up to 24 times to do tests with randomized word lists from the same set of words. Next, Aka and her colleagues combined their dataset with Word2Vec, a natural language model that captures semantic relationships between words. While there were variations between individuals’ memories, the model they built was able to capture the broad associations between memorability and hundreds of different words and categories. Aka and her coauthors found that the types of words we recall are similar to the ones we recognize. “It turns out the semantic determinants are actually fairly stable,” she says, converging around similar characteristics like informality and themes like death.
The researchers validated their results by asking the model to draw on what it had learned from the data to predict the memorability of a new selection of words. Not only was the model able to predict memorability accurately, it was significantly better than humans at predicting which words people would remember. “I think that relates to our everyday experience,” says Aka. “We never think that we’re going to forget where we put the keys.”
Digging Deeper Into Memory
Aka knows that uncovering the words we remember isn’t enough to fully explain what we remember and why. It’s rare to encounter a single word in the wild. Usually, we’re contending with other semantics like whole sentences, jingles, repetition, and rhymes. Yet studies suggest that studying and remembering individual words is a proxy for all kinds of other scenarios where researchers are trying to understand what we remember and why. Whether you’re remembering a list of words or items in your grocery list, Aka says, “The memory dynamics are quite similar.”
Aka is already working on sentence-level models that can predict the memorability of brand slogans. Eventually, she’d love to be able to test more complex real-world stimuli like advertisements with words and images side-by-side.
What’s really exciting in the short term, she says, is that this model can make predictions about words that aren’t in the 600 or so it was trained on. That means researchers can test new hypotheses, like whether pain-related words are more memorable than words associated with pleasure. Her team’s datasets are open-access, so other researchers can use them to test their theories.
For now, progress will mostly happen in incremental steps, Aka says. But this could be a huge methodological shift for the field. “Especially with the recent advances in data science and natural language processing, this type of data-driven work is much more possible,” she says. “And I’m definitely up for leveraging that power.”
For media inquiries, visit the Newsroom.