Stanford Literary Lab uses digital humanities to study why we feel suspense
Stanford English Professor Mark Algee-Hewitt and a team of eight graduate students combine digital tools of textual analysis with the emotional experience of reading to uncover what creates suspense in stories.
You’re sitting on the edge of your seat. Your heart starts racing. You scream aloud, “Don’t open that door!”
We all know the feeling of suspense. But why do we experience such an intense emotion when reading a book?
This is the question that the Stanford Literary Lab’s project on suspense has set out to answer using the tools of digital humanities. The project is led by Stanford English Assistant Professor Mark Algee-Hewitt, and involves a team of eight graduate students from the English Department.
“The big goal of the research is to try and explain why we feel suspense when confronted with certain aesthetic objects, even if we know the outcome of them,” Algee-Hewitt said.
In fact, the continued experience of suspense for readers even when they know what happens in the plot has been a central question for this type of literary study, he added.
Although the project is still ongoing, the group’s central finding so far is that suspense is characterized by the presence of words that convey how things appear to be rather than how they really are, such as “seemed,” “perceived,” or “observed.”
These words generate an “epistemological uncertainty,” Algee-Hewitt said.
“Suspense texts appear to be able to create a virtual space in which the reader can experience uncertainty without necessarily having this kind of ontological uncertainty about the text, or forgetting the ontological certainty of the text that he or she already knows,” he said.
In other words, even if you already know what is going to happen next, the text’s description of how things “seem” still triggers a feeling of uncertainty and suspense.
Developing a virtual reader
The way the group arrived at this finding is just as interesting as the finding itself. Studying suspense with digital humanities methodologies posed a problem from the beginning because suspense is so dependent on the reader’s emotional experience, Algee-Hewitt said.
To incorporate the reader’s response to the text into their data, the group used various methods of tracking their reading experience, such as rating paragraphs by the level of suspense they felt while reading on a scale of one to ten.
Hannah Walser, an English doctoral candidate, said, “It’s an exciting way not to jettison concepts like taste or suspense or attachment in the name of objectivity, but instead to take them as an object of study that you can track and quantify.”
When they compared their individual ratings of suspense in a set of short stories, it turned out that the group did agree on the points at which suspense increased or decreased, though they varied in their ratings of the degree of suspense felt.
“We discovered that we agree in general on what is suspenseful and what isn’t, at least in terms of the ups and downs of the narrative,” said Andrew Shephard, an English doctoral candidate.
They then turned to digital methods to assess what words and topics were most commonly associated with the moments of high suspense they had identified. They found that suspenseful passages were characterized by words relating to the imagination (e.g., “thought”), the senses (“saw”), and movement (“struggled”) and topics such as “assault,” “guns,” “crime,” and “dramatic weather.”
“We wound up figuring out a set of features that do a very good job of predicting suspense,” Algee-Hewitt said. They decided the next step would be to develop a virtual reader using a neural network that could identify suspense based on these features.
A neural network is a computer program that receives inputs on how to categorize certain objects and can be trained to learn how to identify new objects on its own. In this case, they trained the neural network to recognize suspenseful passages based on the sets of words and topics associated with suspense and their data on what readers had identified as moments of high suspense.
The neural network achieved 81 percent accuracy in identifying passages it had never seen before as either suspenseful or non-suspenseful.
“I was shocked. This was not supposed to work,” said Algee-Hewitt. “The very fact that something that is so subjective and affective can be somewhat accurately predicted based on formal features of a text is one of the more surprising things we’ve come across.”
Although creating a program that can detect suspense was not an initial goal of the project, it did help the scholars understand what truly creates suspense. By analyzing what words and topics the neural network relied on most when identifying suspenseful passages, they arrived at their conclusion about words that create what Algee-Hewitt calls epistemological uncertainty.
“This is one of the places where digital humanities research for me is most exciting, because very frequently, either the results will not be what you expect or you wind up finding something you weren’t looking for to begin with,” he said.
Collaboration in the humanities
Because of the wide-ranging historical span of texts that the group looked at, from the 18th through 20th centuries, the project required a large number of scholars with different areas of specialization.
“We ended up with a pretty good historical coverage between British and American expertise, so it’s actually a huge part of what has allowed the project to be possible,” said Tasha Eccles, an English doctoral candidate.
Not only has the project required a large group of participants, it has yielded a collaborative form of research that is unusual for the humanities.
“I think it’s been eye-opening to me to see that academic work can happen in many different ways,” said Abigail Droge, an English doctoral candidate.
The group hopes to produce a co-authored publication, which will include both the results from the group as a whole and essays on individual projects of interest that relate to suspense.