Sudeep Bhatia
University of Pennsylvania

What do people know about real-world risk sources, and how can this knowledge
be uncovered, quantified, and used to study the psychological processes involved
in judging risk? We addressed this question with vector space models of semantic
cognition (Turney & Pantel, 2010, Journal of Artificial Intelligence Research).
These models specify human knowledge representations using high-dimensional
vectors trained on large-scale natural language data. Appropriately trained
representations not only describe behavior in semantic memory tasks, but have
also recently been shown to predict a wide range of high-level judgment
phenomena, including probability judgment, factual judgment, forecasting, and
social judgment (Bhatia, 2017, Psychological Review; Bhatia, 2017, Cognition).

These successes suggest that vector space models can also be used to specify
knowledge representations for real-world risk sources, and subsequently predict
the risk judgments of participants.  We tested this in three studies eliciting
ratings of riskiness from over 1,000 participants, for over 400 different
naturalistic risk sources. Participants were also asked to rate the risk sources
on nine key psychometric dimensions of risk perception (Slovic, 1987, Science).

We uncovered knowledge representations for the risk sources by training the
Word2Vec vector space model (Mikolov et al., 2013, NIPS) on a very large
database of news articles. We found that a support-vector-regression mapping
300-dimensional Word2Vec vectors onto participant ratings was able to predict
out-of-sample risk ratings with R2 between 0.55 and 0.75 across the three
studies. We compared these accuracy levels to those obtained using participant
ratings on the nine psychometric risk dimensions, and found that the two
approaches were roughly equivalent, despite the fact that the Word2Vec vectors
required no additional participant data. Importantly, greatest accuracy rates
were obtained by combing the Word2Vec and psychometric approaches, which
resulted in R2 values between 0.8 and 0.9. This implies near-perfect prediction
of out-of-sample risk ratings.

We modified the above technique to uncover the conceptual associates of risk, as
revealed by word use in natural language.  We found that words such as
"fatal", "dangerous", and "tragic" were strongly associated with
risk for all types of risk sources. Further computational analysis of over
10,000 risk associates showed that words strongly associated with risk were also
words that were high in fear and low in happiness. Emotions such as disgust,
sadness,  and anger did not have a strong relationship with risk.

We also implemented a range of other similar tests, probing the affective and
cognitive structure of risk judgment. Additionally, with the above technique
applied to historical natural language data, we retrospectively measured changes
in risk perception from the 1800s to the present. Due to space constraints, we
do not report these tests here.

Overall, our results show how insights from computational linguistics can be
used to uncover knowledge representations underlying risk perception. These
representations can be used to predict participant risk ratings with very high
accuracy. They can also be used to examine the affective and cognitive
substrates of risk perception, and retrospectively measure changes in risk
perception over time.