Unlocking the Plot Themes Driving Ratings
Over the past few months, we have been busy constructing a dataset of more than 183,000 films with rich detail on their production, marketing, and critical reception. In this article, the first in what will be a monthly series exploring this data, we apply a number of statistical and natural language processing (NLP) techniques as we begin our exploration of the defining trends in film.
One of the first questions we wanted to answer involved plot and popularity. Specifically, we wanted see if we could use the paragraph-length text description of a film’s plot to predict its future audience rating*. To achieve this, we applied NLP methods that standardized each film’s plot description (for example: ensuring that similar words were grouped together, that spelling was consistent, and that words that appeared either too frequently or infrequently were removed), identified the key themes across all films, and transformed these themes into the cogs of a quantitative model.
Before exploring the results, we’d like to address two issues that could arise from using audience ratings from the web as a dependent variable in our model. First, as a content producer, you might be more interested in revenue than ratings and perhaps skeptical that online ratings are truly representative of popular opinion. Second, online ratings are heavily biased toward the opinions of younger viewers (an effect that is likely compounded for older films, where online ratings may have generational and recency bias).
That said, we think these issues can be overcome. On the first point, we found a significant positive correlation between revenue and ratings in our dataset. In other words, on average a higher rating equates to higher revenue, despite the occasional critically-shunned box office hit. On the second point, we argue that contemporary opinion is what matters most. We are selling content today, not 20 or 30 years ago.
Acknowledging potential limitations, it’s time to look at the results. The figure below shows the most impactful themes across our dataset:
Because we controlled for genre in our regression model, we can interpret these themes without any potential confounding influence from a genre’s popularity. Put differently, mentioning an ‘exorcism’ in a plot description is associated with a decrease in a film’s rating above and beyond any underlying audience distaste for horror films. Consequently, we can say that the inclusion of ‘shark’ or ‘exorcism’ in a plot description has historically driven ratings down while the inclusion of ‘Working class’, ‘African American’, or ‘Odyssey’ has driven ratings up. Whilst these themes might be fairly straightforward to interpret, others require some unpacking.
A Social ‘Climb’ in the Himalayas
For example, adding ‘climb’ to a plot was found to have the largest positive impact on a film’s audience rating. While films about climbing are demonstrably popular (see Free Solo, Meru, or 127 Hours), this archetype accounts for only half of the occurrences of ‘climb’. Most of the other occurrences relate to social climbers! The description for the classic 1990 comedy Problem Child includes, “Ben Healy and his social climbing wife,” and the less classic 2003 teen dramedy What a Girl Wants includes, “her long-lost dad is engaged to a fiercely territorial social climber.” The social climbing antagonist appearing to be a trope we love to hate.
‘Thanks’ for nothing
Another theme associated with higher ratings that invites further scrutiny is ‘thank’. Interestingly, the majority of the 44 occurrences of thank are part of the phrase ‘thanks to…’ and convey unforeseen circumstances thrust upon a movie’s protagonist. This can be seen in the plot description of Back to the Future II which reads, “Thanks to bully Biff…”. As most people prefer to believe themselves victims of circumstance rather than their own undoing, it is unsurprising that overcoming unexpected and unavoidable bad luck is a popular theme.
A ‘Menace’ to your ‘Psychopathic’ self
What about ‘psychopathic’ and ‘menace’? These words appear very similar but historically have had opposite effects on audience ratings. A closer examination reveals an important distinction in how they’ve been used. Menace tends to be used to describe group threats or nonhuman entities like in the 2010 movie Birdemic, “It’s not known what caused the flying menace to attack,” while psychopathic tends to be used to describe singular human villains, like in the 2000 movie American Pyscho , “banking executive hides his alternate psychopathic ego from his co-workers.” It would seem that most viewers are more captivated by a dark- but relatable- fantasy than one where they are merely the victims of a half-alligator menace.
‘Sexy’ doesn’t sell?
One of the surprising findings was that both ‘sexy’ and ‘nerdy’ signaled a detriment to ratings. However, we wouldn’t be too quick to toss aside old assumptions. The appearance of sexy in a plot description, rather than solely being an indicator of a lead’s attractiveness, serves as a harbinger for typical lowbrow content. Like in the 1996 film Barb Wire, “A sexy nightclub owner, Barb Wire moonlights as a mercenary…,” or in the 2007 film Virgin Territory, “There are randy nuns, Saracen pirates, and a sexy cow.” Similarly, where nerdy appears in a plot description it appears to be more about attracting unsophisticated moviegoers than denoting an audience distaste for geeky leads, like in 1999’s She’s All That, “Gain the trust of nerdy outcast Laney Boggs.” No knock on Laney Boggs, but a film experience seldom accused of being an intellectual romp.
In the next part of this series we are going to dive deeper into plot themes by examining their influences over time. Plotting important trends from the 1950s to today, we will see how underlying cultural anxieties and aspirations drive each generation’s endorsement or aversion to specific thematic content.
*Technically, after cleaning and lemmatizing plot descriptions, we created TFIDF (Term Frequency–Inverse Document Frequency) features for themes that occurred in a minimum of 10 different descriptions and no more than 80% of descriptions. We also included dummy variables for genre, looked only at English language films, and excluded films that didn’t receive at least 20 unique ratings.