Once More, With Feeling: Measuring Emotion of Acting Performances in Contemporary American Film
In this paper, we conduct a computational exploration of acting performance. Applying speech emotion recognition models and a variationist sociolinguistic analytical framework to a corpus of popular, contemporary American film, we find narrative structure, diachronic shifts, and genre- and dialogue-based constraints located in spoken performances.
Read the paper or explore the website. Make sure your audio is on 🔉!
Keep scrolling to dive in.
A single line can have infinite variation in its delivery; performance has great capacity for meaning-making. In this project, we use computational methods to explore the emotional range of performances in contemporary American film.
We apply the analytical framework of variationist sociolinguistics to explore performance: given a fixed line of dialogue, the performance entails a choice between a set of possible deliveries.
We create a pipeline of speech and text-based computational models to construct a dataset of dialogues and the emotional value of their deliveries.
In this paper, we ask three main questions:
- How does emotional performance relate to narrative structure?
- How has the emotionality of films changed historically?
- How can we measure emotional range?
About the data
We apply our pipeline to a corpus of contemporary American film, taking popular live-action narrative films spanning the years 1980-2022. To supplement, we also include the best picture-equivalent nominees from those years from six different organizations.
Ultimately, this results in 2,283 feature-length films spanning the last four decades.
Emotions over narrative time
We first ask the natural question of how our detected emotions relate to narrative time --- in other words, which emotions are most prevalent at a given point in a film?
We first find that the proportion of utterances that are associated with any emotion steadily increases over the course of the film.
But when we look at specific emotions, we find non-linear trajectories.
Like joy, which tends to be highest at the beginning and end.
And sadness, which has roughly the opposite shape.
Historical shifts in emotionality
Though it’s interesting to look at the trajectories of different emotions, the constructs of measuring specific emotions can be restrictive and contested. In the rest of the paper, we focus on the more general notions, like emotionality.
Computational work on English fiction has shown a decline in emotionality. On the other hand, film theorist David Bordwell has famously written about the visual intensification of cinema.
We find that emotionality has declined since the 1980s. And what’s more, this trend exists even when we control for the dialogue being spoken. This means that the decline in the emotion isn’t due to a change in the writing, but rather a change in the performance.
How do we interpret this finding in the context of Bordwell’s visual intensification?
“…whereas, on the stage, the spoken word makes a stronger rather than a weaker impression if we are not permitted to count the hairs in Romeo's mustache.”
Art historian Erwin Panofsky wrote in his essay “Style and Medium in the Motion Pictures” that the unique characteristics of screen acting are exemplified by the close-up shot. The close-up provides a rich field of action for nuanced acting performances.
What would be almost imperceptible from a natural viewing distance is magnified by the camera. In this context, the subtleties of visual performance create a rich channel of expression.
Indeed, Bordwell has found that close-ups have evolved to become ever tighter on their subjects. Perhaps, when the visual channel increases capacity for emotional performance, the emotionality of the spoken word need not bear so strong a burden.
Measuring emotional range
Finally, we examine the idea of emotional range, which is a highly salient concept in film criticism. For some, range is the hallmark of great acting. We construct a quantitative measure of emotional range based on entropy. Our measure is widely applicable for any set of utterances; for example, we might want to measure the emotional range across a genre, a film, or even a particular phrase of dialogue.
Using this measure, we can compare different lines of dialogue to see which have greater capacity for emotional range.
| Text | Entropy |
|---|---|
| Can I ask you something? can I ask you something? Could I ask you something? | −17.02 |
| Hello Captain. I'm captain. Hey, Captain. Good morning, ladies and gentlemen, this is your captain. Hi, Captain. I'm Captain. Hi Captain. Oh, ladies and gentlemen, this is your captain speaking. This is your captain. I'm the captain. | −16.56 |
| Is that okay? Is that alright? | −16.53 |
| Can I get you something? Can I get you anything? Can I get something for you? Can I get you anything at all? Can I get you anything, is there anything you'd like? | −16.17 |
| Can I get you something to drink? Let me buy you guys a drink, right? Can I get a drink or anything? May I get you something to drink? Hi, can I get you to a drink? Can I get you something to eat or drink? Can I get you a drink or something? can I get you a drink? Yeah can I get a drink? Can I get you a drink? | −16.16 |
| What can I get for you? What can I get you? What can I get you too? Hi, what can I get you? Hi folks, what can I get you? Hey, what can I get you? Anyway, what can I get you? Oh, what can I get you? | −15.91 |
| You want to come? Do you want to come? You want me to come? Would you like to come? Do you want me to come? You wanna come? Oh, you want to come? you want to come? | −15.65 |
| Yeah, that's good. Yeah, that's good! Yeah that's good. Oh yeah that's good man. Oh, yeah, that's good. Oh yeah that's good. Oh, yeah that's good. Yeah, that's good, absolutely. Right, that's good. Yeah, yeah, that's good. | −15.36 |
| Any questions? | −15.26 |
| That's correct. That is correct. | −15.25 |
| Okay, thank you. Alright, thank you. | −15.24 |
| You kidding? | −15.19 |
| No, that's all right. No, that's fine. No, that's okay. No that's fine. No, that's alright. | −15.11 |
| Roger that. Roger that Roger that! | −15.05 |
| Can you do that? | −15.00 |
| Stand by. stand by. Stand by | −14.98 |
| Can I ask you a question? Could I ask you a question? | −14.91 |
| Anything else? anything else? | −14.84 |
| Copy that. copy that. | −14.75 |
| Yeah, okay. Yeah okay. Yeah, okay | −14.68 |
| Text | Entropy |
|---|---|
| All rise. | −7.85 |
| Are you out of your mind? Are you out of your minds? Are you're out of your mind? | −7.88 |
| What the hell is wrong with you? What the hell's wrong with you? What the hell wrong with you? What the heck's wrong with you? What the fuck is wrong with you? What in the hell is wrong with you? What the fucks wrong with you? What the fuck wrong with you? What the fuck's wrong with you? | −7.99 |
| You're alive. you're alive. you are alive. You are alive. | −8.32 |
| You saved my life. You saved me a life. You've saved my life today. You saved me life. You've saved my life! You just saved my life. You saved my life! You saved my life today. You've saved my life. | −8.34 |
| Can't you understand? Don't you understand? | −8.34 |
| Don't be afraid. Oh, don't be scared. Don't be scared. not be afraid. Don't be frightened. Well, don't be scared. Don't get scared. Do not be afraid. Don't be afraid! Don't be afraid | −8.39 |
| You son of a bitch! You son of a bitch. You little son of a bitch. You son of my bitch! You son of a bitch! of a bitch! You son of a little bitch. You son of bitch. You fucking son of a bitch. | −8.40 |
| You scared me. You scared me to death. You scared the shit out of me, man. You scared the life of me. You totally scared me. You frightened me. You scared the shit out of me! You scared the hell out of me. You scared the shit out of me. You scared me! | −8.44 |
| Ow. | −8.44 |
| Where is that money? Where's money? Where's the money? Where's that money? So where's the money? Where is the money? Where's all the money? Where's the money now? Where the money at? where is the money? | −8.44 |
| He wasn't my dad. He's not really my dad. You are not my father. You're not my father. I am not your father. You're not my dad. He's not my father. I'm not my father. You are not your father. He's not my dad. | −8.46 |
| I'm warning you. You're warning me. You're warning me! I'm warning. I'm warning you! I'm just trying to warn you. | −8.47 |
| I was scared. I was so scared. I was terrified. I was frightened. I was afraid. was terrified. I was so frightened. I got scared. I was really scared. I got really scared. | −8.48 |
| You are terminated! You're fired. We're terminated. You are fired. You are terminated. I'm terminated. | −8.49 |
| Who the hell are you? who the hell are you? Who the hell is you? | −8.50 |
| What's going on over here? What's happening over here? What's going on out here? What's happening in here? What is going on down here? What's going on around here? What the hell is going on out there? What the fuck is going on? What is happening out there? What's going on in here? | −8.51 |
| What the hell's wrong with me? What's wrong with me? What's the matter with me? What's the wrong with me? What is wrong with me? What the hell's the matter with me? What is going on me? What is going on with me? What's wrong for me? What the hell is wrong with me? | −8.52 |
| alone. Alone. alone | −8.55 |
| What the hell are you doing here? What the hell are you doing? What the fuck you're doing? What the hell you're doing here? What the hell you're doing? What heck are you doing here? What the fuck are you doing? What the fuck are you doing here? What the hell you doing? What the hell do you think you're doing in here? | −8.56 |
Film scholar James Naremore draws on Goffman when writing about film performance: he theorizes that actors draw on and play against the interactional norms with which the audience is already familiar. When we identify the phrases with greatest and least emotional range, we find support for this idea.
Phrases with low range are mostly functional and belong in highly-directed interactions. Many of the phrases are yes-or-no questions, or answers to them.
Phrases with high emotional range, on the other hand, are often more open-ended or evaluative; in these cases, the delivery of the line has much greater space to lend color to the statement being made.
Conclusion
In this paper, we take a computational approach to studying performance in film. While computational work has traditionally focused on the written text of screenplays, and film theorists have traditionally emphasized editing and cinematography, we find there is rich meaning in multimodal analyses of performance in film.
I encourage you to read the paper, which will be presented at Computational Humanities Research 2024 in Aarhus, Denmark. In it, we detail our computational pipeline, further explain the measure of emotional range, and present additional findings on emotional range by genre.
Some additional notes:
- I used movies from the public domain for the examples on this website, but the corpus in our paper covers the box office top-50 live action narrative films from 1980-2022.
- The prestige films we used were the best picture-equivalent nominees from the Academy Awards, Golden Globes, British Academy of Film and Television Arts, Los Angeles Film Critics Association, National Board of Review, and National Society of Film Critics.