By Miranda Speyer-Larsen, Junior Researcher
I started this project wanting to combine two interests of mine: Natural Language Processing (NLP) and emotional stereotypes related to gender. The latter has long been a subject of both curiosity and frustration for me, and I wanted to explore it in a setting that felt accessible and engaging. Originally, the project was about analyzing movie scripts more broadly, but I soon narrowed the focus to Friends, due to its ongoing popularity.
It’s been over 30 years since Friends first aired, and somehow, it’s still everywhere. People are still watching reruns, quoting Chandler, and debating whether Ross and Rachel were really on a break.
Since Friends has managed to stay relevant to audiences who weren’t even born when it ended, it must have aged pretty well, right? Or at least there must be something about it so relatable, that it transcends generations.
This leads me to the final project, where I explore some of the things we can learn about gender and emotions by using NLP techniques.
But before we get into the technical stuff, let’s talk about the show itself.
Revisiting Friends Through a 2025 Lens
There’s been a lot of discussion over the years about how Friends has aged. On the one hand, it tackled topics that were rare for its time. Take the episode “The One with the Lesbian Wedding,” which aired in 1996, eight years before same-sex marriage was legal anywhere in the US. Ross, despite his heartbreak, stands up for his ex-wife and her partner against his ex-in-laws’ homophobia. That kind of representation is still meaningful to this day.
But there are also parts of the show that haven’t aged as well. The endless fat jokes about Monica, or the almost complete lack of people of color (Aisha Tyler’s character, Charlie, doesn’t show up until the ninthseason). And then there are episodes like “The One with the Male Nanny,” which gets into Ross’s insecurities about masculinity, but in a way that feels half-sincere, half-sitcom punchline.
To sum it up, Friends is complicated, which is part of what makes it so interesting to analyze.
Three Part Analysis
After collecting the scripts of every episode across all 10 seasons of the show (and an absolutely horrendous amount of data-preprocessing) it was time to analyze. I’m focusing my analysis on just the six main characters, because including side characters turned out to be too messy, especially when trying to infer gender or emotional tone from limited data.
Instead of going deep on just one question, I took a broader approach and ran three different analyses:
- What kinds of emotions are expressed?
I used an emotional lexicon (NRCLex) to label words associated with emotions like joy, anger, trust, or fear.
- How does a model interpret emotion in context?
I fed the lines into a pre-trained transformer model (via Hugging Face) to get a more nuanced emotional analysis.
- What do they actually say?
Using Log Odds Ratio (LOR), I tried to uncover which words are more likely to appear in gendered groups.
Emotional Lexicon: What is Being Expressed in Simple Terms
The NRC Emotion Lexicon is a widely used resource in NLP that maps individual words to basic emotions and sentiment categories.
The most striking pattern here is how female characters are more strongly associated with positive emotional expressions. Their dialogue contains more words linked to: Joy, Trust, Anticipation, and overall positive sentiment
Meanwhile, male dialogue contains slightly more anger-related words and is more emotionally balanced between positive and negative tones.
This matches common gendered portrayals in TV writing, where women are often written as emotionally expressive, optimistic, and relational, while men are portrayed with more emotional restraint or tension. Also interesting: fear is more present in female dialogue, which could reflect how vulnerability is more “allowed” or expected from women on-screen.
As a summary point, I decided to calculate the “emotional range” for each gender, based on these values, essentially the average number of emotions expressed. For the male characters, this value is 1.48 – and 1.59 for the female characters. In order to double check this, I decided to also check the range on a character basis, and sure enough, all the male characters are in the range 1.46–1.51, while the female characters are in the range 1.54–1.73. This is a pretty clear difference in the degree to which the characters express their emotions, showcasing further gendered emotional stereotypes.
Pre-trained Model: Emotions in Context
To further explore emotional patterns in the dialogue, I used a pre-trained transformer model that predicts the most likely emotion of each line of dialogue. Then, I grouped these predictions by the gender of the speaker to see if certain emotions were more associated with female or male characters.
Here we see some interesting patterns:
- Curiosity dominates across both genders, but slightly more so for female characters. This might reflect how often female characters ask questions or play more passive roles in scenes. Men are the actors, women are reactors.
- Neutral speech is the most common category for both genders (expected), but male characters show a slightly higher percentage of neutral emotion, which might indicate less overt emotional expression in general.
- Female characters show slightly more “soft” or socially emotional states, such as:
- Caring (2.13% vs. 1.89%)
- Love (1.59% vs. 1.39%)
- Sadness (1.05% vs. 0.81%)
- Remorse (1.67% vs. 1.19%)
- Surprise (4.19% vs. 2.60%)
- Male characters show higher rates of amusement (1.45% vs. 1.17%), confusion (1.42% vs. 1.15%), and joy (0.96% vs. 0.84%), which may reflect the show’s tendency to play male characters (especially Joey and Chandler) for comic relief.
- Emotions like disgust, embarrassment, and fear are very rare overall, but the slight variations are still interesting. For example, disgust is more associated with female dialogue, while embarrassment slightly skews male, possibly reflecting how emotional self-consciousness is played differently across characters.
The differences aren’t dramatic, but they reflect something deeper. Female characters express more relational and vulnerable emotions; male characters show more detachment. This is especially interesting considering these kinds of shows often lean on emotional contrast for humor and tension.
While this is a completely different approach from using the NRC Emotional Lexicon, both techniques mirror some of the same patterns, mainly:
- Women express more positive, relational, and vulnerable emotions
- Men express slightly more neutral, comedic, or restrained emotions
Log Odds Ratio: Which Words Are More Likely to Appear in Gendered Groups
Emotions are interesting, but it’s hard to get a feel for the characters without knowing what they’re actually saying. Log odds ratio (LOR) with smoothing is a powerful way to find distinctive words used by different groups, like gender in this case. Using this technique, we can compare the kinds of words more likely to be used by male versus female characters.
Here are some of the most interesting highlights. Some of the most disproportionately “female” words include:
- “gosh”, “ugh!”, “woo!”, and “uh-hmm” – These reflect expressive interjections often tied to emotion, discomfort, or social energy, which is traditionally associated with femininity.
- “purse”, “client”, “catering” – These suggest that women are more often linked to service-related or appearance-based roles, which again could be echoing gendered stereotypes.
- Names like “Joshua”, “Barry”, and “Gavin” – Romantic interests of Rachel, which could point to how often female characters are defined in relation to dating or romance.
On the male side, we see a different vibe:
- Words like “dude”, “woah”, “tryin'”, “push!”, and “lead” lean into casual bravado, assertiveness, or action-oriented behavior.
- “fear”, “correct”, and “everyday” might hint at more abstract or rational language, or even authority, but it’s hard to tell.
- And then there are some less clear but very “guy talk” terms like “comin'”, “quack”, and “tweet”, which could relate to casual speech or inside jokes.
Of course, none of this is surprising for a sitcom from the ’90s and early 2000s, but it’s fascinating to see how these gendered dynamics show up in the word choices, even down to interjections. What’s powerful about using NLP for this kind of analysis is that it takes something you might intuitively feel when watching the show and actually show it in the data, like how Joey or Ross “talk like guys.”
Conclusion
This analysis shows how Friends may reinforce gender stereotypes through language. By using NLP techniques, I found that the female main characters tend to express more positive and relational emotions, while male characters stick to more neutral or restrained tones. This matches the emotional roles often assigned to men and women in sitcoms.
While these patterns aren’t groundbreaking, they give us a data-backed look at how gender stereotypes can be ingrained in mainstream TV. In the case of Friends, it highlights how emotional expression and word choice reflect broader social and cultural norms.
Finally, by applying NLP techniques to something as popular and enduring as Friends, I’m hoping to show that language analysis isn’t just for academic papers or businesses but can also be a way to understand ourselves and how we’ve changed over time (or not).