💬 NLP

You Talkin to Me?: A Network Analysis of Gendered Speaker-Addressee Patterns in Film Screenplays

This paper utilizes network analysis on film screenplays to demonstrate that gender bias in cinema operates not just through speaking time, but through the structural architecture of conversation, where male characters dominate both speaking and being spoken to, while cross-gender interactions create tighter dyadic bonds compared to the more diffuse same-gender exchanges.

Original authors: Samin Khan, Camilla Griffiths, Shrikanth Narayanan, Dan Jurafsky, Sabyasachee Baruah

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samin Khan, Camilla Griffiths, Shrikanth Narayanan, Dan Jurafsky, Sabyasachee Baruah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're walking into a bustling coffee shop, but instead of just counting how many people are wearing red hats versus blue hats, you're watching who is actually talking to whom. This is the world of computational social science, a field where researchers use computers to find hidden patterns in human behavior, much like a detective using a magnifying glass to spot clues others miss. In this specific corner of science, experts study networks—not the internet, but the invisible web of connections between people. They look at homophily, a fancy word for the idea that "birds of a feather flock together," meaning people often prefer to interact with others who are like them. They also study directed interactions, which simply means tracking who starts a conversation and who is being spoken to. Why does this matter? Because movies aren't just entertainment; they are giant mirrors that reflect our society's rules. If we only count how much time a character is on screen, we might miss the deeper story: who holds the power to direct the conversation, and who gets left out of the circle entirely?

This paper, titled "You Talkin' to Me?", dives deep into 38 movie scripts to answer a question that goes beyond "who speaks the most?" It asks: Who are they talking to? The researchers treated the movie dialogue like a giant game of "telephone" but with a twist: they manually mapped out every single time one character spoke directly to another, creating a massive map of 4,600 conversations. They wanted to see if gender acts like a magnet, pulling certain characters together or pushing them apart.

Here is what they found, and it's a bit more complicated than just "men talk more."

First, they looked at the whole picture. It turns out that men are the "VIPs" of the conversation map. They are the ones doing the most talking, but they are also the ones being talked to the most. It's like a party where the guys are standing in a tight circle, passing the ball back and forth, while the girls are often on the outside looking in. Even in scenes where there are just as many women as men, the women don't suddenly start a "girls' club" conversation. Instead, the women often turn their attention toward the men. The study suggests that simply having more women on screen doesn't automatically change who gets the spotlight; the men remain the central targets of attention, even when they aren't the majority.

Second, the researchers wondered: "When a man and a woman do talk, does the man talk more to her than she talks to him?" You might expect a one-way street where men dominate the conversation. But the data surprised them. When a man and a woman are in a scene together, the conversation is actually pretty balanced. It's a fair exchange, like a game of catch where both players throw the ball equally. The imbalance isn't happening inside the mixed-gender chats; it's happening because those mixed chats happen less often than the all-male chats. The exclusion happens before the conversation even starts, not during it.

Finally, the team looked at what happens after someone speaks. They used a concept called "participation shifts" to see if a conversation opens up to include more people or closes down into a private two-person chat. They found a fascinating pattern: when two people of the same gender talk (which is mostly men talking to men in these movies), the conversation tends to spread out. It's like a ripple effect; a third person jumps in, or the speaker turns to someone else, widening the circle. But when a man and a woman talk, the conversation tends to close up. It becomes a tight, back-and-forth loop between just the two of them, like a private bubble that keeps others out.

So, the big takeaway isn't that men are just shouting over women. It's that the structure of the conversation itself is biased. Men are the ones who keep the party going by inviting others in, while cross-gender chats tend to stay small and private. This means that even when women are present and speaking, the "flow" of the story often keeps them on the outside of the main group dynamic. The paper suggests that true representation isn't just about who gets lines; it's about who gets to be the center of the network, and who gets to decide who else gets to join the fun.

Technical Summary: A Network Analysis of Gendered Speaker-Addressee Patterns in Film Screenplays

Problem Statement
While prior computational research on gender in film has extensively quantified disparities in screen time, speaking time, and linguistic framing (e.g., agency and power), it has largely overlooked the relational structure of dialogue. Existing studies often fail to address the specific question of "who is spoken to," particularly in multi-party scenes where automated identification of the addressee is ambiguous. This paper addresses the gap in understanding how gender structures the architecture of conversation itself—specifically, the directed flow of attention between characters—rather than merely analyzing the attributes of individual speakers or the aggregate volume of speech.

Methodology
The study utilizes a manually annotated dataset derived from 38 film screenplays (spanning 1972–2017) sourced from the Internet Movie Script Database (IMSDB). The dataset comprises approximately 4,600 directed dialogue events where annotators identified the specific speaker and the specific addressee for each line. Character gender was inferred using the gender-guesser Python package based on first names, operationalizing perceived binary gender within the corpus.

The analysis is conducted across three distinct studies employing network analysis, statistical testing, and participation shift (p-shift) taxonomy:

Study 1 (Gendered Addressing Patterns): Analyzes the entire corpus and a subset of "female-dominant" scenes (where female speakers are at least as numerous as males). The study constructs $2 \times 2$ contingency tables of speaker vs. addressee gender and employs Pearson's chi-squared tests to determine if speaker gender predicts addressee gender.
Study 2 (Directionality of Cross-Gender Dialogue): Focuses exclusively on mixed-gender scenes to test for directional symmetry. It calculates the scene-level share of male-to-female ( $M \to F$ ) versus female-to-male ( $F \to M$ ) utterances. The study uses paired t-tests and Wilcoxon signed-rank tests to determine if cross-gender speech is directionally balanced or systematically skewed, while also visualizing scene-level heterogeneity.
Study 3 (Conversational Turn-Taking and Participation Shifts): Examines the downstream structural consequences of initial turns using Gibson's (2003) p-shift taxonomy. The study categorizes transitions between turns (e.g., reciprocation, third-party entry, continuation) and compares the distributions of these shifts following homophilous (same-gender) versus heterophilous (cross-gender) initial turns. Statistical significance is assessed via chi-squared tests and two-proportion z-tests with Benjamini-Hochberg false discovery rate (FDR) correction.

Key Contributions and Results

Dominance of Male Homophily (Study 1): Across the entire corpus, male-to-male ( $M \to M$ ) interactions dominate, accounting for 40.2% of all addressed speech. Crucially, even in scenes where female characters are numerically dominant, the interaction pattern does not shift toward female-to-female ( $F \to F$ ) dialogue. Instead, the dominant pattern becomes cross-gender, specifically female-to-male ( $F \to M$ ), which accounts for 44.6% of dialogue in these scenes. This indicates that increasing female presence reduces male-to-male dialogue but does not necessarily increase female-to-female interaction; conversational attention remains disproportionately directed toward male characters.
Aggregate Symmetry vs. Film-Level Clustering (Study 2): Contrary to the aggregate male dominance found in Study 1, cross-gender dialogue is directionally symmetric at the aggregate level ( $M \to F$ share $\approx$ 0.504). Statistical tests fail to reject the null hypothesis of symmetry. However, significant heterogeneity exists at the scene and film levels. Directional skew is not random noise but clusters by film, suggesting genre-specific conventions (e.g., action films tend toward $M \to F$ dominance, while horror/female-protagonist films show $F \to M$ skew).
Structural Consequences of Homophily (Study 3): The study identifies a mechanism explaining male centrality: the downstream flow of conversation differs by gender pairing. Heterophilous (cross-gender) turns are significantly more likely to result in direct reciprocation ( $AB \to BA$ ), creating tight, closed dyads. Conversely, homophilous (same-gender) turns are significantly more likely to result in third-party entry ( $AB \to CA$ ) or speaker readdress ( $AB \to AC$ ), effectively diffusing conversational attention to the wider group. Since homophilous turns in this corpus are predominantly male-to-male, male characters function as hubs that expand the interaction network, while cross-gender exchanges tend to close off into dyads.

Significance and Claims
The paper argues that gender bias in film dialogue operates through the architecture of conversation rather than solely through directional imbalance within cross-gender exchanges or simple disparities in speaking time. The findings suggest that male characters occupy a structurally privileged role not just as initiators, but as the primary recipients of speech and as the nodes that expand the conversational network.

The authors claim that traditional metrics (screen time, speaking time) fail to capture this relational dimension. By focusing on the "addressee" and the "participation shift," the study reveals that even when women are present and speaking, they are often excluded from the structural mechanisms that distribute conversational attention across a group. The paper concludes that these patterns reflect a deeper structural bias where male characters remain central as recipients and network expanders, while cross-gender interactions tend to be dyadic and self-contained. The authors note that these findings are specific to the stylized nature of film screenplays and call for future work to test these dynamics in naturalistic multi-party conversations.

More like this