More on the Bechdel Test

I gave some theoretical insights on the Bechdel test in a previous post, but silly me, of course there is real data! The Cornell Movie-Dialogs Corpus[1] contains conversations between characters in 617 movies.

Conversations in this corpus are already separated, so it’s easy to tell when two people are talking to each other. Most characters are annotated with a gender. Most, but not all. I inferred gender based on the census’ list of popular boys and girls names[2], this added some more information. All in all there were 9,035 characters: 3,027 male, 1,572 female, and 4,436 unknown. Lots of unknowns unfortunately, which means I wouldn’t trust these numbers too much on an absolute scale.

We do have a natural comparison. The actual Bechdel test requires two women talking to each other about something other than a man. We can easily construct a male version: two men talking to each other about something other than a woman. I’ll be comparing these quantities.

Character Ratios

First a quick pass through to count the number of male/female characters. I took the log2 ratio of male/female characters so that the view would be symmetric. A perfectly balanced cast would be at 0, +1 means twice as many male characters, -1 means twice as many female.


The overall median is a 2:1 ratio of male:female characters, and it’s remarkable consistent across genres. There is a pretty wide variance, which may be due to the incomplete gender-tagging of names in the corpus.


Now the hard part. We need to identify conversations which are between two women only, and about something other than a man. I’m also doing the reverse, identifying conversations between two men which are about something other than a woman, for comparison.

Checking the gender is straightforward (it’s either annotated in the database or its not) and I’m only counting convos that pass if both characters are KNOWN to be women(men). So characters with unknown gender are excluded.

Checking the topic is a bit harder. The method I’m using is simple: check for the presence of a male(female) character name (in the same movie) in the conversation, as well as known male(female) pronouns. Obviously this isn’t perfect, but since I’m doing an apples-to-apples comparison between men and women any flaws should balance out. Technically the Bechdel test only requires 1 passing conversation, for robustness in this analysis I required 2 per movie.


Number of Movies Passing Each Version


Fraction of Movies in Genre Passing Each Version

The top graph shows movies by total count, the bottom shows by fraction. Nearly all movies pass at least 1 version. About 75% of movies (red + blue) pass the male version, while about 40% (blue + purple) pass the female version. Action and adventure movies are the most male-biased (surprise!)[3]

Romance, comedy, and horror come the closest to parity. I’m surprised about the last category, I would’ve that horror would be male-dominated.  And even animation had very few movies passing; won’t somebody think of the children! There were only 10 movies in this genre though so it may not be representative.

Looking only at movies which passed each respective test, we can see how many passing conversations existed:


This may be a bit hard to read. Blue is female, red is male, they’re next to each other by genre, and the y-axis is the number of passing conversations per movie (on a log10 scale). For the most part, movies which pass the male Bechdel test pass a whole lot more than then female. The median number of male-passing conversations is about 40, for female it’s only 10.

That’s a 4:1 ratio, twice as much as the 2:1 ratio we saw of characters. Which is what one might expect given the bias for male charecters, as the number of possible conversation pairs are ~(number of characters)^2. Or it could be that the male characters are more prominent in the story, and hence occupy more screentime.

Other Resources has an enormous manually curated list of movies and their passing status. This post also has some excellent visualizations, based on a much larger set of movies. And near and dear to my heart, there’s an analysis of every Star Trek episode on The Mary Sue Blog.


Facebooktwitterredditpinterestlinkedinmailby feather
  1. [1]Cristian Danescu-Niculescu-Mizil and Lillian Lee. 2011. Chameleons in imagined conversations: a new approach to understanding coordination of linguistic style in dialogs. In Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics (CMCL ’11). Association for Computational Linguistics, Stroudsburg, PA, USA, 76-87.
  2. [2]
  3. [3]Neither of the modern Tombraider movies pass (according to, despite starring a woman, because she’s the only one
This entry was posted in Text Mining. Bookmark the permalink.

Leave a Reply

Your email address will not be published. Required fields are marked *