We have finally arrived at the final question of the Bechdel Test: are two named women speaking about something other than a man? 

 

There are many ways to go about asking this question, computationally, and I would like to forefront this portion of the tutorial by stating that (like other parts of this tutorial) my particular method for identifying conversational topics is not the only method, and likely not even the best. Initially, I considered using topic modeling in order to identify the topics of dialogue between two women. That methodology is still very much open to anyone wishing to modify this code. I eventually decided that topic modeling, because of the unstable nature of the topics produced by topic models, would not be the most effective methodology for getting at the question the Bechdel Test asks. What I mean by this, is that the fact that topics can change drastically given small changes to the input data meant that the topics provided by the code would most likely change on every run of the code. Additionally, topic modeling works best on large sets of data and given that the set of nineteenth-century, fictional dialogue between women is likely to be fairly small in any given novel, I determined that topic modeling was not the most effective methodology with my current understanding of how topic modeling works. You can read more about some of the drawbacks of topic modeling here and some of the strategies for improving topic modeling stability here.

 

Therefore, I determined that the most straightforward way to answer this third Bechdel Test question, is to modify the code from question two to run in reverse. In question two, we filtered the dialogue that we identified to be occurring between two or more characters to only keep dialogue wherein either female pronouns, occupations, or names we have determined to be female. For question three, we are additionally going to filter out any instances where a male name, male pronoun, or male occupation is mentioned. Effectively, we are going to change the bag of words that we developed in question two to reflect maleness. 

 

Some of the limitations of this methodology, is that effectively we will miss almost every instance of inferred dialogue. This methodology only extracts instances where a female character directly mentions another female character by name and only filters out instances of dialogue where that character is speaking directly about a male character. However, while this method certainly has its limitations, it does provide a more definitive solution in the sense that it will accurately determine if a female character is speaking to another off the topic of men (unless they are speaking indirectly and unless there has been a mistake with gendering the name of the character). 

 

Additionally, similarly to how there were many instances in question two where it was crucial to check our own biases as programmers, when constructing our bag of words in order to represent maleness, we must make sure that we are not recreating gender bias. Especially when we are adding occupations to our list of words, be careful to only pick occupations you can say with some certainty are male. For instance, we can say with certainty that the word “Master” is gendered male because of its female counterpart, “Mistress.” Similarly, we can say with some certainty that the word “butler” is male because of the counterpart “maid.” Other terms, like “chef,” “firefighter,” “police officer,” etc are much more vague though we can certainly make the argument that in the nineteenth-century, these occupations were definitively male. However, in making this argument, we are erasing the possibility that perhaps, there might be some nineteenth-century novel in which a woman is also a police officer, a chef, or a firefighter. 

 

The line-by-line explanations of the code below will be relatively brief given that much of the code is coming from question two.