A Stylometric Murder Mystery, or: Poetic Justice by Mitzi Morris (Computers and Literature in Fiction, 3)
It is high time to talk about Poetic Justice by Mitzi Morris, the recent stylometric mystery novel I have been promising to write about in my post on Amanda Crosses Poetic Justice, also in the series on Computers and Literature in Fiction. (Sorry for not linking to the place where you can buy the book, but it’s already bad enough I paid them for my unfree copy of the book.) It is not easy to find out much about the author: there is a blog post under that name, discussing a 1964 children’s novel called Harriet the Spy on a blog called Criminal Element (the blog’s name also appears passingly in the novel, and there are more self-reflective elements in the novel), as well as a curriculum vitae of a software developer based in New York and working in the areas of bioinformatics and natural language processing (well, that would make sense), and one brief review, more of an announcement, really, on a blog devoted to Natural Language Processing and Text Analysis (that makes sense, too). But not much more. Anyway, here we apparently have a programmer who not only writes code, but also fictional prose. Nice!
The story is set in the circles of New York City’s book publishing industry and intelligentsia, which makes for a nicely entertaining read. It all starts (pretty soon at least) with Sheba Miller’s death on Staten Island, and this influential literary agent’s mysterious death then triggers quite a fascinating network of strange and more or less sympathetic characters to appear around her. Of course, it becomes clear pretty quicky that Sheba Miller was killed rather than died of natural death. Jay Alfred, the editor-in-chief of Ars longa, a small publishing house, and his partner Ken, start to investigate.
Besides this literal murder, the ramifications of the story slowly also discover a literary murder, that is a murder by killing several people’s reputation: Sheeba Miller’s supposed memoirs start circulating as pre-publication copies, and almost everyone in the industry looks stupid, perverted, or worse in it. Soon enough, there is quite a number of people with a vested interest in either uncovering or protecting the truth about both murder cases. Based on this basic constellation, the story is very well-told and allows you to build various hypotheses at a very nice pace, so it is neither frustrating nor boring to pick up the clues while the protagonists do the same, and the setting and characters are certainly entertaining.
But what I presume most of my readers here will be more interested in is what role stylometry plays in this novel and how it is being portrayed. Well, first of all, the solution to this novel’s murder mystery would not be possible without stylometry. One of the protagonists, Leni, who runs the online activities of Ars longa, is also a programmer and likes to create and try out fancy text analysis tools. So she does what she does best, and the question of who killed Sheba Miller is soon so closely related to the question of who really wrote these memoirs that it becomes almost the same. And this is what makes it so exciting!
So far, so good. One thing that I found an unnecessary novelistic ploy is the fact that the stylometric tool Leni puts together, and which she calls her “author-ID” programme, is somehow related to Google: of course, this is a legitimizing strategy, because we all trust Google for a bit of algorithmical magic; but in real life, the algorithms stylometry uses are relatively simple and well-understood; some of them are simply transferred from other domains to stylometry, some (like Burrows’ Delta) have been developed specifically for stylometry, but none of them has been developed by Google (yet).
Actually, a colleague’s friend puts Leni on the track of text analysis for author attribution. They are having lunch at some fancy or strange new restaurant (as most characters do most of the time), and the friend explains:
“Using some of the developer tools from Google,” Krishna said. “You can write a program to identify authorship.” He looked around the table at three puzzled faces. He tried to translate into non-technical terms. “For example, you think that maybe your ex-boyfriend wrote a nasty comment on your blog using a fake name. Take a bunch of his emails, stuff you know he wrote. You use this to build a program that can recognize his writing style. Then take the comments from your blog and feed it into your program to find out whether or not he wrote that too.” – “Aren’t you just supposed to delete it and move on?” Bryan said. – “Perhaps, if you’re a perfectly rational, mature, human being,” Krishna said. “For the rest of us, there’s forensic text analysis. Bitter, angry, text analysis.” (chapter 16)
Leni picks up on this quickly. What is really quite cool is how the reliability issue of stylometry is portrayed. First of all, Leni’s sceptical colleague gives her a test case to try the tool on and does not tell her that it is skewed; but Leni and her tool find out; so the tools is submitted to blind review. Later on, there is an episode where Leni means to stop her boss, Jay, from using the stylometric evidence to confound one writer they suspect of being the author of the memoirs (and maybe of the killing), and Jay says:
“Your author-ID program is a real asset. When I told Rupert that we’d analyzed his book via computer and could prove that his fingerprints were all over it, he became extremely worried. The minute I said, ‘irrefutable evidence’ he confessed immediately.” – “But it’s not irrefutable at all,” Leni protested. […]. – “You know that and I know that. Rupert doesn’t,” Jay said. (chapter 38)
So the perceived realiability of the seemingly empirical or scientific method is consciously used by the protagonists who know better. Well, we’d better not take them as our role models. Interestingly enough, the biological fingerprint metaphor comes up here, a metaphor I think is actually flawed, because linking stylometric affinity of two texts to an author’s identity is like linking fingerprints to personality traits, something which we don’t really believe in (at least, not anymore).
Also, the crucial role of parameter setting comes up in the novel, and finally, the issue of insufficient sample size is discussed, when Jay asks Leni to compare the mysterious memoirs to an individual book chapter she found posted on Facebook and written by the person she suspects, at that point, to be the author of the memoirs. This is what happens:
Forty-five minutes later Leni returned, still frowning. “Maybe,” Leni said. “The similarity score is OK, but it’s not significant.” “Significant?” Jay asked. “Statistically significant,” Leni said. “There’s not enough data to tell. There’s only one chapter posted on Facebook.” (chapter 41)
And so the method is allowed not to work its magic, for now, and the result remains unclear. (This is actually more realistic than the role of computers and algorithms in Mr Penumbra’s 24-hour Bookstore, where every thing works out, or almost). Still, later on, Leni does find out more, and this time she has better results, and reports to Jay: “The similarity score is 87. And it’s almost significant.”
I won’t say anymore about how this all continues, but to sum it all up, I think it is fair to say that this is the first stylometric murder mystery in the history of crime fiction, and it works as crime fiction as well! Hope to read more in this vein soon.