Using NLP to Visualize Ulysses, Part One
My data science experience has, thus far, been focused on natural language processing (NLP), and the following post is neither the first nor last which will include the novel Ulysses, by James Joyce, as its primary target for NLP and literary elucidation. In this post I will explain why it’s such a perfect target, since Ulysses will likely be the focus of my next project. This will probably be a multi-part blog post.
About the Book
First off, why this book?
Ulysses, by James Joyce, has elicited just about every kind of response from readers since its publication in 1922, ranging from claims that it’s the pinnacle of modernist literature to claims that it’s a filthy, decadent depiction of obscenity and pornography (nonetheless glittering with Shakespearean intertextuality on nearly every page) which should be, and was, banned until the famous Supreme Court case, United States v. One Book Called Ulysses in 1933 decidedly readmitted it into the United States. Moreover, this decision highlighted serious, longstanding philosophical questions about the role of art and the right to literary expression.
I fall firmly into the former category, and I believe it’s an affirmative work of genius. The reasons for this include many of the reasons I use Ulysses for NLP: the intent of Joyce was to recreate the many different modes of human experience through language. Through the scintillating and narrowing confines of different languages, dialects, subdialects, profanities, connotations, grammars, and idioms, all of which cross different religious and cultural traditions from the Catholic Church to Irish nationalists, to the mundane domestic affairs of a house in Ireland in 1904, to the parallels of ordinary life found on one day in Dublin to the Odyssey, as well as Shakespeare’s Hamlet, through fixations on classical philosophy and human suffering expressed via a hallucinatory, drunken escapade in a brothel manifesting as both the climax of the novel and the resurrection of the dead, Joyce believed that dimensionality would emerge through the parallax of flowing between these different modes of language, or life.
If the ostensible larger project of data science is to provide conceptual clarity via statistical analysis, as well as actionable insight through computing via machine learning, feature engineering, and deep understanding of data structures with the mindset of a scientist, then I can think of no greater interdisciplinary project, at least in the realm of NLP, than Ulysses for the sake of validating Joyce’s larger project. There are deeper parallels between data science and literary criticism than I initially realized when I entered the field, particularly in the importance of understanding the data through exploratory data analysis. And the more experience I’ve gained, the more I’ve realized this is a STEM way of saying ‘cultivate emotional alignment and clarity through conceptually rigorous inspection until dimensionality emerges in the data’.
Hence, the following project will be an attempt to simply visualize Ulysses in a way inspired by a similar project.
The Motivation
I was initially inspired by this project. The project is presented in the form of an academic article on Thomas Pynchon’s V, another difficult English novel characterized by a fragmented plot and an unclear timeline, by Christos Iraklis Tsatsoulis, completed in 2013, and I remain continuously shocked that I haven’t found more projects like it. I presume this is due to a cultural gap between Data Science and Literary Criticism, for reasons most likely due to the ancient war between STEM and liberal arts.
To start, Tsatsoulis presents an overview of the novel from a literary perspective, going over chapter summaries and the two primary ‘storylines’ in the book, the V. storyline and the Profane storyline. For the reader’s sake, I’ll be transparent by saying that I haven’t read V. by Thomas Pynchon, nor do you need to have read Ulysses to understand either project. For this post, I would only like to outline and emphasize the ingenuity behind the overarching project, which is to make a true interdisciplinary effort to use the brilliant tools of contemporary NLP to augment literary analysis through both visualization and deeper understanding of the semantic content.
So often have I seen the combative attitude between ‘machine learning’ and ‘art’, always descending into the same pit of claims that ‘a computer can never make real art’ versus claims that ‘real art is simply a set of fundamental patterns which can be learned and replicated’, whether in the context of AI-produced music, literature, or any number of films about AI-related romance and love. Without getting into the tangential complexities of that debate, I only mean to point out how little cooperation there is between these general poles, which seem to correspond, again, to STEM and liberal arts.
What Tsatsoulis accomplished shows just how useful the tools of NLP can be for healing this strange adversarial relationship.
After presenting a literary overview of V., he then provides some exploratory data analysis, like any good data scientist, via a wordcloud and some of V.’s characterizing vocabulary. He explains his primary methodology for capturing the structure of semantic content throughout the novel, which involves TF-IDF and hierarchical clustering, as well as the interesting and original utilization of ‘distance thresholds’ between chapters, based on Euclidian, Manhattan, and Canberra distances, as well as an independent section on Normalized Compression Distance, a methodology based on Kolmogorov complexity. He uses these distance thresholds to create the bafflingly interesting visualizations for the novel:
This is an incredible application of NLP to literary analysis. Tsatsoulis even mentions:
Somewhat to our surprise, despite this universal agreement regarding the existence of two different storylines in the novel, it seems that there has never been an attempt to exclusively map each chapter to one and only one storyline.
Such a situation screams for the application of the tools of data science, and Tsatsoulis fantastically succeeded in applying them.
Why, then, given that this project was completed in 2013, has this methodology not caught on in the field of literary analysis? James Joyce himself is infamous for having said of Ulysses, to his French translator:
I’ve put in so many enigmas and puzzles that it will keep the professors busy for centuries arguing over what I meant, and that’s the only way of insuring one’s immortality.
And indeed, professors remain busy arguing over Ulysses. I will not even discuss — for now — the ultimate enigma that is Finnegans Wake for the potential application of NLP, though it may indeed be the telos project of NLP and literature.
Hence, my motivation for applying a similar methodology to Ulysses is inspired by the utter success of Tsatsoulis’s project with Thomas Pynchon’s V., not just because of the literary merit provided by such an analysis, but because it demonstrates that productive cooperation between data science, or the field of precision and rigorous statistical dominance, and literary criticism, the refuge of obscurantism and impenetrable vocabulary, is possible.
Look out for Part Two of this post, where I’ll actually attempt similar visualizations with the plot of Ulysses, which will hopefully align with standard interpretations of the novel.
