Measuring the rate at which fiction's genres form and mutate, by modeling literature as a growing network.
Phase 1 — 345 canon novels, one per author (166 authors), unsupervised genre recovery from raw prose. This is the other of the two networks the pipeline builds; Phase 2 (author-to-author influence) is the other interactive view →, and the full method, corpus discipline, and honest limits for both are in the paper →
Phase 2 — 77 authors, Homer to the 1920s, directed edges scored two independent ways. This is one of two networks the pipeline builds; Phase 1 (unsupervised genre recovery from prose) is the other interactive view →, and the full method, corpus discipline, and honest limits for both are in the paper →
English Literature's genres have come to serve as its key mode of classification. However, the rate at which genres and subgenres are formed throughout English Literature's history is hardly definitive and primarily determined by experts in the field in a non-analytical fashion — that is, purely on the basis of opinion. Quantitative research to date has primarily focused on the semantics of literature's text and discerning its attributes. Notably, quantitative formalism at Stanford's Literary Lab has shown that natural-language-processing techniques are capable of recognizing novelistic genres by the semantics of a text [8]. Yet this method — and any other quantitative method, as far as published research is concerned — has yet to be applied to analyzing the formation of subgenres and genres as a whole over literature's history, and more specifically to the rate at which new genres and mutated subgenres form. Because literature's genres are hardly static, and contain multitudes spanning poetry, drama, non-fiction, and fiction, this research is focused solely on fiction. I propose to study English Literature's genre-mutation rate, within the bounds of fiction, throughout history — to form an expected value of this rate for the future.
Up until very recently — relatively speaking — all research in the realm of English Literature was done by non-analytical means. The Western canon, the classification of novels into distinctive genres and subgenres, and even the novels taught in schools fell to the opinionated analysis of the intellectually influential throughout the ages. But for the first time in history, this has begun to change. With the increasing technical aptitude of today's researchers — even those not specifically working in technical fields — we have seen a rise in quantitative techniques applied to fields of historically non-analytical context, more commonly known as the emerging Digital Humanities. This has been further heightened in machine learning and specifically natural language processing, with the increased availability of open-source libraries that allow such analysis to be done with relative ease. And while the foundations of English Literature can hardly be updated to be formed purely on the basis of quantitative analysis — nor should they be — there is still room for quantitative analysis to reveal traits of the field that are veiled to purely qualitative methods.
For starters, how would one go about outlining the mutation of fiction's genres using only qualitative means? It would most likely begin by reading hundreds of books, with subsets representing a sparse variation of each genre and subgenre. From here, you would look for semantic indicators within the works themselves, note publication dates, and research the political context at the time of publication to form a timeline of how such genres formed. The obvious flaw in this method is the inherent dispositions we have as humans. These range drastically and in varying degrees, but ultimately inhibit our ability to say decisively that a novel should or shouldn't belong in a given genre. In the ideal case, quantitative formalism provides the means to forgo such dispositions and derive genres by their semantic form, content, and style — much to the utopic state of the term. And perhaps the most important aspect of this quantitative method is that it allows us to examine thousands of texts on a year-by-year basis, per a work's publication year, to build a complete picture of genre inception and mutation.
While quantitative formalism provides a means of analyzing texts with minimal disposition, it is limited in viability by the availability of necessary data. That said, quantitative methods are still available for this research problem, and ultimately provide the same ability to discern meaning in a field where little to none currently exists.
This research is dependent upon acquiring data surrounding literature published on a year-by-year basis for a given period of time. While it would be ideal to capture every work of fiction published in digital form, spanning as far back as the written word, this is simply not feasible. Instead, there are a couple of ways to build a limited subset of data to serve as the basis of this research.
(1) Public-domain corpora. In the United States, novels published before January 1, 1925 have lost copyright protection and are in the public domain. A number of corpora across the internet maintain tens of thousands of such texts — notably Project Gutenberg and Google Books. While neither provides every work of English Literature for a given period (say 1825–1925), together they provide a core set of literature's most prominent texts, which are generally the kind of works that come to define genres as they are currently understood.
(2) Goodreads shelves. A second dataset uses the Goodreads API [4], or simply web-scrapes the Goodreads genre shelves [5], to build a core dataset of novels, their publication dates, and their modernly defined genres based on the shelves users file them under. The upside is that this dataset is far less sparse than the corpora in option (1) — more titles per year — and also provides data from 1925 to the present day, which helps tremendously in forming an expected value for genre mutation as it stands today. As Galton showed [3], it is perhaps more credible than one would initially think to leave the construction of genres to the everyday reader. The downside, when using a novel's Popular Shelves specifically rather than Goodreads' assigned genres, is that the data must be cleansed before being fed into the model — some user shelves such as "Want to Read" do not carry genre qualities the way a shelf like "1900s Westerns" does.
Due to the limited nature of the dataset in option (1), the basis of this project is founded upon the data built via option (2). That said, option (1) and its use case are discussed here, since it is the ideal dataset for this type of research and is worth contrasting with option (2) in order to more clearly define the basis upon which the network is formulated and analyzed.
Building the network from the acquired data is the next logical step. The network is modeled as a graph G = (V, E), where the nodes in V represent novels and the edges in E represent a relationship between two given novels. Because the relationship between two novels represents how similar they are, the edges are undirected.
The graph is iteratively populated on a year-by-year basis: a single new node is added to V for each novel published in a given year, and edges are added to E for each valid relationship between the new node v ∈ V and some other node u ∈ V, where the validity of the relationship — the similarity measure — is determined by a threshold function t(u, v) → {0, 1}.
G = (V, E), t(u, v) → {0, 1}
The natural question becomes how this threshold function is defined. Ultimately, it needs to discern novels that are similar — specifically in terms of the genre(s) each is classified as. Using the Goodreads dataset, we have each novel's shelves available. Goodreads shelves are the way users classify novels when saving them to their library; examples range from fiction's widely accepted genres such as mystery, romance, and fantasy to more exclusive subgenres like urban fantasy, gothic romance, and thrillers. From these shelves we build a factor. For example, if 1,000 users have a novel on their "Mystery" shelf and 3,000 users have the same novel on their "Thriller" shelf, the genre factor for this novel would be {Mystery: .25, Thriller: .75}. From this factor we can determine a level of similarity that serves as the basis for the threshold function deciding whether an edge e ∈ E is present between u, v ∈ V.
As the data is iterated on a year-by-year publication basis, the network continues to grow and clusters begin to develop — the largest clusters being fiction's genres, and sub-clusters being the subgenres of a given genre. From this arises perhaps the most pivotal consideration of the proposal: what properties define a cluster and sub-cluster in this network such that it accurately represents genres and subgenres within fiction?
One consideration is tuning the threshold function t(u, v) to generate a graph whose number of clusters corresponds to the number of widely accepted genres of fiction, using partitional clustering methods [1]. However, classifying definitive fiction genres is an area of intense debate and minimal consensus in English Literature. Additionally, a partitional method organizes data into non-overlapping groups, which seems like an over-generalization for a body of data that is excessively diverse and inherently spans multiple genres and subgenres. That said, it does provide a clear method of forming a desired number of genre clusters and tuning the threshold function to meet that criterion.
Alternatively, a hierarchical clustering method [1] seems more natural, since genres and subgenres are intrinsically related. This raises the further question of whether a top-down (divisive) or bottom-up (agglomerative) approach should be used — which leads to the consideration of whether genres are inherently formed by their subsequent subgenres, or whether subgenres form as an effectuation of overarching genres [6]. Perhaps this could be answered by watching the graph form as a whole on a year-by-year publication basis.
Watching it, 2026. The animation at the head of this page ↑ is that year-by-year formation, run on the rebuilt corpus — and it answers the question by dissolving it. Divisive against agglomerative presupposes that genres have formation events at all, and seven of the eight communities the pipeline recovers do not: they are perennial modes spread across the whole 250 years, not born at a point. Only detective fiction is a datable emergence (§8). The mutation ledger running beneath the animation does lean divisive — 33 splits against 25 merges over the full 345-novel run — but it cannot carry that weight, because shuffling every publication year produces just as many events. What the graph shows is the shape of genre formation: not its direction, and not its rate.
From the formation of a network of fictional novels based on genre similarity, we can ultimately discern the rate at which clusters fabricate and subgenres mutate. This rate, compared on a year-by-year publication basis, provides insight into the expected number of genre mutations within fiction's genus in the years to come.
The obvious questions concern how this expected valuation is determined. Is it simply a linear regression that forms this expected value? Or, due to the increasing variability of subgenres, should it be an application of the power law? Determination of an appropriate expected-value function would ultimately depend heavily on the properties of the resulting data itself.
It turned out to be neither. The question this section poses — what function fits the mutation rate — assumes there is a rate to fit. Built and measured, there is not: the apparent acceleration toward 1900 is the canon simply holding more books per later year, and per book the rate is flat across 270 years. This was the proposal's central quarry, and it is the one thing the rebuilt pipeline decisively fails to find (§8).
As noted earlier, an alternative dataset for this research would be digital copies of novels analyzed using quantitative-formalism techniques to discern genre from the ground up — that is, with as minimal human disposition as possible. This dataset would instead utilize Language Action Types (LATs) to quantify the type, content, and style of novels. LATs, as used by Jonathan Hope and Michael Witmore [7] via the text-tagging device Docuscope [2], are a collection of functional linguistic categories that classify more than 200 million possible strings of English. Thereby, LATs would replace Goodreads shelves when classifying a novel's inherent qualities.
Whether using LATs or Goodreads shelves, once the underlying novels have their genre properties exposed quantitatively, the process of building the network, performing cluster analysis, and evaluating expected values for genre mutation would be nearly identical — with the exception of the threshold function, which would be tuned specifically to the data being used.
While there are uncertainties surrounding the acquisition and preparation of the data required for this research, there is a straightforwardness of method and an extensiveness of impact that make intriguing results a strong possibility. Moreover, for hundreds — if not thousands — of years, English Literature has perhaps been the most kaleidoscopic examination of the human condition we have come to acquaint ourselves with. Divulging its inherent genre patterns therefore seems a worthwhile and edifying venture.
The proposal above was written in 2021, before the tools to run it properly existed. In 2026 it was rebuilt end to end — and the honest result is one part vindication, one part cautionary tale.
The Goodreads-shelf plan (§2) was abandoned: shelves are reader labels, so clustering them only rediscovers the categories you fed in. Instead the corpus was defined top-down — a cross-referenced canon of 345 pre-1929 English novels, each confirmed across several canon lists and two independent language models, then matched to its real full text on Project Gutenberg. Genres are learned from the prose itself (distinctive vocabulary → a k-nearest-neighbor similarity graph → community detection), never from labels. Held-out Gutenberg subject tags are used only to check the result.
Finding one — genres are real, and recoverable from words alone. The network rediscovers the recognizable genre system — detective fiction, science fiction, nautical adventure, historical romance, the courtship novel — and each emergent cluster matches the held-out label it was never shown. Explore it live above: click any novel, or a genre pill, to trace its cluster — every genre shows its temporal-concentration z-score and top distinctive vocabulary, plus which single genre (detective fiction) is a genuine, datable emergence rather than a perennial mode. Switch to the interactive genre network ↑
Finding two — there is no global "mutation rate." This was the proposal's central quarry (§5), and it does not survive scrutiny. The apparent acceleration toward 1900 is an artifact of the canon simply containing more books per later year; measured per book, the rate is flat. A null model settles it: shuffling every novel's publication year yields the same number of "mutation" events as real history (z = −0.27). The events are not chronology-driven.
Finding three — but one genre genuinely emerges. After controlling for three confounds that each faked a signal in turn — corpus density, the secular drift of prose style, and prolific-author voice (a "science-fiction" cluster turned out to be H. G. Wells alone) — exactly one genre is significantly concentrated in time: detective fiction (z ≈ −3.0, ~1840s–1920s), robust across parameters and validated by its held-out label. Everything else is a perennial mode, present across all 250 years. Genre birth is real, but rare and genre-specific — not the smooth universal rate the proposal imagined.
Tested again, harder. That test measures concentration as the standard deviation of a community's publication years, which is blind to a cluster with two tight modes far apart — and two communities are visibly bimodal, one with a 51-year hole at 1837–1888. So the whole graph was re-cut at seven Louvain resolutions across ten seeds and every sub-community re-tested. The split comes from vocabulary alone; publication year never enters the graph, so it cannot be circular. The conclusion holds: detective fiction is still the only survivor. The Gothic never separates at all — Otranto, Vathek and The Monk stay bound to Frankenstein, Dracula and The Time Machine under every resolution and seed, because their vocabulary genuinely is shared. The hole at 1837–1888 is a gap in the corpus, not a boundary between genres. Two candidates come close without clearing correction — Victorian sensation fiction (1848–1899) and American naturalism (1899–1920) — and are what a larger corpus should settle.
Full pipeline, controls, and the negative results are open source: github.com/2016judea/literature-mutations ↗
A separate question, asked with the same discipline: forget genre — does textual similarity between two specific authors track real, documented influence, once you control for the fact that a later author is chronologically free to resemble everyone who came before them? Phase 2 reuses the corpus-building rigor above but builds a different graph: directed, author-to-author, candidate edges permitted only forward in time, carrying two independent similarity scores that are never merged into one — stylistic (word choice, syntax) and conceptual (ideas, themes, via embedding).
build_bibliography.py cross-references 2,411 works across 108 authors; build_corpus.py resolves 583 of those to real Gutenberg prose across 77 authors, Homer through the 1920s; build_influence_graph.py builds 2,915 directed candidate edges. Explore it live above: click any author to trace their edges — every connection shows both scores side by side, plus the real citation where one is independently documented.
Finding one — conceptual similarity is real, and it replicates. Held out against 130 documented influence pairs (Pound on Eliot, Emerson on Thoreau, Wagner on Nietzsche), conceptual similarity is significantly elevated over a shuffled-timeline null (z = 9.47). A second, independent, non-LLM validation source — Wikidata's structured "influenced by" property, 102 more pairs — replicates it (z = 7.16). A density control, restricted to the 47 best-represented authors, confirms it isn't primarily a density artifact (z = 6.25). Three independent checks, same direction: documented influence really does look like elevated conceptual similarity.
Finding two — stylistic similarity is a genuinely open question. On the full 130-pair sample it is not significant (z = 0.91) — word choice and syntax alone don't track documented influence. But it turns significant in both narrower, independent checks: the Wikidata pairs (z = 2.45) and the well-represented subset (z = 2.97). This is left deliberately unresolved rather than adjudicated: either the full-sample null is genuinely flat and both narrower checks are small-N noise landing the same lucky direction, or the full-sample null was itself partly an artifact that both narrower checks happen to correct. That dedicated pass has now been run, and it tilts the answer.
Raising the minimum books per author and re-testing at each level — against a null drawn from that same subset, since more text raises every pair's similarity, real or random — produces no trend. Stylistic z runs 0.9, then −0.7, then 0.3, before a single excursion at ≥4 books that decays again by ≥6. Conceptual similarity, run identically as a control, is significant at every level and declines smoothly as pairs are lost, which is what a real effect losing power looks like. The published 0.91 / 2.45 / 2.97 were three non-nested samples, not a series in representation. Dropping 48 pairs between ≥3 and ≥4 books flips z from 0.31 to 2.97; that is fragility, not signal. The honest reading is that the stylistic effect is most likely a null, and the earlier refusal to adjudicate was the right call.