Computational Narratology: Moving Beyond Textual Close Reading
The approach to narrative analysis that Elspeth Keller developed changed how I actually work with texts. It is not about reading more carefully. It is about measuring things that previous narratologists treated as intuitive judgments — sequence transitions, focalization shifts, temporal jumps — and turning them into observable data points. Most academic programs still teach narratology as a purely interpretive exercise. I stopped treating it that way around 2018 after spending a year trying to reconcile qualitative readings with actual textual evidence that contradicted the readings. Keller's central contribution was formalizing narrative units into analyzable segments and then mapping the transitions between them with a degree of rigor that had not existed before her work. The model she built, often called the Keller model or Keller formalism, breaks a narrative into three primary components: states, actions, and events. A state describes a condition held by a character or setting. An action is what a character does. An event is a change that occurs in the narrative world, usually triggered by an action or external force. The transitions between these units follow rules — not grammatical rules, but logical constraints about what can follow what in a coherent narrative sequence. The practical value of this is immediate once you start applying it. I used it on a corpus of roughly 40 novels for a comparative project, and what happened was that the model exposed patterns invisible during standard reading. Two novels that seemed identical in theme produced dramatically different transition maps when formalized. One had frequent state-to-state sequences with minimal action units bridging them. The other had almost no sustained states — every segment forced a transition. That difference explained the reader experience gap better than any thematic essay could have.
Here is the part nobody warns you about. The Keller model assumes that narrative coherence follows identifiable transition rules, but not all texts cooperate with that assumption. Modernist and postmodernist works, particularly those using fragmented timelines or unreliable narration, will produce formalizations that look like noise. I spent weeks trying to force a narrative by W.G. Sebald into the standard Keller framework before accepting that the text was deliberately violating the transition constraints the model relies on. The workaround was to treat the violations themselves as data — mapping which rules were broken and where, rather than trying to smooth them out. That shift in approach actually produced more interesting results than the formalization itself. The terminology matters because you will encounter different labels across departments. What Keller called state-action-event sequences, some narratologists refer to as functional units or narrative propositions. The underlying structure is the same. The labeling difference mostly reflects whether the researcher comes from a computational background or a literary theory one. If you are reading secondary literature and feeling lost, check whether the author is describing the same formalism under a different name before assuming two competing theories are in play. Applying the model correctly requires a specific workflow. You do not start by analyzing the whole text. You segment first. Choose a passage — ideally something between 2,000 and 5,000 words for a first attempt — and identify every state, action, and event. This is slower than it sounds. A single paragraph might contain three states, one action, and two events interwoven. Write them out as labeled tokens in order. Then map the transitions between tokens. If token A is a state and token B is an action, record that as a state-to-action transition. Continue until the segment is fully mapped.
The output is a sequence diagram or a transition matrix. From there you can count frequencies, identify dominant patterns, and compare across texts. A typical formalization of a 3,000-word passage takes about 45 minutes on the first pass. After you have done five or six, it drops to roughly 20 minutes. The bottleneck is always the initial segmentation. Experienced analysts spend more time debating whether a given sentence represents a state or an implicit event than they spend labeling it. There are software tools that claim to automate this process, and most of them are inadequate for anything beyond the simplest texts. The model requires human judgment on boundary cases — sentences that describe a character thinking about doing something, for example, which sits somewhere between a state and an action depending on contextual framing. A program cannot reliably make that call without extensive manual training data. I tried three different automated tools before settling on a manual coding approach with a spreadsheet-based tracking system. It was slower but more accurate, and accuracy is what matters when your analysis gets peer-reviewed. One counter-intuitive finding from using the Keller formalism repeatedly: longer texts do not necessarily produce more complex transition maps. In fact, highly plot-driven novels with frequent action events often produce simpler maps because the transitions follow predictable patterns. Texts with dense interiority — many sustained states, few physical actions — produce more varied transition profiles because the model has to account for more state-to-state sequences, which are harder to classify consistently. This is why the Keller model tends to favor certain genres over others when used for quantitative comparison. It is not a flaw in the model. It is a property of the model that you need to account for when drawing conclusions.
Get the Full Details
The biggest limitation of the Keller approach is that it describes structure, not meaning. A perfectly mapped transition sequence will tell you how a narrative is organized, but it will not tell you what the narrative means. I have seen researchers treat the formal output as a complete analysis, which is a category error. The model is a tool for structural description, not interpretation. Use it alongside traditional close reading, not as a replacement. The combination produces stronger results than either method alone, but only if you treat them as complementary rather than competitive. If you want to study the original formulations, look for Keller's work on narrative sequence analysis and her publications related to the University of Hamburg's narratology projects. The English-language secondary literature is thinner than you would expect for how influential the model became. Most of the detailed technical discussions are in German. If you read German, you will find more nuance. If you do not, the translation gap is a real obstacle that slows adoption in Anglophone programs. A common mistake beginners make is applying the Keller model to very short texts — under 500 words. The transition patterns require enough data to become visible, and short excerpts produce maps so sparse that any conclusion drawn from them is unreliable. Start with complete chapters or novella-length sections. Build confidence with the method before attempting shorter passages. The model works at smaller scales, but the signal-to-noise ratio drops sharply, and the effort required per word of output becomes unsustainable.
