Foundations of Computational Thinking for Humanists: Paradigms, Pitfalls, and Provenance

The transition from humanistic inquiry to computational engagement is not merely an adoption of new tools; it represents a fundamental epistemological and methodological realignment. For the expert humanist—steeped in traditions of close reading, hermeneutic spiral, and contextual nuance—the computational paradigm initially appears alien, even antithetical, to the core values of the discipline. The essence of this chapter is to codify that realignment, to articulate the new grammar of research that emerges when scholarly questions are reframed as computable operations. This is not a call to abandon the interpretive richness of the humanities, but rather to augment it with systematic rigor, scalability, and new forms of evidence. We will establish that computational thinking for humanists is characterized by three interlocking pillars: the formalization of research questions into algorithmic processes; the critical awareness of scale inherent in "distant reading" methodologies; and the non-negotiable principle of data provenance, which anchors computational results in scholarly accountability. The chapter culminates in a technical deep-dive, demonstrating how the quintessentially humanistic process of editorial scholarship—the construction of a variorum edition—can be modeled as a directed acyclic graph (DAG) of transformations, thereby making explicit the implicit logic of the editorial workflow.

The Epistemological Shift: From Interpretation to Specification

At its heart, the computational turn demands a shift from a primarily *interpretive* epistemology to a *specification-driven* one. Traditional humanistic research often proceeds via a hermeneutic circle: a dynamic, recursive movement between part and whole, where understanding is gradually deepened through repeated engagement with a text or artifact. The output is an argument, a nuanced reading, an interpretation that acknowledges multiplicity and contingency. Computational systems, however, do not "interpret" in this sense; they execute specifications. A research question must be translated into a precise, unambiguous set of instructions—an algorithm—that operates on formally represented data. This act of translation is the first and most critical skill of the computational humanist.

Consider the seemingly simple question: "How does the portrayal of urban space evolve in Victorian novels?" A traditional approach might involve selecting a handful of canonical texts, performing close readings of relevant passages, and crafting an argument about industrialization, class, or narrative perspective. A computational approach requires specification: What operational definition will be used for "urban space"? (e.g., frequency of words from a predefined geographic lexicon, mentions of specific street names, co-occurrence networks of architectural terms). Which corpus of "Victorian novels"? (e.g., all novels in Project Gutenberg published between 1837-1901, filtered by language and length). What metric defines "evolve"? (e.g., trend analysis of term frequencies over publication year, topic modeling to discover latent semantic shifts, or change point detection in network centrality). Each of these sub-questions must be answered not with a scholarly intuition but with a concrete, implementable rule. The humanist's intuition guides the *design* of the specification—the choice of lexicon, the definition of the corpus—but the execution is deterministic.

This specification imperative exposes a common pitfall: the "black box" syndrome. A humanist might run a topic model and produce a list of topics, then proceed to interpret them. This is computationally naive. The responsible scholar must understand the *provenance* of those topics: Which algorithm was used (Latent Dirichlet Allocation, Non-Negative Matrix Factorization)? What were its hyperparameters (number of topics, alpha, beta)? What was the preprocessing pipeline (tokenization, lemmatization, stop-word removal, stemming)? How was the number of topics $k$ chosen (coherence scores, manual inspection, elbow method)? The output is not a revelation but a *synthesis conditioned by a series of explicit, and often contestable, technical decisions*. The computational humanist must therefore become a critic of their own methods, tracing the lineage of every result back to its foundational assumptions.

Formalizing the Hermeneutic Spiral: From Spiral to State Machine

The hermeneutic spiral—the idea that understanding moves from a preliminary guess, through testing against the text, to a revised, more sophisticated guess—can be formalized as a state machine. In its simplest deterministic form, a finite state automaton (FSA) consists of a set of states, a set of inputs (symbols from an alphabet), a transition function, a start state, and a set of accept states. We can model a basic interpretive act as traversal through such a machine.

For instance, consider the interpretive task of identifying narrative perspective in a novel. A scholar might proceed through states: `InitialReading` -> `HypothesisFormation` (e.g., "This is third-person limited") -> `EvidenceGathering` (search for first-person pronouns, focalization shifts) -> `HypothesisRevision` (e.g., "Actually, it's omniscient with free indirect discourse") -> `FinalInterpretation`. Each state transition is triggered by specific textual evidence or analytical moves. The FSA model makes this process explicit and, crucially, *reproducible*. Another scholar, given the same transition rules and the same text, should traverse the same state sequence and arrive at the same final state (interpretation). The ambiguity and intuition of the humanist are not eliminated but are *compiled into the transition function itself*.

In practice, the states and transitions are far more complex and often probabilistic. A more realistic model uses a Hidden Markov Model (HMM). Here, the *observations* are the textual features (word choices, syntactic patterns), but the *hidden states* are the interpretive categories (e.g., `NarrativeMode: FirstPerson`, `NarrativeMode: ThirdPersonLimited`, `NarrativeMode: Omniscient`). The transition probabilities between hidden states capture the likelihood of shifting perspective, while the emission probabilities capture the likelihood of a given textual feature appearing under a particular narrative mode. Training such a model on a manually annotated corpus allows it to predict the most probable sequence of hidden states (interpretive labels) for a new, unannotated text. The HMM does not "understand" perspective; it computes the statistically most likely sequence of labels given the observed words and the learned parameters. The humanist's role is to define the state space, curate the training data (itself a hermeneutic act), and, most importantly, *critique the model's failures*—the places where the predicted state sequence diverges from what a human reader would deem plausible. These divergences are not merely errors; they are sites of new insight, revealing complexities in the textual phenomenon that the simplified state model cannot capture.

Distant Reading vs. Close Reading at Scale: A Spectrum, Not a Dichotomy

The dichotomy between "distant reading" (Moretti, 2000) and "close reading" is often overstated. Computational thinking does not replace close reading; it recontextualizes it and creates a new, symbiotic relationship. Distant reading is not "reading without reading"; it is reading *differently*, at a scale where macro-patterns become visible. Close reading, in turn, becomes the essential tool for *grounding, interpreting, and validating* the patterns discovered from a distance. The computational humanist must be fluent in both modalities and understand how to move between them.

Distant reading employs aggregate, statistical, or network-based methods to reveal structures invisible at the level of the individual text. Examples include:

• **Corpus-level stylometry**: U