Computational Analysis of Prose

• Prose is defined as extended written language analyzed via word-adjacency networks to reveal structural, rhythmic, and stylistic patterns.
• The approach leverages local network features of function words, achieving up to 95% accuracy in distinguishing informative from imaginative texts.
• Structural analysis of prose complements lexical methods, offering practical insights for genre detection, style transfer, and broader computational linguistics tasks.

Searching arXiv for the core paper and closely related work on prose classification and style transfer. Prose, in the context of computational analysis, denotes extended written language whose organization can be examined not only through semantics and lexical frequency but also through structural, rhythmic, and stylistic regularities. One influential line of work models prose as a word-adjacency network and asks whether broad prose classes can be distinguished from topology alone. In “Classifying informative and imaginative prose using complex networks” (Arruda et al., 2015), prose is represented as a complex network whose local measurements—especially symmetry and accessibility around function words—support supervised discrimination between informative and imaginative texts with accuracy of up to 95%. Related work further situates prose as a target of style transfer (Carlson et al., 2017), rhythm-based discrimination against poetry (Arruda et al., 2021), and transformation into other literary forms (Khanmohammadi et al., 2021), suggesting that prose can be characterized at multiple levels beyond semantics.

1. Prose as a computational object

In (Arruda et al., 2015), prose is treated not via its semantic content but via the structural patterns of how words are arranged. The study asks whether informative prose and imaginative prose can be distinguished purely from the topology and dynamics of a word network, especially focusing on function words. The two target classes are defined broadly. Informative prose is expository, factual, and non-fiction material designed primarily to convey information; in the experiments, this class is composed of 80 scientific manuscripts, 30 “miscellaneous” texts, and 16 biographies and related subjects. Imaginative prose is narrative, literary, and fictional material, where the primary goal is storytelling or aesthetic effect rather than explicit exposition; this class includes general fiction, romances, love stories, and other imaginative subgenres (Arruda et al., 2015).

The corpus is drawn from the Brown Corpus, and the labels “informative” and “imaginative” are derived from its genre annotations. To avoid bias due to different text lengths, the study balances the classes by taking only the 126 longest informative texts and matching them with 126 imaginative texts. Each pre-processed text becomes a network instance for classification (Arruda et al., 2015).

This framing places prose within a broader research program in which stylistic variation is operationalized as a transformation or discrimination problem. “Evaluating prose style transfer with the Bible” (Carlson et al., 2017) defines prose style transfer as “the task of rewriting a sentence such that we preserve the meaning but alter the style,” while “A pattern recognition approach for distinguishing between prose and poetry” (Arruda et al., 2021) analyzes prose through aural and rhythmic properties alone. Taken together, these works suggest that prose can be modeled as a structured signal whose class membership is recoverable from form as well as content.

2. Corpus construction and preprocessing

The preprocessing pipeline in (Arruda et al., 2015) consists of four steps: removal of punctuation, line breaks, spaces, numbers, and special characters; optional removal of stopwords, depending on the strategy; lemmatization; and POS tagging, following Greene and Rubin, to solve lemmatization ambiguities such as “works” as noun versus verb. The stopword decision is methodologically central because some experiments remove function words before network construction, whereas the best-performing setting keeps them (Arruda et al., 2015).

The distinction between content words and function words is decisive in the paper’s argument. Some experiments remove stopwords to focus on content words and their syntactic skeleton, but the most discriminative features ultimately turn out to be structural properties of specific function words such as “the,” “by,” “an,” “have,” and “it.” This makes the representation largely independent of explicit semantic modeling: semantics is not encoded directly; rather, the pattern of word-word connections defines a prose style signature (Arruda et al., 2015).

A related, though methodologically different, view appears in (Arruda et al., 2021), where prose is modeled from phones, rhyme events, and rhythmic windows rather than lexical semantics or syntax. That study uses North American English prose samples from technical books, novels, and news pieces, matched to poetry samples for length in phones and characters, and also controls punctuation frequencies. This suggests that prose classification can remain feasible even under aggressive abstraction away from conventional lexical content (Arruda et al., 2021).

3. Network representation of prose

The core representation in (Arruda et al., 2015) is the word adjacency model. Each distinct word after preprocessing is a node, and an edge is established between words that appear adjacent in the running text. The network is defined by an adjacency matrix $\mathbf{A}=\{a_{ij}\}$, where

$$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$

When the network is treated as undirected and unweighted, $a_{ij}=a_{ji}$; when weighted, edge weight can reflect frequency of adjacency (Arruda et al., 2015).

This model is explicitly structural. Degree corresponds to the number of distinct bi-grams a word participates in, and weighted degree becomes related to word frequency. The study emphasizes that adjacency acts as a syntactic-style approximation and that most syntactic relations can be recovered considering adjacency as context. The resulting graph is therefore not a semantic network in the strict sense, but a stylized representation of how prose organizes local word succession (Arruda et al., 2015).

The paper contrasts this network-centered perspective with more standard content-based approaches such as bag-of-words and latent semantic analysis. Those methods capture semantical content effectively, but the paper argues that textual structure analysis yields complementary information. A plausible implication is that prose style is only partially reducible to lexical choice; topological arrangement contributes an additional discriminative layer (Arruda et al., 2015).

4. Topological and dynamical measurements

The study computes a broad set of graph-theoretic and random-walk-based measurements, focusing on local properties per node and then aggregating or directly using them as features. Classical quantities include vocabulary size $V$, degree $k_i$, neighborhood connectivity, clustering coefficient $cc_i$, betweenness centrality $B_u$, closeness centrality $C_i$, eccentricity $E_i$, eigenvector centrality, PageRank, and modularity $Q$ (Arruda et al., 2015).

Two measurements receive particular emphasis. The first is accessibility, defined for node $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$0 and level $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$1 as

$$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$2

where $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$3 is the probability that a self-avoiding random walker starting at $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$4 reaches node $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$5 in exactly $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$6 steps. Accessibility is interpreted as the effective breadth of structural reach: if a node reaches many nodes with equal probability, accessibility is high; if access is concentrated on few nodes, accessibility is low. The generalized accessibility $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$7 extends this idea across walks of all lengths via $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$8 and $$a_{ij} = \begin{cases} 1 & \text{if there is a link } i \rightarrow j,\ 0 & \text{otherwise.} \end{cases}$$9 (Arruda et al., 2015).

The second is symmetry, defined using concentric levels $a_{ij}=a_{ji}$0 and the same random-walk probabilities: $a_{ij}=a_{ji}$1 Here $a_{ij}=a_{ji}$2 is the number of dead ends at level $a_{ij}=a_{ji}$3. Two variants are used: backbone symmetry $a_{ij}=a_{ji}$4, which removes edges between nodes in the same concentric level, and merged symmetry $a_{ij}=a_{ji}$5, which merges nodes connected within the same level into single nodes. These variants emphasize radial and angular regularity, respectively (Arruda et al., 2015).

A central finding is that symmetry and accessibility measurements are among the most prominent network measurements for prose discrimination. This suggests that what separates informative and imaginative prose is not merely local connectivity density or centrality, but the regularity and breadth of access patterns around particular words. The paper’s interpretation is that these measurements play a complementary role in characterizing texts (Arruda et al., 2015).

5. Classification methodology and quantitative results

The supervised task in (Arruda et al., 2015) is binary classification of each text network as informative or imaginative. The authors explore three feature-construction strategies. The global strategy without stopwords computes summary statistics such as mean, standard deviation, median, minimum, and maximum over local measures and includes modularity. The local strategy without stopwords uses per-word local measurements as features but restricts to topological measures that do not correlate with frequency. The local strategy with stopwords is identical to the local strategy except that stopwords are kept, allowing measurements for function words to enter the feature set (Arruda et al., 2015).

Feature selection is performed with information gain

$a_{ij}=a_{ji}$6

The classifiers are $a_{ij}=a_{ji}$7-nearest neighbors, classification and regression trees, and Naive Bayes with Gaussian feature likelihoods. Evaluation uses leave-one-out cross-validation over 252 documents, and the main performance metric is classification accuracy (Arruda et al., 2015).

The principal quantitative results are concise and strong.

Complex network approach	KNN	CART	Bayes
Global strategy without stopwords	72%	78%	75%
Local strategy without stopwords	92%	92%	92%
Local strategy with stopwords	95%	95%	95%

These results show that global network summaries reach up to 78% accuracy, local structure without stopwords reaches 92%, and local structure with stopwords reaches 95% (Arruda et al., 2015). The best setting is therefore the local strategy with stopwords, and the improvement from the global to the local setting is substantial across all three classifiers.

The paper also compares network-only methods with more traditional baselines. Stopword frequency achieves 97% accuracy with KNN, and character bigrams yield 98% accuracy. The network-only approach is therefore slightly weaker in absolute accuracy, but the paper stresses that symmetry and accessibility are largely uncorrelated with word or character frequency, indicating complementary stylistic information (Arruda et al., 2015). This suggests that prose classification may benefit from combining lexical and topological features rather than opposing them.

6. Interpretation: what the measurements reveal about prose

The interpretation offered in (Arruda et al., 2015) is that informative and imaginative prose differ in how function words are structurally embedded in the network of words. Informative prose tends to display more regular, stereotyped syntactic patterns, including recurring noun phrase patterns and prepositional structures. Function words such as “the” and “by” are used in a more stable set of syntactic frames, producing higher symmetry and more predictable neighborhoods. In PCA plots of stopword features, the cluster of informative texts is more compact, indicating less stylistic variability (Arruda et al., 2015).

Imaginative prose, by contrast, tends to exploit a larger variety of syntactic constructions and narrative patterns. Function words appear in more creative, context-dependent ways, giving lower symmetry and more heterogeneous local neighborhoods. The same function word can occur in structurally diverse contexts, including dialogues and varying clause structures (Arruda et al., 2015).

The most relevant features in the best setting are all symmetry or accessibility measurements of specific function words. The ranking reported in the paper begins with $a_{ij}=a_{ji}$8, followed by $a_{ij}=a_{ji}$9, $V$0, $V$1, $V$2, $V$3, $V$4, and $V$5 (Arruda et al., 2015). This makes the classification interpretable in stylometric terms: it is not arbitrary vocabulary but the local topology around high-frequency grammatical items that most strongly separates prose classes.

A partially analogous conclusion appears in (Arruda et al., 2021), where prose forms a dense, compact cluster under rhythmic features, while poetry is more dispersed. That study finds that prose tends to have fewer distinct rhyme classes, fewer repetitions, and many texts with $V$6, indicating sparse or highly regular rhyme windows (Arruda et al., 2021). Although the modalities differ—network topology in one case, rhythm and rhyme in the other—both works converge on the idea that prose occupies a narrower structural space than more stylistically flexible literary forms.

7. Context, limitations, and related directions

The work in (Arruda et al., 2015) belongs to a broader intersection of computational linguistics and complex network theory. Earlier network-based studies often emphasized global properties such as scale-free degree distributions, small-world effects, clustering, and modularity. The distinctive contribution here is the shift from global graph summaries to local, node-specific multi-scale features, especially around function words, and the application of symmetry and generalized accessibility as stylistic descriptors (Arruda et al., 2015).

Several limitations are explicit. The experiments are based on the Brown Corpus and present-day American English; the study treats only a binary classification problem; classifiers are used with default parameters; and much of the discriminative power is concentrated on a small set of function words (Arruda et al., 2015). These constraints limit direct generalization. A plausible implication is that performance and feature salience may vary across languages, translation regimes, or editorial normalization practices.

Nonetheless, the paper argues that the approach is general in principle: the word adjacency model and the symmetry and accessibility measures are language-agnostic provided tokenization and lemmatization are available, and the local topology around function words may serve as a stylistic fingerprint in other tasks such as authorship attribution, genre detection, readability and complexity assessment, and literary style analysis (Arruda et al., 2015).

This broader relevance is consistent with adjacent research. “Evaluating prose style transfer with the Bible” (Carlson et al., 2017) treats prose style transfer as rewriting that preserves meaning while altering style, using Bible versions as aligned stylistic variants. “Prose2Poem” (Khanmohammadi et al., 2021) treats prose as the semantic source in translation to ancient Persian poetry, emphasizing that prose is grammatically regular, relatively literal, and unconstrained by meter and rhyme. Such work suggests that prose is not a residual category opposed to poetry, but a structured domain whose organization can be modeled, transformed, and classified through multiple formal lenses (Carlson et al., 2017, Khanmohammadi et al., 2021).