Narrative Pattern Mining

• Narrative Pattern Mining is a computational approach that extracts recurring event patterns and structures from diverse narratives.
• It employs multiple representations—including event schemas, summary graphs, linguistic substrings, serial episodes, and dynamic network motifs—to capture core narrative structures.
• Various extraction workflows and scoring methods optimize for interpretability and non-redundant summarization across domains such as biomedical, fraud, and literary analysis.

Narrative pattern mining denotes a family of computational procedures for extracting recurrent structure from narratives, event-centric texts, and document collections. In the cited work, a narrative pattern is variously formalized as a latent set of “event patterns” mined from long reports, a weighted summary graph around query entities, a contiguous substring over a sequence of linguistic choices, a serial episode with gaps, or a dynamically evolving network motif over characters, sentiments, and topics (Luo et al., 29 May 2025, Kroll et al., 18 Aug 2025, Cortal et al., 9 Oct 2025, Tatti et al., 2019, Min et al., 2016). This suggests a field unified less by a single representation than by a shared objective: to recover compact, informative regularities that mediate between raw narrative material and higher-level interpretation.

1. Formal representations of narrative patterns

The literature defines narrative patterns at multiple representational levels. In event-centric generalized category discovery, the corpus is $X=\{x_1,\dots,x_N\}$, the latent pattern set is $P=\{p_1,\dots,p_K\}$, and narrative pattern mining is framed as learning a mapping $g_\phi:X\to P$, where $g_\phi$ is implemented via an LLM guided extraction process parameterized by prompt-templates and refinement steps (Luo et al., 29 May 2025). The pattern is not merely a cluster label: it is a succinct event schema intended to capture the core event structure underlying a category.

In biomedical literature discovery, the basic object is graph-theoretic. Each document $d$ is transformed into a labeled directed graph $G_d=(V_d,E_d)$, where $V_d\subseteq V$ is the set of detected entities and $E_d\subseteq V_d\times R\times V_d$ is the set of extracted binary statements. Given a query $Q\subseteq V$, the narrative pattern $P(Q)$ is a weighted summary graph $P=\{p_1,\dots,p_K\}$0, with $P=\{p_1,\dots,p_K\}$1, edges incident to at least one query entity, and weights $P=\{p_1,\dots,p_K\}$2 defined by a global score $P=\{p_1,\dots,p_K\}$3. By construction, $P=\{p_1,\dots,p_K\}$4 is a star-like subgraph around the query entities (Kroll et al., 18 Aug 2025).

In stylistic analysis of personal narratives, the representation is symbolic and sequential. A narrative with $P=\{p_1,\dots,p_K\}$5 clauses is mapped to a word $P=\{p_1,\dots,p_K\}$6, where $P=\{p_1,\dots,p_K\}$7. A narrative pattern is any contiguous substring $P=\{p_1,\dots,p_K\}$8, and the corpus-wide set of length-$P=\{p_1,\dots,p_K\}$9 patterns is denoted $g_\phi:X\to P$0 (Cortal et al., 9 Oct 2025).

In sequential pattern set mining, the relevant unit is the serial episode $g_\phi:X\to P$1, which occurs in a sequence $g_\phi:X\to P$2 if there exist positions $g_\phi:X\to P$3 such that $g_\phi:X\to P$4. Gaps are allowed, so the pattern captures ordered recurrence rather than strict adjacency (Tatti et al., 2019). In network-based narrative analysis, the pattern may instead be a dynamic configuration in $g_\phi:X\to P$5, augmented by sentiment trajectories and topic associations (Min et al., 2016).

Representation	Formal object	Source
Event schema	$g_\phi:X\to P$6, $g_\phi:X\to P$7	(Luo et al., 29 May 2025)
Summary graph	$g_\phi:X\to P$8	(Kroll et al., 18 Aug 2025)
Linguistic substring	$g_\phi:X\to P$9 over $g_\phi$0	(Cortal et al., 9 Oct 2025)
Serial episode	$g_\phi$1 with gaps	(Tatti et al., 2019)
Dynamic narrative network	$g_\phi$2	(Min et al., 2016)

These formalisms are heterogeneous, but all treat a pattern as an intermediate structure: more abstract than raw text, yet still interpretable enough to support retrieval, classification, clustering, or literary analysis.

2. Extraction and mining workflows

A central workflow in event-centric contexts is the “generate $g_\phi$3 refine” procedure implemented by PaMA. Narratives are first embedded by $g_\phi$4 and clustered into $g_\phi$5 groups via K-means. For each cluster $g_\phi$6, the LLM generates a candidate pattern $g_\phi$7 summarizing the dominant event sequence, and the pattern is then refined against labeled examples so that the final $g_\phi$8 aligns with human annotation criteria. PaMA fine-tuning proceeds through clustering, matching and ranking; pattern generation and refinement; pseudo-label reassignment; and contrastive and prototype learning (Luo et al., 29 May 2025).

The biomedical workflow is retrieval-centered. A user enters keywords, an autocomplete step matches them to canonical entities via dictionary lookup including synonyms, and an inverted index retrieves the set of documents containing all query entities. For each relevant document, the system loads a precomputed document graph, scores each edge incident to a query entity, aggregates scores across documents, sorts edges by descending score, and displays the top-$g_\phi$9 edges in an interactive network view with provenance access (Kroll et al., 18 Aug 2025). Here, mining is inseparable from exploratory search.

In personal narratives, the workflow begins with sentence segmentation using SpaCy’s English transformer model, followed by clause extraction via few-shot prompting of Llama 3.1 8B Instruct. Each clause is then labeled by the same model for process type, participant roles, and circumstances, after which process types are encoded symbolically as $d$0, $d$1, $d$2, and $d$3. Pattern mining itself is implemented by straightforward substring enumeration up to a maximum pattern length $d$4, producing exact counts and high-dimensional sequence vectors for later comparison or clustering (Cortal et al., 9 Oct 2025).

The MDL-based serial-episode framework separates search from evaluation. For a fixed candidate set $d$5, the subroutine Sqs$d$6 iteratively computes usages and gaps from minimal windows, performs interval-scheduling alignment to maximize overall gain, and recomputes counts until convergence. On top of this, Sqs-Candidates filters a pre-mined candidate set, whereas Sqs-Search is a parameter-free any-time algorithm that grows a pattern set directly from the data (Tatti et al., 2019).

The network framework for literary narratives begins with text segmentation into units such as chapters or scenes, character-timeline extraction, and construction of the dynamic network $d$7. Sentiment analysis produces chapter-level polarity values, while topic modeling produces topic-document and character-topic associations. Mining then proceeds through the joint examination of growth curves, centralities, community evolution, sentiment-weighted interaction maps, and topic exchange patterns (Min et al., 2016).

Taken together, these workflows show that narrative pattern mining can be retrieval-driven, clustering-driven, statistically driven, or compression-driven. A plausible implication is that the choice of workflow is determined primarily by the target inference task rather than by the narrative medium alone.

3. Objectives, scoring functions, and optimization criteria

In event-centric generalized category discovery, the objective is explicitly shaped by two challenges: divergent clustering versus classification groupings and unfair alignment for minority classes. PaMA addresses cluster reliability and dominance using

$d$8

and

$d$9

with $G_d=(V_d,E_d)$0 by default. High-confidence samples are selected using assignment probabilities from a Student-$G_d=(V_d,E_d)$1 kernel,

$G_d=(V_d,E_d)$2

and entropy

$G_d=(V_d,E_d)$3

The final objective combines instance-level InfoNCE, prototype-level contrastive losses for novel and known classes, and supervised cross-entropy: $G_d=(V_d,E_d)$4 Prototypes are updated by smoothing between the cluster centroid $G_d=(V_d,E_d)$5 and the pattern embedding $G_d=(V_d,E_d)$6 (Luo et al., 29 May 2025).

In the biomedical graph setting, local edge relevance is scored by

$G_d=(V_d,E_d)$7

and global pattern relevance by

$G_d=(V_d,E_d)$8

The three factors privilege edges that occur frequently in a document but are rare across the collection, whose mentions are spread through the document, and whose extraction method has higher confidence (Kroll et al., 18 Aug 2025). This is a relevance-ranking criterion rather than a learning loss.

In personal narrative style analysis, significance is statistical rather than predictive. For a pattern $G_d=(V_d,E_d)$9, a $V_d\subseteq V$0 contingency table is constructed from presence or absence in a target series and a norm baseline. The odds ratio is

$V_d\subseteq V$1

Fisher’s exact test yields a $V_d\subseteq V$2-value, and $V_d\subseteq V$3-values are Holm-Bonferroni-corrected across all patterns. Patterns with corrected $V_d\subseteq V$4 and $V_d\subseteq V$5 are flagged as significantly over- or under-expressed. Sequence comparison uses cosine similarity on exact pattern-frequency vectors, followed by hierarchical agglomerative clustering using Ward linkage and silhouette-based model selection (Cortal et al., 9 Oct 2025).

The MDL framework uses two-part coding. It seeks a model $V_d\subseteq V$6, consisting of a code table $V_d\subseteq V$7 and cover $V_d\subseteq V$8, minimizing

$V_d\subseteq V$9

The code table assigns a pattern-stream code $E_d\subseteq V_d\times R\times V_d$0, a gap code $E_d\subseteq V_d\times R\times V_d$1, and a no-gap continuation code $E_d\subseteq V_d\times R\times V_d$2. Pattern quality is therefore assessed by total encoded length rather than support alone, and the resulting optimization is over both the selected pattern set and its non-overlapping alignment on the data (Tatti et al., 2019).

In dynamic network analysis, scoring is distributed across topology, sentiment, and topics. The framework tracks $E_d\subseteq V_d\times R\times V_d$3, $E_d\subseteq V_d\times R\times V_d$4, degree distributions, clustering coefficients, betweenness centrality, and modularity. Sentiment uses the chapter-level polarity index

$E_d\subseteq V_d\times R\times V_d$5

pair-level averages $E_d\subseteq V_d\times R\times V_d$6, and cosentiment edge weights $E_d\subseteq V_d\times R\times V_d$7. Topic modeling uses NNMF on a TF-IDF matrix $E_d\subseteq V_d\times R\times V_d$8, followed by character-topic association scores $E_d\subseteq V_d\times R\times V_d$9 (Min et al., 2016).

These criteria reveal a major conceptual divide. Some approaches optimize for alignment with annotation conventions, some for exploratory ranking, some for statistical distinctiveness, some for compression, and some for structural dynamics. Narrative pattern mining is therefore not reducible to “finding frequent patterns.”

4. Domains, datasets, and empirical findings

Event-centric narrative pattern mining has been evaluated on two EC-GCD benchmarks and several base GCD datasets. The Scam Report dataset contains 14 known and 8 novel categories, more than 8K reports, and is highly imbalanced; the Telecom Fraud Case dataset has 8 known and 4 novel categories and consists of police transcripts. Evaluation uses $Q\subseteq V$0, $Q\subseteq V$1, and the harmonic $Q\subseteq V$2-score $Q\subseteq V$3. On Scam Report, PaMA achieves an $Q\subseteq V$4-score of $Q\subseteq V$5 versus a best baseline of $Q\subseteq V$6, a gain of $Q\subseteq V$7. On Telecom Fraud Case, it reaches $Q\subseteq V$8 versus $Q\subseteq V$9, a gain of $P(Q)$0. On BANKING, StackOverflow, and CLINC, it maintains competitive $P(Q)$1-scores, including $P(Q)$2 on BANKING. Ablations report that pattern refinement yields an approximately $P(Q)$3–$P(Q)$4 $P(Q)$5-score lift, ranking and filtering add approximately $P(Q)$6, and removing $P(Q)$7 or $P(Q)$8 causes drops greater than $P(Q)$9 (Luo et al., 29 May 2025).

In biomedical literature discovery, the evaluation is qualitative. Five pharmaceutical researchers, already familiar with PubPharm’s narrative service, participated in approximately 30-minute remote sessions via Zoom. The procedure comprised introduction and consent, think-aloud exploration of the “Pattern Discovery (Beta)” tab, and a semi-structured interview. Reported positive reactions included rapid graph-based overviews, appreciation for filtering by entity type and top-$P=\{p_1,\dots,p_K\}$00 edges, and trust arising from provenance access. Reported usability challenges included unclear autocomplete behavior, insufficiently descriptive “Add” and “Search” buttons, loss of graph context when an edge opened a new tab, and small or similar colors and icons. No quantitative precision or recall metrics were reported (Kroll et al., 18 Aug 2025).

The stylistic framework was applied to the DreamBank corpus, including five series of single-author dream collections: blind ($P=\{p_1,\dots,p_K\}$01), ed ($P=\{p_1,\dots,p_K\}$02), izzy ($P=\{p_1,\dots,p_K\}$03), merri ($P=\{p_1,\dots,p_K\}$04), and viet ($P=\{p_1,\dots,p_K\}$05). The norm baseline consisted of 10 randomly sampled dreams from each series, totaling 720. Extraction was validated on 50 gold-standard clauses from textbooks, and the LLM agreed 100% with reference labels. Substrings of size $P=\{p_1,\dots,p_K\}$06 were extracted. In the “viet” series, quantitative results include $P=\{p_1,\dots,p_K\}$07, $P=\{p_1,\dots,p_K\}$08, $P=\{p_1,\dots,p_K\}$09, and $P=\{p_1,\dots,p_K\}$10. Clustering yields two patterns, one highly action-dominant and one action-state mix, and the silhouette score is highest when $P=\{p_1,\dots,p_K\}$11 substrings and two clusters are used for “viet” (Cortal et al., 9 Oct 2025).

The MDL serial-episode framework was tested on synthetic data, addresses from US presidential speeches, JMLR abstracts, and Moby Dick. On independent synthetic data with no structure, both methods return no non-singleton episodes. On planted-pattern synthetic datasets, both recover nearly all planted episodes, with only a few fragments due to overlap noise. In JMLR abstracts, the top MDL-most-important patterns discovered by Sqs-Search include “support vector machine” with $P=\{p_1,\dots,p_K\}$12 bits, “machine learning” with 646 bits, “state [of the] art” with 480 bits, and “cross validation” with 279 bits (Tatti et al., 2019).

The network framework is illustrated with Les Misérables. Stages S1-S3 show rapid $P=\{p_1,\dots,p_K\}$13 growth associated with exposition, S4 shows flat growth associated with digressive subplots, and S5 shows $P=\{p_1,\dots,p_K\}$14 growth with fixed $P=\{p_1,\dots,p_K\}$15, associated with convergence in the climactic barricade sequence. The young revolutionaries exhibit sentiment cycles from optimistic to struggle to tragic climax to resolution. For two Marius-Valjean interactions, the correlation of their topical states rises from $P=\{p_1,\dots,p_K\}$16 to $P=\{p_1,\dots,p_K\}$17 after the first meeting and to $P=\{p_1,\dots,p_K\}$18 after the second meeting, accompanied by topic transfer and emergence of shared exogenous topics (Min et al., 2016).

5. Interpretation, misconceptions, and methodological tensions

A common simplification is to treat narrative pattern mining as a variant of keyword extraction. The event-centric generalized category discovery results directly counter this view: EC-GCD is characterized by long, complex narratives that are 14 times longer than BANKING or CLINC, by divergent clustering and classification groupings, and by class imbalance as severe as 100-fold between dominant and rare scam types (Luo et al., 29 May 2025). In this setting, human labels may hinge on finer context than surface or temporal similarity, so narrative patterns must encode annotation-relevant event structure rather than lexical co-occurrence alone.

Another misconception is that narrative patterns are inherently contiguous or sentence-local. The literature contains both contiguous and non-contiguous formalisms. Personal-narrative style patterns are contiguous substrings by definition, whereas serial episodes explicitly allow gaps between matched events (Cortal et al., 9 Oct 2025, Tatti et al., 2019). This suggests that adjacency is a modeling choice, not a defining property of the field.

A further tension concerns interpretability versus evaluation. The biomedical system produces human-readable graphs with provenance, and domain experts valued the resulting exploratory overviews, but the evaluation reported no quantitative precision or recall (Kroll et al., 18 Aug 2025). Conversely, the stylistic framework uses exact counts, odds ratios, corrected significance testing, and clustering, yet its psychological claims remain correlational, and the authors state that clinical validation is needed before diagnostic use (Cortal et al., 9 Oct 2025). Narrative pattern mining therefore spans both exploratory interfaces and inferential pipelines, with different standards of evidence.

There is also a tension between frequency and informativeness. The MDL work explicitly begins from the observation that standard frequent pattern miners suffer from pattern explosion and return large numbers of highly redundant patterns. Its response is to evaluate results as a whole, using encoded length as a global quality score (Tatti et al., 2019). PaMA addresses a related issue under class imbalance by ranking clusters, filtering out samples already covered by earlier patterns, and refining generated patterns against labeled positives and negatives so that cluster boundaries better match annotation conventions (Luo et al., 29 May 2025). In both cases, mining is organized around non-redundant summarization rather than raw recurrence counts.

The network framework introduces a final methodological tension: whether narrative structure is best modeled as event order or as social and thematic dynamics. By moving beyond simple occurrence-based character co-appearance and incorporating sentiment analysis and topic modeling, it treats narratives as dynamically unfolding systems whose structure is visible in growth patterns, interaction polarity, and topic exchange (Min et al., 2016). A plausible implication is that “narrative pattern” is best understood as a family resemblance term spanning several levels of abstraction.

6. Limitations and prospective developments

Across the cited work, limitations arise from extraction quality, representational scope, and domain dependence. In the biomedical setting, nearly-unsupervised extraction methods yield moderate precision and recall, motivating future integration of supervised or semi-supervised relation extractors or specialized biomedical LLMs. Proposed interface and functionality extensions include clearer search-bar design, typo tolerance, better ranking of suggestions, inline provenance instead of context-losing tab switches, exclusion queries, query variables, bulk operations, export or save functions, incorporation of additional sources such as clinical guidelines, news feeds, or patent literature, LLM-based recommendation of candidate patterns, automatic summarization of pattern subgraphs, quantitative evaluation against a gold standard, and user studies on whether pattern mining accelerates hypothesis generation (Kroll et al., 18 Aug 2025).

In event-centric generalized category discovery, future work may explore multimodal event descriptions such as video plus text, meta-learning prompt variations for few-shot pattern induction in truly zero-shot novel classes, jointly learning the ranking criterion $P=\{p_1,\dots,p_K\}$19, or adapting $P=\{p_1,\dots,p_K\}$20 when $P=\{p_1,\dots,p_K\}$21 is unknown (Luo et al., 29 May 2025). These directions follow directly from the current reliance on prompt-guided LLM extraction, a fixed number of total categories, and handcrafted ranking weights.

In personal narrative style analysis, the current framework focuses on process types only. The authors note that richer analyses could incorporate participants, circumstances, tense, and aspect, but this would inflate the alphabet and combinatorial complexity. Mining only contiguous substrings may miss non-contiguous subsequence patterns, and application to other personal-narrative domains would require redefining the norm baseline and possibly retuning prompts and feature sets (Cortal et al., 9 Oct 2025).

In MDL-based sequence mining, modeling only serial episodes omits partial orders and simultaneous events, the current cover forbids overlapping windows, encoding design choices affect which patterns are preferred, preprocessing quality strongly affects pattern reuse, and multimedia narratives would require new code-table designs (Tatti et al., 2019). In network analysis, the framework itself is presented as an advance beyond simple occurrence-based representations, and its broader significance is the prospect of a more systematic modeling and understanding of narratives for social interactions, expression of human sentiments, and communication (Min et al., 2016).

Taken together, these limitations indicate that narrative pattern mining remains a methodological umbrella rather than a settled paradigm. The current literature supports at least five stable directions: latent event-schema mining, graph-based exploratory discovery, symbolic stylistic sequence analysis, MDL-based pattern set mining, and dynamic network-text modeling. The coexistence of these directions suggests that future consolidation, if it occurs, will likely proceed through shared evaluation problems and cross-representation alignment rather than through a single universal formalism.