How can PMI-weighted narrative networks reveal peripheral story structures?

Rare entities hide in the long tail of a narrative

Narrative networks model a text as a graph in which nodes are entities and edges are relations between them. The paper notes that this entity-centric approach has proven useful for foregrounding central structures: salient entities and their connections at a macro level, typically through frequency-based or centrality-based measures.

That strength is also a blind spot. According to Zipf's law, the distribution of words or entities extracted from a corpus has a long tail of rare ones. Rare entities are easily overshadowed by more salient entities, and they are also lost among each other in the long tail of many but rare entities, alongside noise from extraction errors. The paper frames the problem not as finding needles in a haystack but as recovering how the straws of the haystack fit together.

The target of the method is therefore the everything else: brief sub-plots, small clusters of descriptions, and associations between minor characters that go under the radar at the macro level.

PMI inflation becomes the signal instead of the noise

Pointwise mutual information has a known tendency to inflate for rare events. Conventional analyses treat that inflation as a weakness because maximum-likelihood probability estimates are unreliable at low frequencies. This paper turns the weakness into a strength by weighting edges with PMI so that the resulting network favors rarities.

The pipeline is deliberately simple: build a co-occurrence graph in which an edge between two entities indicates that they co-occur within a limited text span, weight those edges with PMI, and extract communities. Coreference resolution is included as the linguistically principled default. The analytical focus is on PMI weighting and community detection rather than on sophisticated entity extraction.

Communities extracted from the resulting network are treated as structural traces of underlying narrative elements. The paper describes them as molecules of a larger narrative structure that contribute texture and specificity to the whole.

K-clique and Louvain differ mainly in overlap

The authors compared two community detection algorithms using hyperparameters chosen to yield comparable community sizes: k = 4 and w = 2 for k-clique percolation, and resolution = 35 for Louvain. At comparable sizes, k-clique produced 1193 communities with a mean size of 7.22, while Louvain produced 797 communities with a mean size of 7.0.

The gap is expected because k-clique allows overlapping communities, meaning the same entities can appear in multiple communities. On average, an entity belongs to 1.9 k-clique communities. The authors note that k-clique often yields at least one dominant community considerably bigger than the rest, following Palla et al. (2005); they observed these and discarded them from the analysis.

Manual inspection found that both methods produce communities corresponding to recognizable narrative elements, but both also tend to produce fragments of a broader narrative configuration. The authors adopt k-clique percolation as the primary method because its allowance of overlap better reflects the narrative reality that characters and locations participate in more than one configuration, while noting that the reported findings generally hold for Louvain communities as well.

Three quarters of communities are single-spike episodic

Running k-clique percolation on the PMI-weighted narrative network of The Lord of the Rings yields 1193 communities. Roughly three quarters are single-spike communities, activated in only one section, and therefore fall under the episodic type. Of the remaining multi-spike quarter, roughly a quarter are episodic with multiple spikes in close proximity, roughly half are echoing, and the last quarter are recurring.

Episodic structures are concentrated activations within a limited span. They are generally idiosyncratic in terms of entities, which the paper expects from an information-theoretic perspective: rare entities co-occur with few others, and small clusters in proximity surface as communities with high internal PMI. What they correspond to varies, including short descriptions, events, songs, or fragments of dialogue.

A subset of episodic structures is sustained across multiple consecutive sections, such as concentrated sub-plots or scenic descriptions. These are few but distinct, and the more sustained they are, the less they are favored by PMI, because their entities will be more frequent.

Echoing structures link passages hundreds of pages apart

Echoing structures, roughly a quarter of the multi-spike communities, form a distinct class in which an idiosyncratic configuration reappears at a distant point in the text: a poem or song revisited, an event or quote repeated or referenced, or a specific lore reference between two or few entities.

The paper argues these long-distance connections are particularly interesting in a hybrid close-distant reading setting, where the computational method surfaces connections a human reader may struggle to pick up across hundreds of pages. One example is a vision by Sam in the Mirror of Galadriel in book one that is realized by the end of book two, where the community a winding stair, a dark veil, pale faces, a great dark cliff appears in both contexts.

The typology table lists further echoing examples, including distinct versions of a song sung by Frodo in the same geographical location early in book one and late in book three, and an echoing reference to The Smials in Tuckborough, ancestral home of the Tooks.

Recurring threads span the work and vary internally

Recurring structures span much of the work and show internal variation. A small tail of many-spiked communities tends to be diluted by generic nouns, such as pairs like left side and right side that co-occur frequently but are questionable as meaningful narrative elements, or by aggregated connections between entities that do not form a meaningful cluster.

Among the meaningful communities, the paper identifies three patterns. Some function as lore threads reflecting recurring geographic or world-building references, such as descriptions of and references to fiefdoms and landmarks in the land of Gondor. Others track recurring sub-plots across distant parts of the narrative, such as the legend of the dead men of Dunharrow. A third pattern emerges where a later passage gathers threads dispersed across earlier sections into a retrospective account, such as Gandalf recollecting events at the Council of Elrond that tie back to short passages in the preceding half of book one.

The paper notes that the bulk of single-spike communities might be questioned from a macro perspective, but their sheer prevalence may hint at a structural property of narrative: stories constantly introduce and resolve small elements alongside their central progression.

Sensitivity limits the method to exploratory use

The method deliberately leverages the instability of PMI at low frequencies, which follows from the unreliability of maximum-likelihood probability estimates for rare events. Extracted communities depend on PMI values, which depend on marginal and joint probability estimates, which depend on entity counts. Operating in low-frequency territory, even small differences in individual entity counts propagate through each step.

This makes the method highly susceptible to extraction errors: at the level of rare entities, a single false positive or negative can distort PMI associations drastically, unlike macro-level patterns where signal strength masks such errors. Changes to count input, such as coreference resolution or different entity resolution strategies, affect PMI values considerably. The method is also sensitive to hyperparameter choices; k and the weight threshold for k-clique percolation can shift individual community assignments, though the authors observed that some communities are more stable under such variation than others.

The paper concludes that this inescapable sensitivity means the method is best understood as a bottom-up lens for surfacing structure in the long tail of rare entities, rather than a robust pipeline for extracting predefined latent structures. The authors also note that no benchmark exists for the task and that evaluation is qualitative: they examine whether extracted communities correspond to recognizable structures in the source text.

Treat communities as candidates, not conclusions

The paper positions the approach as a complement to existing methods that focus on salient entities, offering a means of exploring structure in the long tail rather than replacing established techniques. The extracted communities are candidates, often even fragmentary ones, not conclusions. They surface specific entity configurations that trace back to passages which can then be examined.

For readers deciding how to use this work, the most decision-relevant split is between the single-spike episodic bulk and the multi-spike minority. The single-spike communities are the most numerous but the least individually informative, and their value lies in their prevalence as a hint about how narratives constantly introduce and resolve small elements. The multi-spike communities, especially echoing and recurring ones, are where the method surfaces weak-signal connections across long textual distances that salience-based methods would miss.

The authors suggest that other association measures such as log-likelihood, which is more robust at low frequencies, may complement or improve the approach, and that applying it to more minimalist works with fewer characters or sparser descriptions may yield different results. Whether such differences reflect fewer peripheral structures or a different relationship between peripheral and central elements remains open.

Sources used in this answer