Latent Structure Narrative Model (LSNM)

• LSNM is a family of probabilistic, neural, and graph-based models that uncover hidden narrative structures such as sentiment arcs, actant relationships, and turning points.
• It employs methods like Switching Linear Dynamical Systems, graph-based inference, and attention-driven soft labeling to model narrative dynamics in texts.
• Empirical evaluations show that LSNMs lead to improved narrative coherence, control in generation, and more complete, interpretable summaries compared to traditional approaches.

The Latent Structure Narrative Model (LSNM) refers to a family of probabilistic, neural, or graph-based models that aim to infer, represent, and utilize the latent scaffolding underlying narrative texts. LSNMs make explicit the hidden structure of stories—such as sentiment arcs, actant relationships, or pivotal turning points—and use such structures to improve narrative generation, extraction, or analysis. Across recent literature, LSNM implementations include switching linear dynamical systems with explicit discrete control, graph-based actant-relationship networks, and latent-event extractive models for summarization. This overview synthesizes the primary LSNM variants and their methodologies, with emphasis on recent models for generation (Weber et al., 2020), relational inference (Tangherlini et al., 2020), and summarization (Papalampidi et al., 2020).

1. Formal Modeling Frameworks

LSNMs formalize narrative structure using different mathematical and algorithmic primitives, matching the domain and purpose.

1.1. Switching Linear Dynamical Systems (SLDS)

In generative narrative modeling, LSNM is instantiated as a Switching Linear Dynamical System. This model integrates a discrete sequence of scaffold variables $s_t$ (e.g., sentiment states) and continuous narrative content variables $x_t \in \mathbb{R}^d$ ("state vectors"). The generative process involves:

• Discrete scaffold chain: $s_t \in \{1,\dots,K\}$ (e.g., $K=3$ for sentiment).
• Latent dynamics: Each $s_t$ selects a bank of $K$ linear Gaussian transitions $$x_{t+1} = A_{s_{t+1}} x_t + b_{s_{t+1}} + \varepsilon,\; \varepsilon \sim \mathcal{N}(0,Q_{s_{t+1}})$$.
• Markovian scaffold prior: $p(s_{t+1} \mid s_t) = \Pi_{s_t, s_{t+1}}$.
• Conditional emission: Sentences are produced via a recurrent neural network LLM modulated by $x_t$ and $s_t$:

$x_t \in \mathbb{R}^d$0

1.2. Latent Narrative Graph Model

In narrative framework discovery, LSNM denotes a latent actant-relationship network within contexts $x_t \in \mathbb{R}^d$1. Each $x_t \in \mathbb{R}^d$2 is a context-specific directed weighted multigraph among actants ("supernodes" and "subnodes") and their semantic relations. Story fragments (e.g., posts) are modeled as samples from these latent subgraphs, where participation probabilities and relationship frequencies are estimated empirically.

1.3. Extractive Summarization via Latent Turnpoints

For long-form narratives, LSNM is implemented as a latent "turning-point" assignment model over document scenes. Each scene $x_t \in \mathbb{R}^d$3 is given a softmax-based probability $x_t \in \mathbb{R}^d$4 of realizing one of $x_t \in \mathbb{R}^d$5 canonical turning points (e.g., Climax, Setback). These latent assignments are then regularized and combined with content features for extractive summarization.

2. Inference and Learning Algorithms

Inference in LSNMs involves recovering latent structure from ambiguous or incomplete observations. The technical approaches vary by model type.

2.1. Variational Inference and Semi-supervised Learning

In SLDS-based LSNMs, inference is amortized through a variational encoder, yielding tractable approximations:

$x_t \in \mathbb{R}^d$6

The unsupervised objective is a VAE-style ELBO with KL terms for latent states and switch variables, and (if semi-supervised) a cross-entropy term for any available switching labels. Gumbel-Softmax and reparameterization allow end-to-end gradient optimization (Weber et al., 2020).

2.2. Gibbs Sampling for Fill-in and Constrained Generation

Post-training, SLDS-LSNMs support constrained (non-ancestral) generation via Gibbs sampling. Missing latent states and observations are resampled alternately, using the product-of-Gaussians posterior for $x_t \in \mathbb{R}^d$7 and greedy decoding for missing $x_t \in \mathbb{R}^d$8.

2.3. Pipeline and Clustering for Latent Network Models

In the relational LSNM variant, the pipeline consists of:

• Triplet extraction via syntactic parsing and SRL (Semantic Role Labeling).
• Entity clustering to create actant supernodes/subnodes (BERT embeddings, $x_t \in \mathbb{R}^d$9-means, frequency heuristics).
• Edge weighting and filtering by KL-divergence scores between actant-pair-specific and global verb distributions. No EM or variational inference is performed; the approach is unsupervised and deterministically data-driven (Tangherlini et al., 2020).

2.4. Attention-based Soft Labeling for Summarization

For the extractive screenplay LSNM, turning-point assignments are computed as latent softmax attentions. Regularization encourages mutual exclusivity (KL orthogonality) and temporal localization (priors anchored by summary theory). Parameters are optimized by gradient descent either jointly (supervised) or with pretraining plus test-time TextRank-style centrality (unsupervised) (Papalampidi et al., 2020).

3. Controlling and Exploiting Latent Narrative Structure

An essential function of LSNMs is to allow controlled manipulation and analysis of narratives:

• In SLDS-LSNMs, user-specified scaffolds (e.g., sentiment trajectories) can be imposed to generate stories with desired narrative arcs.
• In fill-in tasks, clamping partial observations enables interpolation or story completion by sampling coherent latent states and emissions.
• In relational LSNMs, actant domains and relationship macro-structure (community detection) elucidate patterns such as cross-domain linkage (e.g., conspiracy theory stabilization via "hidden knowledge").
• In summarization LSNMs, interpretability is achieved via inspection of which scenes correspond to canonical narrative events, improving completeness and diversity of summaries.

4. Empirical Evaluation and Comparative Results

Comprehensive comparisons demonstrate the practical benefits of LSNM strategies.

LSNM Type	Application Domain	Key Empirical Findings
SLDS-based (Weber et al., 2020)	Story generation, fill-in	Higher ROUGE/METEOR, lower perplexity, better coherence and control vs. Seq2Seq, RNN-LM baselines
Graph-based (Tangherlini et al., 2020)	Conspiracy framework mapping	83–84% relation recall to human gold, 100% supernode coverage (high-freq), interpretable domain structure
Turning-point (Papalampidi et al., 2020)	Screenplay summarization	+3–4 absolute F1 over strong baselines, improved inclusion of key aspects, confirmed by human evaluation

Significance: LSNMs yield both higher quantitative performance (e.g., fill-in, classification, summarization) and qualitatively more accurate or interpretable representations of underlying narrative dynamics.

5. Interpretation, Theoretical Insights, and Limitations

LSNMs provide interpretable latent semantics—scaffolds, events, or relationship networks—but differ in formalism and methodological rigor.

• The SLDS model (Weber et al., 2020) is a fully probabilistic graphical model; inference supports both maximum-likelihood and Bayesian variants.
• The graph-based pipeline (Tangherlini et al., 2020) encodes the generative intuition mathematically (joint probabilities of context, actant sets, relations) but currently lacks explicit priors or a complete EM/VI routine.
• The turning-point LSNM (Papalampidi et al., 2020) integrates latent structure within modern neural architectures, leveraging auxiliary regularization for mutual exclusivity and temporal localization.

A plausible implication is that while LSNMs increase interpretability and control, formal Bayesian treatment remains lacking in some variants, particularly in relational extraction pipelines.

6. Empirical Insights on Narrative Structure and Stability

Findings across domains reveal that:

• Conspiracy theory narratives (Pizzagate) rapidly align multiple otherwise-unconnected domains using keystone events, stabilizing early and resisting pruning.
• Actual conspiracies (Bridgegate) evolve more slowly, with network structure accreting as investigations unfold; the resultant actant graph is robust to node/edge removal (Tangherlini et al., 2020).
• In screenplay and story summarization, explicit modeling of latent narrative events (turn-points) vastly improves both automatic and human evaluations of summary completeness and structure (Papalampidi et al., 2020).

This suggests that LSNMs offer a unified analytical toolset for investigating both the dynamics of narrative creation and the mechanisms by which stories are constructed, propagated, and summarized.

7. Theoretical Directions and Model Extensions

If one wishes to extend LSNMs into a truly parametric latent-variable family, Dirichlet or other priors could be placed on contexts, participation probabilities, and relationships, enabling classical EM or variational-Bayes inference—a direction outside the scope of currently published models but promising for future work (Tangherlini et al., 2020). This would further align LSNMs with established probabilistic graphical modeling traditions, potentially enhancing robustness and interpretability across narrative domains.