Landscape Reading Model and Its Text Network Representation

The Landscape Model proposed by van den Broek and colleagues in 1999 [8] proposes a following representation of how we perceive text: when we read, concepts rise and fall in activation like a terrain of shifting peaks and valleys. Based on our limited attention span, we hold the concepts that we're immediately perceiving in our short-term memory and connect them to our knowledge so the resulting landscape of activation will be affected both by the text and by our pre-existing knowledge. That's why we use a similar approach to represent connections in text using InfraNodus. Using a sliding 4-gram window, a graph of a text is built that approximates how we perceive information and also shows how the concepts will form patterns over time in the context of a particular document. 

However, what the landscape reading model misses is the that once the landscape emerges, there will also be gaps in it. Or that there will be peaks where we'd expect valleys. Especially if a particular text is highly original. This may lead to creative tension that can be highly productive for generating ideas. That's why network representation, while being very useful to represent how we perceive text, can also be very useful for helping us see how to develop it further: because it lets us see those gaps (or clusters, where they should not exist) and generates a creative tension that can be used for ideation.

Here's how it works.

What is a Landscape Reading Model?

The Landscape Model emerged from a productive tension between two irreconcilable accounts of reading. 

The memory-based view held that comprehension is automatic and effortless — concepts activate passively through pre-existing associations, and the reader simply rides the wave [4]. The constructionist view insisted on the opposite: readers strategically and consciously retrieve prior text and background knowledge to build coherence, and this costs effort [1]. Both accounts were partially right, and each collapsed precisely where the other was strong. If reading is purely passive, how do you recover when you lose the thread? If reading is purely effortful, how does anyone get lost in a novel for three hours?

Van den Broek's resolution was not a compromise but a structural integration. Both processes operate, but on different triggers. Passive cohort activation fires continuously — when a concept enters the reading cycle, its semantic neighbors light up automatically. Strategic coherence-based retrieval fires conditionally — only when the reader's standards of coherence are not met by the passive process alone [8]. The reader's purpose, domain knowledge, and available cognitive resources modulate the threshold. The result is a dynamic landscape of fluctuating activations across reading cycles [7], and this landscape — accumulated over the entire text — becomes the memory representation.

The metaphor of a landscape is really interesting. But what if the landscape could be made visible?

Why four? — The cognitive basis of the sliding window

Before mapping the Landscape Model onto text network analysis, it is worth asking a precise question: why should a co-occurrence window of four words capture anything cognitively meaningful? 

It has to do with our working memory.

The dominant framework for working memory since the 1970s has been Baddeley and Hitch's multi-component model [15], which divides the system into specialized subsystems: 

• a phonological loop for verbal rehearsal, 
• a visuospatial sketchpad for spatial information, 
• a central executive for attentional control, and — added later — 
• an episodic buffer for integrating information across domains [16].

This architecture explains how working memory operates but is less precise about how much it can hold at once. Miller's classic estimate of "seven plus or minus two" items [17] long served as the capacity benchmark, but it conflated raw storage with the effects of rehearsal and chunking strategies.

Nelson Cowan's influential revision of Miller's classic "magical number 7" proposed that the focus of attention — the part of working memory where items are simultaneously available for processing — holds approximately four independent chunks in the average adult [2].  Cowan's revision cuts through this conflation. By measuring capacity under conditions where rehearsal and grouping are experimentally prevented, Cowan [2] showed that the true limit of the focus of attention — the part of working memory where items are simultaneously available for association and comparison — is closer to four independent chunks. T

his is not a revision of Baddeley's architecture but a specification within it: the central executive and episodic buffer coordinate the system, but the scope of what can be held in concurrent awareness at any single moment is approximately four items [3]. For reading comprehension, this distinction matters. Baddeley's model explains that working memory supports reading through rehearsal, storage, and executive control. Cowan's constraint specifies the bottleneck — the maximum number of concepts a reader can hold simultaneously active, and therefore the maximum scope of a single reading cycle's co-activation window.

What makes this relevant to reading is that these four chunks define the scope within which new associations can form. Elements that are simultaneously present in the focus of attention become linked — they form new structures that are then consolidated into long-term memory [11]. This is the associative engine of comprehension: meaning is built not from individual concepts in isolation but from the co-presence of concepts within the same attentional window.

The Landscape Model's reading cycle operates on the same principle. At each cycle (roughly a clause), a set of concepts is simultaneously active, and concepts that are co-active strengthen their mutual connections [8]. But the Landscape Model defines its cycle by linguistic unit — clause or sentence — without specifying a capacity constraint on how many items can be co-active. Cowan's work supplies exactly that constraint: approximately four.

There is also a temporal dimension. Cowan and colleagues found that information held in sensory memory decays dramatically unless attended — memory for unattended spoken syllables showed sharp decline as the interval between presentation and recall cue increased from 1 to 10 seconds [12]. This ~10-second window defines the temporal horizon within which unattended information remains available for potential integration. Beyond it, what has not been attended is effectively lost.

These two constraints — four chunks in spatial scope, roughly ten seconds in temporal duration — converge on a cognitive bottleneck through which meaning must pass. The 4-gram co-occurrence window used in the text network analysis used in InfraNodus  [5] [6] is not an arbitrary parameter. It is a computational operationalization of the attentional focus through which readers construct meaning: approximately four items, simultaneously available, forming associative links through co-presence. The sliding window moves through the text the way the focus of attention moves through the reading process — capturing local co-activations and encoding them as structure.

Cowan makes the connection to reading comprehension explicit: in understanding an essay, a reader might need to hold simultaneously the major premise, the point from the previous paragraph, and a fact and an opinion from the current paragraph — and only when all of these have been integrated into a single chunk can the reader continue to read and understand [3]. This is precisely the process the sliding window captures and the resulting graph makes visible.

InfraNodus performs two distinct cognitive operations that helps reduce the complexity and addresses the natural limitation of our perception. 

First, it chunks the text's concepts into topical clusters, compressing the dimensionality from potentially hundreds of individual nodes to a handful of groups — bringing the structural overview closer to what the ~4-chunk attentional focus can hold. 

Second, and more importantly, it identifies the gaps between these clusters and proposes specific bridging concepts or questions. This second operation is the critical one: it transforms an absence (which requires holding two clusters plus the missing link simultaneously — at least three chunks devoted to representing something that doesn't exist) into a presence — a concrete node or question that the reader can attend to directly. The gap, which was cognitively invisible because its perception exceeded the bottleneck's capacity, becomes a single attendable item.

The sliding window as a reading cycle

Consider what happens when a text is processed through a network analysis algorithm that uses this co-occurrence window. The window moves through the text one word at a time. At each position, the words within the window are connected to each other. As the window slides forward, new connections form, old ones either strengthen through repetition or fade into the periphery.

This is structurally analogous to the Landscape Model's reading cycle. In van den Broek's framework, each cycle activates a set of concepts simultaneously, and concepts that are co-active within a cycle strengthen their mutual connections [8] [9]. The 4-gram window performs the same operation — it defines a local scope of co-activation and encodes it as edges in a graph [5] [6]. The accumulated graph after processing the entire text is, in effect, the frozen trace of all reading cycles combined.

The difference is one of visibility. The Landscape Model produces a theoretical activation matrix [10]. The text network — through InfraNodus — produces an observable graph.

This operation also resonates with how attention mechanisms work in transformer-based LLMs. A transformer has simultaneous access to all tokens in its context window, but the attention mechanism must still select — and many attention mechanisms learn a proximity bias, assigning higher weights to nearby tokens, much as cohort activation in the Landscape Model privileges local semantic neighbors. Other attention mechanisms specialize in long-range bridging, connecting distant parts of the sequence — a learned analogue of strategic coherence-based retrieval. 

InfraNodus performs a similar division of labor explicitly: clusters capture local co-occurrence (the cohort), while betweenness centrality and gap detection identify where long-range bridges exist or are missing. The critical difference is that the transformer builds its bridges implicitly and cannot represent their absence, while InfraNodus makes both the bridges and the gaps structurally visible.

Clusters as cohorts

One of the Landscape Model's key mechanisms is cohort activation: when a concept is activated, it passively triggers related concepts that belong to its associative neighborhood. These cohorts are not defined by the text alone — they are shaped by the reader's prior knowledge and the patterns of co-occurrence accumulated during reading [9].

In a text network, community detection algorithms (such as modularity-based clustering) identify groups of nodes that are densely interconnected — concepts that tend to co-occur frequently within the same local windows. These clusters are the topological equivalent of cohorts. They represent the semantic neighborhoods that the text itself constructs through repeated co-activation [5] [6]. Graph visualization helps emphasize those cohorts both using color-coding and Force-atlas layout algorithms, which can then be used both by humans for ideation and by LLMs.

The isomorphism is not trivial. The Landscape Model posits cohorts as a cognitive mechanism inferred from behavioral data. The text network makes them structurally explicit — visible as colored clusters in a graph, measurable in their density and boundary sharpness.

Betweenness centrality — not peaks, but bridges

Here is where the mapping becomes most interesting, and where the network approach goes beyond what the Landscape Model describes.

In the Landscape Model, concepts that maintain high activation across many reading cycles become central to the resulting memory representation [7] [8]. They are the peaks of the landscape — the ideas the reader is most likely to recall. This is essentially a frequency-based account of centrality: more activation cycles, stronger trace.

Betweenness centrality captures something different [5] [6]. A node with high betweenness is not necessarily the most frequently activated — it is the one that sits on the shortest paths between different clusters. It is a bridge, not a peak. It connects semantic territories that would otherwise remain separate.

This distinction matters. A text can have a concept that appears constantly but only within one cluster (high frequency, low betweenness). It can also have a concept that appears less frequently but serves as the sole connection between two major thematic regions (lower frequency, high betweenness). The Landscape Model would predict the first concept dominates the memory representation. The network analysis suggests the second concept is structurally more important — it is the crossroad where meaning from different parts of the text can flow and recombine.

The reader who grasps the bridging concepts understands the text's architecture. The reader who remembers only the peaks understands its surface.

The gap the Landscape Model cannot see

Van den Broek's model tracks what is activated [7]. It maps the presence and intensity of concept activation across reading cycles. What it does not — and structurally cannot — represent is what is absent.

In a text network, the gaps between clusters are not empty space. They are structural features: regions where connections could exist but do not [6]. These gaps represent unexplored relationships between thematic territories, questions the text implicitly raises but does not answer, conceptual bridges that the reader (or the author) might construct but has not yet.

This is perhaps the most significant extension. The Landscape Model gives you the terrain as it is. The network gives you the terrain and the spaces between the mountains — the valleys that might be crossed, the paths that might be built. Gap detection transforms reading from a process of absorption into a process of generative inquiry.

However, what if the real unit of comprehension is not the concept, nor the link, but the felt absence that reorganizes attention?

In that view, a gap is not a defect in the graph. It is a higher-order operator: a negative node that does not store meaning but redirects activation across clusters. Passive activation spreads through what is present; strategic retrieval begins when what is missing becomes cognitively tangible. 

Meaning does not emerge from connected concepts alone, but from the mind’s ability to stabilize an unconnected possibility long enough for a new bridge to form. This shifts the model from a graph of associations to a topology of latent absences: 

- clusters = what thought can already hold
- betweenness = what thought can traverse
- gaps = what thought is not yet structured enough to think

Attention binds presence; inquiry binds absence. Comprehension follows links; insight follows the shape of what links refuse to exist.

ADHD, Working Memory, and Broader Scope of Semantic Activation

One text, one landscape?

There is a productive tension worth acknowledging. The Landscape Model emphasizes that the same text produces different landscapes depending on the reader's standards of coherence, reading purpose, and domain knowledge [8] [9]. A scientist and a casual reader generate different activation patterns from the same paragraph. The landscape is reader-dependent.

A text network, as typically constructed, produces one graph per text [5]. It represents the text's inherent co-occurrence structure — something closer to the potential landscape that any reader might traverse, rather than the actual landscape any specific reader produces.

This could be seen as a limitation or as a feature. The single graph shows the structural affordances of the text itself — the connections and gaps available to any reader. Different analytical perspectives (e.g. starting the reading from clusters, gaps, or from the latent nodes) can then be used to simulate different reading stances. 

From temporal dynamics to spatial topology

The Landscape Model is fundamentally temporal. It tracks activation fluctuations cycle by cycle, and the memory representation emerges as a cumulative trace of these dynamics [7] [10]. The text network collapses this temporal dimension into a spatial topology — a graph that can be inspected, measured, and manipulated all at once.

Normally, the sequence of activation may lost in this transformation — the dynamic unfolding that makes reading a process rather than a product. However, InfraNodus makes it possible to retain the narrative through its dynamic trend activation feature that helps see evolution of a graph in time. Also, the network provides something that the temporal model cannot provide: a structural overview that reveals the architecture of meaning in its totality [5] [6]. The reader cannot easily see the landscape while traversing it. The graph lets you see the entire terrain from above — including the parts you have not yet walked through. 

The graph is like a map, the various interface features and interpretation hints are like the GPS that direct your attention through the landscape.

This is not a replacement for the Landscape Model. It is a different projection of the same underlying phenomenon. The Landscape Model captures reading as it unfolds in time. The text network captures the structure that reading, over time, builds.

There is a neural architecture that occupies precisely the intermediate position the article has so far left empty. 

The Long Short-Term Memory network (LSTM), introduced by Hochreiter and Schmidhuber in 1997 [18], processes text sequentially — one token at a time, maintaining a hidden state that carries information forward through the sequence. At each step, the LSTM faces a decision that maps almost exactly onto the Landscape Model's reading cycle: a forget gate determines which elements of the prior state to discard (analogous to activation decay between cycles), an input gate determines which new information to incorporate (analogous to the current cycle's text input), and an output gate determines what becomes available for the next step (analogous to carryover to the next reading cycle) [18]. 

The LSTM traverses the text. It builds its representation cumulatively, step by step, and what it "remembers" at any point is shaped by the gating decisions made at every prior step — a direct computational parallel to the Landscape Model's claim that the memory representation emerges from the dynamics of activation and deactivation as reading unfolds [7] [8]. 

The transformer architecture in LLMs, by contrast, abandoned this sequential architecture entirely. By computing attention over all tokens in parallel, it gained the ability to see the entire text at once — but lost the temporal dynamics that make the LSTM a plausible model of actual reading. The shift from LSTM to transformer mirrors, in compressed form, the shift from the Landscape Model to the text network: from sequential traversal to simultaneous topology, from process to structure. The LSTM reads. The transformer sees. The text network maps.

Beyond the bottleneck — variability as a navigation strategy

The solutions proposed so far — text network visualization, organizational bridge-detection, cross-functional teams — share a common logic: externalize the landscape so it can be perceived despite the bottleneck. They are prosthetics. They work. But they leave the bottleneck itself untouched.

There is another approach. Not expanding the window, but changing how the window moves.

The EightOS framework, developed as an adaptive movement practice, proposes that the key to navigating complexity is not holding more — it is cognitive variability [25]. 

The framework identifies a progression of states: 

• uniform variability (shared rhythm, common ground), 
• regular variability (patterned differences that produce engagement and controlled tension), 
• fractal variability (variation distributed across scales so that local patterns echo global structure), and 
• complex variability (novel combinations that exceed any single agent's prediction). 

The system cycles through these states, returning to uniform variability for regeneration before beginning again.

This is a navigation protocol for the ~4-chunk bottleneck. 
A reader — or a thinker, or a strategist — who stays in one cluster sees no gaps. 
One who jumps chaotically between clusters sees no coherence. 
But one who oscillates through the variability sequence can progressively build a representation of the full landscape without ever needing to hold it simultaneously. 
Uniform variability establishes a base — the reader settles into a cluster, builds local coherence. 
Regular variability introduces controlled disruption — a bridging concept pulls attention toward an adjacent cluster. 
Fractal variability is where the critical operation occurs: the reader recognizes the same structural pattern at a different scale or in a different cluster, and this recognition itself becomes the bridge. 
Complex variability allows genuinely novel connections to emerge. 
Then return, consolidate, and begin again.

The bottleneck remains. Four chunks is still four chunks. But the trajectory through the landscape compensates for the constraint.

This connects to a principle the EightOS framework calls the A-R-D sequence: assimilation, redirection, dissipation. When the reader — or the organization, or the body in movement — encounters a structural gap, the gap manifests first as tension: something does not fit, a transition fails, coherence breaks. 

The standard response is to force coherence (assimilate harder) or abandon the thread (dissipate prematurely). The A-R-D sequence proposes a third path: absorb what fits the current frame, redirect toward what does not integrate directly (approach the gap obliquely rather than head-on), and release the rigidity that prevents the network from reorganizing.

In reading terms: when a text resists comprehension, the reader who forces local coherence misses the gap entirely. The reader who abandons the thread loses the local structure. The reader who redirects — pauses, shifts to a different entry point, approaches the difficult passage from the perspective of a different cluster — allows the gap to become perceptible without exceeding the bottleneck's capacity. The gap does not need to be held in working memory. It needs to be circled, approached from multiple directions, until its shape becomes inferable from the surrounding terrain.

This is what fractal variability offers that linear reading cannot: the recognition of structural isomorphisms across scales. When the same dynamic appears in the Landscape Model's reading cycles, in the text network's co-occurrence topology, in the transformer's attention patterns, and in organizational decision-making, the reader does not need to hold all four domains in the focus of attention simultaneously. They need only recognize the recurring pattern — co-activation within a limited window producing cumulative structure — and let the pattern itself serve as the bridge between domains. The cross-domain echo is the bridge. The isomorphism replaces the need for simultaneous representation.

This article has been performing exactly this operation. The 4-chunk bottleneck, described by Cowan [2] [3], was first recognized in the Landscape Model's reading cycle [8]. Then the same structure was identified in the text network's sliding window [5] [6]. Then in the transformer's sparse attention [18]. Then in organizational strategy [24]. At no point did the argument require holding all four domains simultaneously. It required recognizing the same pattern moving through different substrates — and letting each recognition deepen the understanding of the ones that came before.

The text network makes the landscape visible. The organizational protocol makes the gaps actionable. But the practice of cycling through variability — moving between scales, between clusters, between states of tension and release — makes the mind itself a more capable navigator of the landscapes it cannot fully see.

The bottleneck does not go away. But the landscape, approached fractally, reveals more of itself than any single view could contain.

References

[1] Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101(3), 371–395.

[2] Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–185.

[3] Cowan, N. (2010). The magical mystery four: How is working memory capacity limited, and why? Current Directions in Psychological Science, 19(1), 51–57.

[4] Myers, J. L., & O'Brien, E. J. (1998). Accessing the discourse representation during reading. Discourse Processes, 26(2–3), 131–157.

[5] Paranyushkin, D. (2011). Identifying the pathways for meaning circulation using text network analysis. Nodus Labs, Published online at noduslabs.com.

[6] Paranyushkin, D. (2019). InfraNodus: Generating insight using text network analysis. Proceedings of the World Wide Web Conference (WWW '19), 3584–3589. ACM.

[7] van den Broek, P. (1995). A 'landscape' model of reading comprehension: Inferential processes and the construction of a stable memory representation. Canadian Psychology, 36(2a), 53.

[8] van den Broek, P., Young, M., Tzeng, Y., & Linderholm, T. (1999). The landscape model of reading. In H. van Oostendorp & S. R. Goldman (Eds.), The construction of mental representations during reading (pp. 71–98). Erlbaum.

[9] Yeari, M., & van den Broek, P. (2011). A cognitive account of discourse understanding and discourse interpretation: The Landscape Model of reading. Discourse Studies, 13(5), 635–643.

[10] Yeari, M., & van den Broek, P. (2016). A computational modeling of semantic knowledge in reading comprehension: Integrating the landscape model with latent semantic analysis. Behavior Research Methods, 48(3), 880–896.

[11] Cowan, N. (2005). Working-memory capacity limits in a theoretical context. In C. Izawa & N. Ohta (Eds.), Human learning and memory: Advances in theory and applications. The 4th Tsukuba International Conference on Memory. Erlbaum.

[12] Cowan, N., Lichty, W., & Grove, T. R. (1990). Properties of memory for unattended spoken syllables. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(2), 258–269.

[13] Linderholm, T., Virtue, S., Tzeng, Y., & van den Broek, P. W. (2004). Fluctuations in the availability of information during reading: Capturing cognitive processes using the landscape model. Discourse Processes, 37(2), 165–186.

[14] Singer, M., Graesser, A. C., & Trabasso, T. (1994). Minimal or global inference during reading. Journal of Memory and Language, 33(4), 421–441.

[15] Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. H. Bower (Ed.), The Psychology of Learning and Motivation (Vol. 8, pp. 47–89). Academic Press.

[16] Baddeley, A. (2000). The episodic buffer: A new component of working memory? Trends in Cognitive Sciences, 4(11), 417–423.

[17] Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2), 81–97.

[18] Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780.

[19] Martinussen, R., Hayden, J., Hogg-Johnson, S., & Tannock, R. (2005). A meta-analysis of working memory impairments in children with attention-deficit/hyperactivity disorder. Journal of the American Academy of Child & Adolescent Psychiatry, 44(4), 377–384.

[20] Kasper, L. J., Alderson, R. M., & Hudec, K. L. (2012). Moderators of working memory deficits in children with attention-deficit/hyperactivity disorder (ADHD): A meta-analytic review. Clinical Psychology Review, 32(7), 605–617.

[21] White, H. A., & Shah, P. (2016). Scope of semantic activation and innovative thinking in college students with ADHD. Creativity Research Journal, 28(3), 275–282.

[22] Hoogman, M., Stolte, M., Baas, M., & Kroesbergen, E. (2020). Creativity and ADHD: A review of behavioral studies, the effect of psychostimulants and neural underpinnings. Neuroscience & Biobehavioral Reviews, 119, 66–85.

[23] Martinussen, R., Hayden, J., Hogg-Johnson, S., & Tannock, R. (2005). A meta-analysis of working memory impairments in children with attention-deficit/hyperactivity disorder. Journal of the American Academy of Child & Adolescent Psychiatry, 44(4), 377–384.

[24] Dalio, R. (2017). Principles: Life and Work. Simon & Schuster.

[25] Paranyushkin, D. (2016). EightOS: Bodily Adaptive Intelligence. Nodus Labs. Published online at 8os.io