Chapter 10 · Text and Situation

A Network of Propositions and Text

When text comes in, working memory turns sentences into propositions. A proposition is one slice of meaning that binds objects in one relation, such as "Minsu left his umbrella behind." One composite representation weaving "Minsu," "umbrella," and "leave behind" in one relation is one proposition.

Propositions connect to one another where they point again to the same object. If an earlier sentence says "umbrella" and a later one follows with "that umbrella was red," the two propositions are connected through the shared "umbrella"—because when the node "umbrella" lights up, the two propositions that have it as a member wake together. The network of propositions woven by what is written in the text alone is called the text base. Since you need only follow the connections the text gives directly, no separate inference is required, and it completes cheaply, using almost none of working memory's resource.

Beyond the Text, to the Situation

But the text base is only "what the text said," not the situation the text points to. Read "Minsu left his umbrella behind. That afternoon he came back soaked," and we know what is not written between the two sentences: that it rained, and that, having no umbrella, he was caught in the rain. Neither "rain" nor "the reason he got wet" is written in the text. The reader filled that gap by inference, drawing on background knowledge in long-term memory—the knowledge that an umbrella blocks rain and that being caught in rain makes you wet.

The integrated representation that binds propositions to one another, and to background knowledge drawn from long-term memory, with relations made by inference, filling even the gaps the text left blank, is called the situation model. It is the situation the text means to depict, built as one model in the head. The situation model is the product of carrying out the synthesis of the inferential route across the whole text. If the text base is propositions written in the text linked shallowly and finished cheaply, the situation model is what is deeply synthesized, drawing in even background knowledge to fill the gaps. It takes that much more resource.

The Processing That Fills the Gap

We said the gap is filled by inference; let us follow, at one gap, exactly what operations happen and in what order within that filling.

Inside the text base, "left the umbrella behind" and "came back soaked" share no node but "Minsu," so the reason connecting the two is empty. This gap sends working memory to long-term memory. The surface nodes now lit, "umbrella" and "wet," become cues and bring up from long-term memory the representations that have them as members—representations like "an umbrella blocks rain" and "being caught in rain makes you wet." This retrieval raises into working memory the bridge node "rain" that will connect the two propositions.

When the bridge comes up, working memory binds it to both propositions in new relations: it rained, there was no umbrella to block it, and so he got wet. A causal edge is newly built and inserted at the gap, and the two propositions that lay apart are integrated into one chain. This retrieval and formation, repeated at every gap, yields the integrated representation in which even the bridges the text did not write are all laid—the situation model.

So what tells the text base and the situation model apart is processing. The text base links propositions only by the cues the text gives; the situation model retrieves from long-term memory at every gap and forms a new edge. That is, the situation model is the inference of binding by an already-known relation, repeated along the text. Because retrieval and formation are repeated, it takes more resource, and if there is no background knowledge to draw on, the bridge cannot come up and the gap goes unfilled. So a reader who does not know the relation between umbrella and rain stays at the text base reading the same text.

Can You Take One Step Further?

What a representation can do is set by which nodes and edges are in it. To answer a question is to follow activation from the representation's nodes and reach the answer; if there is no node to reach and no edge to follow, no answer comes.

The text base has only the nodes the text wrote and the edges the text connected directly. So there is no obstruction up to retracing what is written. "What did Minsu leave behind" comes straight out by following the edge from the "Minsu" node to "umbrella." A question that asks what the text said has its answer on the path the text laid. But that is exactly the limit. Ask "why did Minsu get wet" and there is no answer within the text base. Neither the node "rain" nor the causal edge connecting having-no-umbrella and getting-wet was in the text, so neither entered the text base, and the path for activation to follow is cut. The text base's limit is exactly here: it retraces all the text wrote, but with what the text did not write it cannot take one step.

The situation model makes exactly that one step possible. Because the "rain" drawn in while filling the gap and the newly built causal edge are inside the model, the path to answer "why did he get wet" is laid. Moreover, those edges are lodged even into long-term memory's background knowledge, so they connect, through the same background nodes, even to new situations not in the text. So one who has built a situation model infers what the text did not say and applies what was learned to a new situation. Inference and application are, in the end, sending activation to nodes beyond the text, and the situation model is what laid the path out beyond.

So the usefulness of the two products divides by the question asked. "What did it say" is answered by the text base; "why is it so" and "then what about this case" demand the situation model. That one who has memorized the wording whole nonetheless stalls before a question that takes one step further is not because what was memorized is short, but because the path out beyond the text was not laid.

The Test That Tells the Two Apart

This difference in usefulness shows straight away in measurement. Even for the same person, a recall test that has you reproduce the wording shines only on the text base, while a test that has you apply or infer to a new situation shines on the situation model. Depending on which you give, one shows and the other is hidden.

Someone who has memorized an appliance manual whole yet, set before the machine, cannot operate it is a model case of this divide. Able to recite the manual's sentences as is, the text base stands firm. But they did not bind which button, in what order, of the actual machine those words correspond to, and build it into a situation. The situation model did not stand. Of this person—full marks on a recall test, zero when made to operate—one who saw only recall says "understood perfectly," and one who saw the operation says "did not understand at all."

What the Opening's McNamara experiment measured was exactly this divide. Those who studied the gapped text and those who studied the smooth text were similar on the recall test and divided only on the inference-and-application test. This is because filling the gaps oneself grew the situation model while the text base was similar on both sides. That state—the same surface, differing only in depth—was a state where the text base was the same but the situation model differed. To have understood is not the state of having memorized the wording but the state of having built a model of the situation the text depicts, and the two divide by what you measure with.