some of the early stuff and then go to the popular later ones. We will start with the work of Kintsch in
the 70s and 80s and then go on to later research which bases on this.
An interactive Model of Comprehension
This model was already developed in the 80s, it is the basis for many later models like the CI-
Model, or even the Immersed-Experiencer Framework. According to Kintsch and van Dijk (1978) text
comprehension proceeds in cycles. In every cycle a few propositions are processed, this number is
determined by the capacity of the Short-Term Memory, so 7 plus or minus 2. In every cycle the new
propositions are connected to existing ones, they therefore form a connected and hierarchical set.
Early Computational Model
This computational model from Miller and Kintsch tried to model earlier theroies of
comprehension, to make predictions according to these and compare them to behavioural studies and
experiments. It consisted of several modules. One was a chunking program: It's task is to read in one
word at the moment, identify if it is a proposition and decide whether to integrate it or not. This part of
the model was not done computationally. The next part in the input order was the Microstructure
Coherence Program (MCP). The MCP sorted the propositions and stored them in the Working Memory
Coherence Graph. The task of the the Working Memory Coherence Graph was then to decide which
propositions should be kept active during the next processing cycle. All propositions are stored in the
Long Term Memory Coherence Graph, this decided which propositions should be transferred back in
to the Working Memory or it can construct a whole new Working Memory Graph with a differen
114 | Cognitive Psychology and Neuroscience
Situation Models and Inferencing
superordinate node. The problem with this Computational Model was that it show a really low
performance. But still it led to further research which tried to overcome it's shortcomings.
Construction-Integration Model
Event-Indexing Model
The Event-Indexing Model was first proposed by Zwaan, Langston and Graesser (1995). It makes
claims about how the incoming information in comprehension is processed and how it is represented in
the long-term memory.
According to the Event-Indexing Model all incoming actions events are splitted into five indexes.
The five indexes are the same as the five situational dimensions, though Zwaan&Radvasnky(1998)
claim that there are possibly more dimensions. These might be found in future research. One basic
point of this model is the processing time of integrating new events into the current model. It is more
easier to integrate a new incoming event if it shares indexes with a previous event. The more
contiguous the new event is, the easier it is integrated into the new Situation Model. This prediction
made by Zwaan & Radvanksy (1998) is supported by some prior research (Zwaan, Magliano and
Graesser, 1995). The other important point of the Event-Indexing Model is the representation in long-
term memory. Zwaan & Radvasnky (1998) predict that this representation is a network of nodes, these
nodes encode the events. The nodes are linked with each other through situational links according to
the indexes they share. This connection does not only encode if two nodes share indexes but it also
encodes the number of shared indexes through its strength. This second point already hints what the
Event-Indexing Model lacks. There are several things which it does not include. For example it does not
encode the temporal order of the events nor the direction of the causal relationships. The biggest
disadvantage of the Event-Indexing Model is clearly that it treats the different dimensions as different
entities though they probably interact with each other.
Radvansky & Zwaan (1998) updated the Event-Indexing Model with some features. This new
model splits the processed information into three types. These three types are the situational
framework, the situational relations and the situational content. The situational framework grounds the
situation in space and time and it's construction is obligatory. If no information is given this framework
is probably build up by standard values retrieved from prior world knowledge or some empty variable
would be instantiated. he situational relations are based on the five situational dimensions. These are
analysed through the Event-Indexing Model. This kind of situational information includes not the basic
information, which is given in the situational framework, but the relationships between the different
entities or nodes in the network. In contrast to the situational framework the situational relations are not
obligatory. If their is no information given or their are no possible inferences between entities, then
there is simply no relationship there. There is also an index which addresses importance to the different
relations. This importance consists of the necessity of the information to understand the situation, the
easiness to inference it when it would not be mentioned and how easy the information can later be
remembered. Another distinction this theory makes is the one between functional and non-functional
relations (Carlson-Radvansky & Radvansky, 1996; Garrod & Sanford, 1989). Functional relations
describe the interaction between different entities whereas non-functional relations are the ones
between non-interacting entities. The situational content consists of the entities in the situation like
protagonists and objects and their properties. These are only integrated explicitly in the Situation
Model, like situational relations, if they are necessary for the understanding of the situation. None the
Chapter 11
less the central and most important entities and their properties are obligatory again. It is proposed that,
in order to keep the processing time low, non-essential information is only represented by something
like a pointer so that this information can be retrieved if necessary.
The Immersed Experiencer Framework
The Immersed Experiencer Framework (IEF) is based on prior processing framework models (see
above for a detailed list) but tries to include several other research findings too. For example it was
found that during comprehension brain regions are activated, which are very close or even overlap with
brain regions which are active during the perception or the action of the words meaning (Isenberg et al.,
2000; Martin & Chao, 2001; Pulvermüller, 1999, 2002). During comprehension there is also a visual
representation of shape and orientation of objects (Dahan & Tanenhaus, 2002; Stanfield & Zwaan,
2002; Zwaan, Stanfield, & Yaxley, 2002; Zwaan & Yaxley, in press a, b). Visual-spatial information
primes sentence processing (Boroditsky, 2000). These visual representations can interfer with the
comprehension (Fincher-Kiefer, 2001). Findings from (Glenberg, Meyer, & Lindem, 1987; Kaup &
Zwaan, in press; Morrow, Greenspan, & Bower, 1987; Horton & Rapp, in press; Trabasso & Suh,
1993; Zwaan, Madden, & Whitten, 2000) suggest that information which is part of the situation and the
text is more active in the reader's mind than information which is not included. The fourth research
finding is that people move their eyes and hand during comprehension in a consistent way with the
described the situation. (Glenberg & Kaschak, in press; Klatzky, Pellegrino, McCloskey, & Doherty,
1989; Spivey, Richardson, Tyler, & Young, 2000).
The main point of the Immersed Experiencer Framework is the idea that words active experiences
with their referents. For example "an eagle in the sky" activates a visual experience of a eagle with
stretched-out wings while "an eagle in the nest" activates a different visual experience. According to
Zwaan (2003) the IEF should be seen as an engine to make predictions about language comprehension.
These predictions are then suggested for further research.
According to the IEF the process of language comprehension consists of three components, these
are activation, construal and integration. Each component works at a different level. Activation works
at the world level, construal is responsible for the clause level while integration is active at the
discourse level. Though the IEF shares many points with earlier models of language comprehension it
differs in some main points. For example it suggests that language comprehension involves action and
perceptual representations and not amodal propositions (Zwaan, 2003).
Levels of Representation in Language and Text
Comprehension
A lot of theories try to explain the situation model or so called mental model in different
representations. Several theories of the representation deal with the comprehension from the text into
the situation model itself. How many levels are included or needed and how is the situation model
constructed, is it done by once like:
Sentence → Situation Model
Or are there levels in between which have to be passed until the model is constructed? Here are
116 | Cognitive Psychology and Neuroscience
Situation Models and Inferencing
three different representations shown which try to explain the construction of the situation model by a
text.
Propositional Representation
The propositional Representation claims that a sentence will be structured in another way and then
it is stored. Included information does not get lost. We will have a look at the simple sentence:
“George loves Sally” the propositional representation is [LOVES(GEORGE, SALLY)]
It is easy to see that the propositional representation is easy to create and the information is still
available.
Three levels of representation
This theory says that there exist three levels of representation the surface form, text base and the
situation model. In this example the sentence “The frog ate the bug.” Is already the surface form. We
naturally create semantically relations to understand the sentence (semantic tree in the figure). The next
level is the “Text base”. [EAT(FROG, BUG)] is the propositional representation and Text base is close
to this kind of representation, except that it is rather spatial. Finally the situation model is constructed
by the “Text base” representation. We can see that the situation model does not include any kind of
text. It is a mental picture of information in the sentence itself.
Two levels of representation
This theory is like the “three levels of representations” theory. But the “Text base” level is left out.
The theory itself claims that the situation model is created by the sentence itself and there is no “Text
base” level needed.
Further situation model theories directing experiences exist. So not only text comprehension is
done by situation models, learning through direct experience is handled by situation models, too.
KIWi-Model
One unified model the so called KIWi-Model tries to explain how text representation and direct
experience interact with a situation model. Additionally the domain knowledge is integrated. The
domain knowledge is used by forming a situation model in different tasks like simple sentence
comprehension (chapter: Why do we need Situation Models). The KIWi-Model shows that a permanent
interaction between “text representation → situation model” and between “sensory encoding →
situation model” exists. These interactions supports the theory of a permanent updating of the mental
model.
Chapter 11
Inferencing
Inferencing is used to build up complex situation models with limited information. For example: in
1973 John Bransford and Marcia Johnson made a memory experiment in which they had two groups
reading variatons of the same sentence.
The first group read the text " John was trying to fix the bird house. He was pounding the nail
when is father came out to watch him do the work"
The second group read the text " John was trying to fix the bird house. He was looking for the nail
when is father came out to watch him do the work"
After reading some test statements were presented to the participants. These statements contained
the word hammer which did not occur in the original sentences, e.g.: " John was using a hammer to fix
the birdhouse. He was looking for the nail when his father came out to watch him". Participants of the
first group said they had seen 57% of the test statements while the participants from the second group
had seen only 20% of the test statements.
As one can see, in the first group there is a tendency of believing to have seen the word hammer.
The participants of this group made the inference, that John used a hammer to pound the nail. This
memory influence test is good example to get an idea what is ment by making inferences and how they
are used to complete situation models.
While reading a text inferencing creates information which is not explicitly stated in the text,
hence it is a creative process. It is very important for text understanding in general, because texts
cannot include all informations needed to understand the sense of a story. Texts usually leaves out what
is known as worldknowledge. Knowledge about certain situations, persons or items, that most people
share and which therefore doesn't need to be explicitly stated again. Each person should be able to infer
this kind of information, as for example that we usually use hammers to pound nails. It would be
impossible to write a text, if it has to include all information it deals with; if there was no such thing
like inferencing or if it was not automatically done by our brain.
There is a number of different kinds of inferences:
Anaphoric Inference
This kind of Inferencing usually connects objects or persons from one to another sentence.
Therefore it is responsible for connecting cross-sentence information. E.g. in " John hit the nail. He was
proud of his stroke", we directly infer that "he" and "his" relate to "John". We make this kind of
inferences quite easy normally. But there can be sentences where more persons and other words
relating to them are mixed up and people have problems understanding the story at first. This is
normally regarded as bad writing style.
Instrumental Inference
This type of Inference is about the tools and the methods used in the text, like the hammer in the
example above. Or for example if you read about somebody flying to New York you would not infer
118 | Cognitive Psychology and Neuroscience
Situation Models and Inferencing
that this person has built a dragon-flyer and jumped off a cliff but that he or she used a plane, since
there is nothing else mentioned in the text and a plane is the most common form of flying to New York.
If there is no specific information about tools, instruments and methods we get this information from
our General World Knowledge
Causual Inference
Causal Inference is the conclusion that one event caused another in the text, like in "He hit his nail.
So his finger ached". The first sentence gives the reason why it came to the situation described in the
second text. It would be more difficult to draw a causual inference in an example like "He hit his nail.
So his father ran away", although one could create an inference on this with some fantasy.
Causual inferences create causual connections between text elements. These connections are
separated into local connections and global connections. Local connections are made within a range of
1 to 3 sentences. This depends on factors like the capacity of the working memory and the
concentration due reading. Global connections are drawn between the information in one sentence
together with the background information gathered so far about the whole text. Problems can occur
with Causal Inferences when a story is inconsistent. For example vegans eating a steak would be
inconsistent. An interesting fact about Causal Inferences (Goldstein, 2005) is that the kind of Inferences
we draw here who are not as easy seen at first, are easier to remember. This may be due to the fact that
they required a higher mental processing capacity while drawing the inference. So this "not-so-
easy"-inference seems to be marked in a way that it is easier to remember it.
Predictice / Forward Inference
Predictive/Forward Inferences uses the General World Knowledge of the reader to build his
prediction of the consequences of what is currently happening in the story into the Situation Model.
Integrating Inferences into Situation Models
The question how models enter inferential processes is highly controversial in the two desciplines
of cognitive psychologie and artificial intelligence. A.I. gave a deep insight in psychological
procedures and since the two disciplines crossed their ways and give two man bases of the cognitive
science. The arguments in these are largely independent from eacht other but they have much in
common, though.
Johnson-Laird (1983) makes a distinction between three types of reasoning-theories in which
inferencing plays an important role. The first class geares to logical calculi and have been implemented
in many formal system. The programmin language Prolog arises from this way of dealing with
reasoning and in psychologie many theories postulate formal rules of inference, a "mental logic". These
rules shall work in a purley syntactic way and so are "context free"; blind for the context of its content.
A simple example clarifies the problem with this type of theorie:
If patients have cystitis, then they are given penicillin.
and the logical conclusion:
If patients have cystitis and are allergic to penicillin, then they are given penicillin
Chapter 11
This is logically correct, but seems to fail our common sense of logic.
The second class of theories postulate content specific rules of inference. Their origin lies in
programmming languages and production systems. They work with forms like "If x is a, then x is b". If
one wants to show that x is b, showing that x is a is a subgoal of this argumentation. The idea of basing
psychological theories of reasoning on content specific rules was discussed by Johnson-Laird and
Wason and various sorts of such theories have been proposed. A related idea is that reasoning depends
on the accumulation of specific examples within a connectionist framework, where the distinction
between inference and recall is blurred.
The third class of theories base on mental models and do not use any rules of inferencing. The
process of building mental models of things heard or read. The models are in an permanent change of
updates. A model built, will be equipped with new features of the new information as long as there is
no information, which generates a conflict with that model. If this is the case the model is generally re-
built, so that the conflict generating information fits into the new model.
Important Topics of current research
Linguistic Cues versus World Knowledge
According to many researchers language is the set of processing instructions on how to build up
the Situation Model of the represented situation (Gernsbacher, 1990; Givon, 1992; Kintsch, 1992;
Zwaan & Radvansky, 1998). As mentioned readers use the lexical cues and informations to connect the
different situational dimensions and integrate them into the model. Another important point here is
prior world knowledge. World knowledge also influences how the different informations in a situation
model are related. The relation between linguistic cues and world knowledge is therefore an important
topic of current and future research in the area of Situation Models.
Multidimensionality
Another important aspect of current research in the area of Situation Models is the
Multidimensionality of the Models. The main aspect is here how the different dimensions relate to each
other, how they influence and interact. The question here is also if they interact at all and which
interact. Most studies in the field were only about one or a few of the situational dimensions.
References
Ashwin Ram Kenneth Moorman (1999) Understanding Language Understanding - chapter 5
Baggett, P. (1979). Structurally equivalent stories in movie and text and the effect of the medium
on recall. Journal of Verbal Learning and Verbal Behavior, 18, 333-356.
Bertram F. Malle Louis J. Moses Dare A. Baldwin (2001) Intentions and Intentionality - chapter 9
Boroditsky, L. (2000). Metaphoric Structuring: Understanding time through spatial metaphors.
120 | Cognitive Psychology and Neuroscience
Situation Models and Inferencing
Cognition, 75, 1-28.
Carlson-Radvansky, L. A., & Radvansky, G. A. (1996). The influence of functional relations on
spatial term selection. Psychological Science, 7, 56-60.
Carreiras, M., Garnham, A., Oakhill, J., & Cain, K. (1996). The use of stereotypical gender
information in constructing a mental model: Evidence from English and Spanish. Quarterly Journal of
Experimental Psychology, 49A, 639-663.
Dahan, D., & Tanenhaus, M.K. (2002). Activation of conceptual representations during spoken
word recognition. Abstracts of the Psychonomic Society, 7, 14.
Dupoux, Emmanuel,(???) Language, Brain, and Cognitive Development - chapter5
Ericsson, K. A., & Kintsch, W. (1995). Long-term working memory. Psychological Review, 102,
211-245.
Farah, M. J., & McClelland, J. L. (1991). A computational model of semantic memory
impairment: modality specificity and emergent category specificity. Journal of Experimental
Psychology: General, 210, 339-357.
Fincher-Kiefer (2001). Perceptual components of situation models. Memory & Cognition, 29 ,
336-343.
Fincher-Kiefer, R., Post, T. A., Greene, T. R., & Voss, J. F. (1988). On the role of prior knowledge
and task demands in the processing of text. Journal of Memory and Language, 27, 416-428.
Garrod, S. C., & Sanford, A. J. (1989). Discourse models as interfaces between language and the
spatial world. Journal of Semantics, 6, 147-160.
Gernsbacher, M.A. (1990), Language comprehension as structure building. Hillsdale, NJ: Erlbaum.
Glenberg, A. M., & Kaschak, M. P. (2002). Grounding language in action. Psychonomic Bulletin
& Review, 9, 558-565.
Glenberg, A. M., Meyer, M., & Lindem, K. (1987) Mental models contribute to foregrounding
during text comprehension. Journal of Memory and Language 26:69-83.
Givon, T. (1992), The grammar of referential coherence as mental processing instructions,
Linguistics, 30, 5-55.
Goldman, S.R., Varma, S., & Coté, N. (1996). Extending capacityconstrained construction
integration: Towards "smarter" and flexible models of text comprehension. Models of understanding
text (pp. 73-113).
Goldstein, E.Bruce, Cognitive Psychology, Connecting Mind, Research, and Everyday Experience
(2005) - ISBN: 0-534-57732-6.
Graesser, A. C., Singer, M., & Trabasso, T (1994), Constructing inferences during narrative text
Chapter 11
comprehension. Psychological Review, 101, 371-395.
Holland, John H. , Ke