Gernsbacher's structure building framework has many appealing features, that it has demonstrated significant predictive utility, and that it captures insights which seem consistent with both traditional psycholinguistic constructivism and post-modern connectionist modelling.
1.1 By the time you get to her introduction, Gernsbacher (1990, 1992) has already admitted a number of times, not only in her preface, but in her dedication, that she is "an avid user of language." Her loquacity does indeed strike one as one reads these preliminaries, but the text itself is terse and tight, while remaining fluent and readable.
1.2 The preface also acknowledges help and advice in relation to layout and fonts, but I fear it is in this area I must make my first criticisms. Indeed, others have also commented on the "heaviness" of the presentation -- which would have been much better in a lighter font and either more leading or a smaller point-size. (The choice of the fonts she mixes is also a distraction, though it is only in the table of contents that the lack of balance is readily apparent. It is the heaviness and denseness of the text, however, that strikes one every time one opens the book.)
1.3 Another problem of presentation concerns the figures, which are often virtually identical, with no caption to distinguish the different experimental conditions in effect and similar or identical attributions regarding original sources. As the differences between experiments, or phases of an experiment, can be quite subtle, and since figures do not always appear on the same page as the covering text, it can be difficult to see the significance of each figure without painstaking comparison of text and figures (this is particularly marked in Chapters 4 and 5, e.g. figs. 4.10 to 4.12, 4.20 to 4.25, and amongst 5.7 to 5.15). There are two dangers to avoid in presenting masses of information graphically: one is putting too much information into one figure, the other too little. Gernsbacher (and her colleagues) have fallen into the latter trap. Perhaps they are bound by the particular software package they used. Nonetheless, it should be possible to present an entire set of contrasting figures on a single page with captions that allow us to focus on the differences they are meant to convey.
1.4 Now that I have gotten these technical irritations off my chest, let me say that the volume is well worth reading, and that violating four centuries of typesetting and publication style doesn't preclude one's deriving both enjoyment and profit from it.
2.1 The whole theory presented in the book is based on a single experimental technique for estimating the time a subject takes to read some text (under various experimental and contextual conditions representing the variables under study). The particular computer-age variant used by Gernsbacher's group involves displaying the text (e.g., word by word or sentence by sentence) on a computer screen and measuring the time taken before the subject presses a button to trigger the next presentation. Comparison is made with the work of other researchers using similar techniques, but no reference is made to a century of related research using more sophisticated experimental setups based on considerably less-sophisticated hardware (e.g. eye-movement experiments; see e.g. Huey 1908).
2.2 The starting point for the research program is the observation that "initial sentences take longer to read than subsequent sentences" (p.5) and "initial words take longer to read than later-occurring words."
2.3 For someone familiar with the aforementioned reading research, this brings one up short with a jolt: the eye-movement research tends to indicate that readers speed up around the beginning or end of a sentence, clause or phrase (in English at least), and bog down in the middle. In addition, eye-movement research documents that the eye skips backwards and forwards (it is hypothesized to get an idea of where the author is heading). (It is also ironic, in view of my technical complaints earlier, that reading research also correlates reading speed with the standards of presentation style developed by the printing industry -- except in one respect, it is actually better if the text is not justified precisely, but successive lines are alternately offset slightly left or right.)
2.4 To make this criticism is not to deny the significance of the technique used or the arguments and experimental framework that have been built on that foundation. It does make it clear, however, that the experimental paradigm is significantly different from what occurs in unconstrained reading, although the Gernsbacher paradigm is arguably closer to the sequential nature of the natural conversational and monologue understanding paradigms of speech comprehension -- as in speech we have no corresponding ability explicitly to direct our focus backwards and forwards (although implicit direction of our focus of attention is possible in both audition and vision).
2.5 Above all, it would seem scientifically imperative to compare results from the different reading paradigms, both with each other and with data about natural speech rates, stress and prosody, which should have some relation to the rates of comprehension we are capable of sustaining.
2.6 Reading on, one finds that the words "word" and "sentence" have been used fairly loosely, and what is meant by "word" is something that must be either the head of a noun phrase or at least the first open-class or content word of a linguistic unit -- which may again fail to be an entire sentence but a clause (as in the example Gernsbacher provides subsequently). This point is never clarified, partly because the majority of the examples use single proper names (in which case all three interpretations coincide) or adjectiveless noun phrases are used in which at most one pesky article must be ignored and the two interpretations I have proposed remain unexplicated.
2.7 Gernsbacher interprets these results as evidence that some sort of additional cognitive activity is going on during the reading of these words; she uses this to motivate her structure-building framework: The delays occur precisely at the times when we have to lay new foundations for mental structures corresponding to new participants or themes or episodes, and they are attributed to the cost of doing so.
2.8 Thereafter, delays are equated with costs, either those of building new structures or those of accessing existing structures, and a clever and well motivated sequence of experiments is built to test predictions made within the Structure Building Framework.
2.9 Both Gernsbacher and I need to remember that generalizations about words and sentences must be treated with care, because these are demonstrably the most ill-defined of linguistic units (as is particularly clear from cross-linguistic and diachronic studies). The imprecision with respect to articles is understandable, and fatal, because these words have primarily a functional role (syntactic and pragmatic) and they signal information such as whether or not it is necessary to build a new structure (cf. p.26)! I would predict that these would be skipped over rather quickly, contrary to the precise wording of Gernsbacher's claim.
3.1 The phenomenon which is most in focus in Gernsbacher's research is the advantage of first mention: "After comprehending a sentence involving two participants, it is easier to remember the participant who was mentioned first in the sentence than the participant mentioned who was second." Being easier to remember is, of course, correlated with the delay in their experimental paradigm.
3.2 This is a result which both writers and linguists would do well to bear in mind. Linguistics still hasn't seriously come to grips with such pragmatics, and the question of the choice between apparently equivalent "surface structures."
3.3 The advantage of clause recency, also well-attested and apparently standing in contradiction to the advantage of first mention, can also be explained quite plausibly in terms of the structure building framework, assuming that one can access most easily the substructure one is currently building, and that the foundation laid by the first clause is also preferentially accessible. Predictions from these hypotheses are tested and confirmed rather convincingly using the same experimental paradigm. Note that the theory suggests that subjects shift between structures (and, in particular, to new structures). This too seems to be confirmed in Gernsbacher's paradigm.
3.4 Other important phenomena investigated include the activation of words with related meanings or forms, facilitation due to anaphor, and mechanisms of enhancement and suppression in general. The chapter dealing with all this (Chapter Four) is worthy of close attention.
3.5 To round off this review and give a more complete impression of the flavour of the book, it may be useful to list some of the other areas where structure building proves to have a useful explicative power in terms of the experimental paradigm used. These include evidence for hierarchical structure and the representation of clausal dependency and evidence for mapping facilitation resulting from coherence of several kinds: referential (previous mention of the concept), temporal (overlapping time frames), locational (spatial overlap) and causal (logical or cause-effect relationships).
4.1 In the introduction to Chapter 3, Gernsbacher wraps her theory in a papier-mache model which serves to emphasize its Piagetian aspect and thus its possible connections with learning mechanisms. Comprehenders develop their mental structures by mapping information onto the developing structure, as if adding another layered strip. Chapter 4 opens, by contrast, with a neural analogy, in which the mental structures are more active than just mere representation.
4.2 I believe it is useful to capture and compare these analogies, because there is some neural wetware in there somewhere, and the neural model captures some interesting features of language, but current connectionist and neurophysiological models have not focussed on structures, and indeed encourage a contrast with more symbolic or constructionist models.
4.3 The papier-mache model has the attractive feature that it captures the insight that as you add more layers, you occlude some existing details while adding to the overall weight and accenting particular features. Neural networks and neuroanatomical systems also have a layered structure, however. We have very little idea how mental structures are actually represented, retained and refined -- or whether indeed they can rightly be called structures. Layers aren't physically added, but perhaps the neural layers do refine the concept or schema in some analogous way.
4.4 What is clear though, is that Gernsbacher's structure building framework has many appealing features, that it has demonstrated significant predictive utility, and that it captures insights which seem consistent with both traditional psycholinguistic constructivism and post-modern connectionist modelling.
Gernsbacher, Morton Ann (1992) Precis of: "Language Comprehension as Structure Building." PSYCOLOQUY 3(69) language-comprehension.1.
Gernsbacher, Morton Ann (1990) Language Comprehension as Structure Building. Erlbaum
Huey, E. B. (1908) The Psychology and Pedagogy of Reading. MIT Press.