ReadbyGrade3.com
Reading and Reading Disabilities

Re: three-cueing system: a failed model

Posted By: Dave Ziffer <DaveZiffer@projectpro.com>
Date: Monday, 22 October 2001, at 7:56 a.m.

In Response To: Re: three-cueing system (kerry hempenstall)

I read Marilyn Adams' article on the "three-cueing" system and found it very interesting. She does a good job, but unfortunately she doesn't go into quite enough depth to truly divulge the failures of the three-cueing model. I would guess that like most educators, Adams does not have enough background in information theory (making mathematical models of processes such as reading) to elucidate the issue thoroughly. Or at least, she recognizes that her readers do not generally have such background and therefore couldn't understand a thorough explanation.

Like Adams, I was somewhat perplexed when I first found the "three-cueing" Venn diagram in Regie Routman's "Invitations to Literacy" book. I was even more perplexed to discover how popular it was in the ed schools and within the teaching profession in general.

This diagram reminded me of my early attempts to analye antennas in my "Introduction to Antennas" class way back in college. Basically the way you predict an antenna's output signal pattern is to make a mathematical model of the physical shape of the antenna (almost always a set of calculus integrals), and then apply certain transformations to the model to determine the shape and intensity of the emitted signal, given any particular electrical input stimulus to the antenna.

My problem at the start of the class was that I was an amateur at making mathematical models of physical shapes. So I would walk through all the procedures I learned in the class, only to discover that my model of a given antenna's output signal was incorrect, because my initial model of the antenna itself was incorrect.

Routman's "three-cueing" Venn diagram suffers from the same problem. It is an extremely amateurish attempt to model the reading process. Too bad she didn't bother consulting anyone with any background in information theory before publishing it. If you are going to fool around making mathematical models, you'd better get the math right. Not surprisingly, neither Routman nor any of her colleagues in the teaching profession (who have latched onto this travesty and promoted it to the point where you can find hundreds of web pages on "three-cueing") seem to have enough background in information theory to understand the problem.

The three-cueing Venn diagram depicts "comprehension" at the intersection of three apparently equally-weighted and independent processes: symantic analysis, syntactic analysis, and lexical analysis (she uses slightly different terms, but I'll use the information-theory equivalents here). The primary failure of the model is that these three processes are actually intensely interdependent, whereas Routman depicts them as being independent. This is not just a minor detail. Like my amateurish antenna models, it leads us to entirely wrong conclusions.

Aside from the lack of depiction of the interdepencence of the three processes, there is another glaring failure in Routman's Venn diagram. Routman conveniently doesn't label the pairwise intersections of the Venn circles. What, exactly constitutes the intersection of semantic and syntactic analysis? She doesn't say. It doesn't fit her purpose, which is to convey her personal feeling about the way things work, and so she simply ignores it. The diagram also suggests that there is a direct connection between semantic analysis and lexical analysis, for example, but what is it? Does the semantic analyzer have some direct interaction with the lexical analyzer? Why would the semantic analyzer, which receives sentence fragments from the syntactic analyzer and converts them into meaningful phrases, deal directly with the lexical analyzer, which converts characters into tokens (i.e. lists of possible words represented by character sequences)? This doesn't happen in any information-theory model I've ever seen.

This latter problem becomes even more apparent if we try to expand Routman's diagram to include lower-level concepts (of course Routman's diagram denies that there are any "levels" at all). For example, character recognition is certainly an essential element of the reading process. Routman has arbitrarily chosen to ignore it. Perhaps she means to embed it within the lexical analyzer. Fair enough, but if none of the other levels are subordinate to each other, then why should character recognition be subordinate to lexical analysis? Why not just stick it in as a fourth independent circle, as Routman did with the other three?

If we do that, then we run into serious problems. Routman could conveniently ignore the positions and intersections of the circles in the three-circle triangular diagram because her diagram is fully symmetric (i.e. no matter how you exchange the labels on the circles, it's the same diagram). But a fourth circle would introduce problems. The four circles could not all pairwise intersect each other. Positioning of the four supposedly independent circles would now become an issue. Adjacent processes would share 2-way intersections but diagonally opposed processes would not. What would be the significance of that? How would we choose the two other processes that each of our four processes would pairwise interact with? And of course now we would have four sets of 3-way intersections to explain, in addition to the now nonsymmetrical pairwise intersections. My guess is that Routman would conveniently ignore all these problems as well.

Information theorists never model the information decoding (i.e. "reading") process in as a Venn diagram. They model it as a stack, because the reading process consists of a series of highly dependent subprocesses. Here's what the stack looks like:

5. Interpretation (i.e. comprehension of the complete message)

4. Semantic Analysis (i.e. piecewise construction of meaning)

3. Syntactic Analysis (i.e. word recognition & sentence structure)

2. Lexical Analysis (i.e. token recognition)

1. Character Recognition

Information in this model is imagined to flow primarily from bottom to top. The character recognizer encounters characters. When it has assembled enough of them to form a token (i.e. something that might be a word), it invokes the lexical analyzer to recognize the token. The lexical analyzer generally assembles some arbitrary number of tokens (the number is called the "look-ahead") and then invokes the syntactic analyzer to resolve the tokens into words and to create phrases. When the syntactic analyzer has assembled some number of phrases, it calls upon the semantic analyzer to deduce a meaning. Finally, after the semantic analyzer produces a possible meaning, usually for an entire sentence, it passes this meaning on to the interpreter, which maintains a global perspective and assembles a comprehensive database, containing not only the material being read now but possibly also all the material that has ever been read before.

Level 5 above corresponds to Routman's "Comprehension" at the intersection of the three Venn circles. The next three levels correspond to Routman's three circles. I tacked on level #1 (character recognition) to help illustrate how inapproprate Routman's Venn model is. Character recognition is the next level down below lexical analysis. Any sane person would realize that lexical analysis is almost completely dependent upon character recognition. Yet if Routman had chosen to descend to this level, she would probably have plopped it in as a fourth seemingly independent Venn circle, as described above. Had she done this, the absurdity of the supposedly independent nature of her Venn circles would have been even more apparent: obviously you can't accurately perform token recognition (i.e. determining a set of possible words [i.e. homographs] that might be represented by a sequence of characters) unless you know what the characters are.

A primary source of the entire "meaning emphasis / code emphasis" confusion is the fact that the stack shown above is not in itself an adequate model of the reading process (this whole issue consumes a significant portion of Adams' book, "Beginning to Read"). This is due to the fact that all readers (including both computer programs and humans) have a limited amount of "look-ahead"; i.e. they have a limit on the number of items they can process before passing the assembled structure on higher up the stack. The problem with limited look-ahead is that English sentences (like "sentences" in most computer languages) can contain phrases whose syntactic and semantic structurecan be modeled in multiple ways depending upon what happens later in the sentence or in the story- but since the analyzer at any particular level doesn't yet know what it will find later on, it can only take a guess at its interpretation as it goes. In most cases the guess will be correct, because the vast majority of sentences don't contain this sort of problem. Let me illustrate:

Sentence #1. Underneath the cover was the title of the book.

Let's assume that our syntactic analyzer (level #3) has a look-ahead limit of four tokens, which is fairly typical for humans when reading simple materials. Sentence #1 can be interpreted in two ways, each implying a different sentence syntax. The two possible interpretations are:

1a. The title of the book was "Underneath the cover." 1b. The title of the book could be found underneath the cover.

Here it is not possible, even given the entire sentence, to determine which interpretation is correct (although if #1a is correct, it suggests bad form on the part of the author because the title was not quoted - but in real life we have to deal with these situations). Thus the interpreter, which has no look-ahead limit and knows what preceded the sentence and will eventually know what follows it, may have to resubmit the sentence back down to lower levels for reprocessing if the initial guess at the meaning of the sentence does not fit the situation. Specifically, the interpreter requests new semantics from the semantic analyzer, which in turn submits a request for a new syntax from the syntax analyzer (quoting the book title constitutes a new syntax). Note that each processor can communicate only with the one next to it.

Now let's try another higher-level example:

Sentence #2: I washed my hands of the whole affair.

This is an example where the semantic analyzer must restart itself after making an initial mistake. If the semantic analyzer is looking ahead just four tokens after the start of the sentence, then it will make a mistake - it imagines hands being washed. But at the end of the sentence, the semantic analyzer locates "wash my hands of" in its idiom list, and it must reprocess almost the entire sentence as an idiom. This same sort of problem happens on a higher level as well, where entire sentences or even entire phrases must be reinterpreted on the basis of new information (as when the ending of the story casts an entirely new significance on prior events).

Ultimately, if we want to make our reading processor powerful enough to interpret text that might contain errors, or the reader's sight may be imperfect (as is the case with real-world reading), then we must allow the processes to feed all the way back down to the lowest level. For example, the lexical analyzer must requery the character analyzer if a mistake was made in reading a character (bad eyes). Similarly, the syntactic analyzer may need to request a new interpretation from the lexical analyzer if the book contained a misprint. The more intense the need for error recovery, the more intelligent each subprocess must be.

Thus to complete the stack model, we must insert some arrows. The interpreter must be able to resubmit portions of sentences or entire sentences back to itself and down to the semantic analyzer for re-analysis, to allow for the problem in sentence #2. Similarly, the semantic analzyer must be able to request new interpretations from the syntax analyzer, since quoting the book title in sentence #1 produces a different syntactic structure. And so on down the line, as far as we wish to go. These resubmission capabilities are usually represented by drawing an arrow from each level back down to the level below it. Implicit in all of this is the fact that the higher level processors can pass meaningful information back down to the lower level processors (or back to themselves) that will help those processors arrive at different conclusions than they produced the first time.

------

The failure of the meaning-emphasis movement lies in the degree of importance it assigns to this feedback from higher to lower levels (i.e. to the context sensitivity of reading). Any observer of a typical reading process will quickly realize that the vast majority (probably over 99%) of the reading task can be accomplished with no context sensitivity whatsoever, so long as the lower-level processes are implemented completely and correctly. But the meaning-emphasis folk imagine reading as a nearly totally context sensitive operation, to the point where Ken Goodman refers to it as a "guessing game." Probably Goodman has never actually implemented or even observed a "reading" computer program, or a human that decodes successfully.

The stack model leads us to an entirely different conclusion than Routman's Venn diagram does. The Routman model allows us to arbitrarily assign any degree of importance we wish to the three processes represented by her circles. In its vagueness, we can choose dispense with as much of the lexical analysis (e.g. token and word recognition) as we like, just as Routman enthusiastically does in her book.

In the stack model, syntactic analysis simply does not occur without comprehensive lexical analysis, and so on and so forth up the line, since the processes are heavily dependent upon one another. To the degree that we wish to implement context sensitivity and error correction, the dependencies are bidirectional, although the bottom-up information flow is always much stronger than the top-down flow (i.e. it occurs much more frequently). Yes it is possible that higher-level information may end up causing a reinterpretation at lower levels, but this is done only as necessary (i.e. yes you should rely on context clues if your initial interpretation of phonetically derived tokens does not make sense, but only in that case and as a last resort).

Of course in the computer world we could build a very sophisticated syntactic analyzer that attempts to identify words without fully knowing whether those words might conceivably be represented by the tokens in the sentence (i.e. that guesses words based on context, in the absence of reliable information from the lexical analyzer) - but we would be fools to do so. The amount of work required to construct such an analyzer would be orders of magnitude more complex than simply building a reliable lexical analyzer, and similarly the amount of work it does and the amount of time it consumes would be vastly increased. And of course just as with the meaning-emphasis reader, our syntactic analyzer would have a strong possibility of producing the wrong words althgether.

The "stack" model of the reading process has a lot going for it. It provides a comprehensible model of a complex process (i.e. reading), one that withstands close scrutiny. It is remarkably efficient. As far as I know it is the only model that has ever been used to actually implement real processes that actually "read" complex languages (i.e. both computer languages and natural languages) and successfully convert them into correct interpretations.

In contrast, the "three-cueing" model is merely a vague and poorly thought out notion, so vague in fact that the authors don't even bother to identify three of its seven sections. No sane programmer or mathematician would use it as the basis of an actual working model, first because it contains unspecified sections, and second because it clearly doesn't correspond to the nature of the problem. It is a perfect example of the difference between the teaching profession and all other professions - vagueness versus specificity, laxity versus rigor, blind acceptance versus critical thinking, and convenient bromides versus working models. Oh yes, and laziness versus hard work.

Finally, when we actually implement a "reading" computer program, rather than just thinking about reading in vague terms, we (once again) see the folly of the "meaning emphasis" method - namely that it is much harder and significantly less accurate to infer the identities of words than to simply decode them. No programmer or mathematician in his right mind would subscribe to the idea that the sub-processes involved in reading are somehow independent, or that decoding is anything less than the foundation of a successful "reading" process.

-Dave Ziffer

Marilyn's Analysis of Three-Cueing

Messages In This Thread

three-cueing system (views: 29)
connie weinstein -- Monday, 8 October 2001, at 8:04 a.m.

Re: three-cueing system (views: 24)
kerry hempenstall -- Tuesday, 9 October 2001, at 2:53 a.m.

Re: three-cueing system: a failed model (views: 11)
Dave Ziffer -- Monday, 22 October 2001, at 7:56 a.m.