Pre-Application Meeting Videos and TranscriptsApril 22, 2002
DR. SWEET: We have a guest lecturer with us, his name is Thomas Trabasso. His topic is Dynamic Comprehension, Learning and Memory and I particularly like his title, Understanding Understanding. And before I turn the mike over to Tom I'd like to introduce him.
Tom is the Irving B. Harris Professor in Psychology at the University of Chicago. And, as you know, he's a cognitive scientist who is internationally recognized for his research on reading comprehension. He's been Chair of the University of Chicago's Department of Psychology and most relevant for this session he did serve as Chair of the National Reading Panel's subcommittee on reading comprehension. He's also served as editor or editorial board member of numerous prestigious journals such as Cognition Instruction, Cognitive Psychology, Discourse Processes. I could go on and on about Tom, but we've run out of time. So, without any further hesitation I present to you Tom Trabasso.
DR. TRABASSO: Now, the title Understanding Understanding, my word processor kept on rejecting it saying, you know, you've repeated the word. I said, no, this is really what I'm very serious about. So, part of our goal is really to understand what understanding is about and that's why the title is there. The subtitle describes the approach which I'm taking which is dynamic in process and contrasts with probably the typical notion, though, of comprehension.
Someone asked earlier about a definition for comprehension and typically in the operation you find it's someone reads something and then they do something with what they read. Okay? So, you might answer some questions, you might be asked to remember it and so on.
So, especially to ETS I'm sure that operational definitions of that sort are often used to define what you mean by reading comprehension, but you'll see it's much more than that. I'm concerned with what happens while you're learning to read or why you're reading and reading to learn and how this, how understanding, itself, leads to learning and leads to changes in memory.
Now, I'm concerned with both while you're, during reading, that is activity during reading sometimes referred to as on-line comprehension, but I'm also concerned with off-line. I'll give you examples of both throughout the talk so you understand my notion of understanding understanding.
So, understanding, someone asked earlier what is it? One definition which is traceable back to Bartlett in 1932 is the search for meaning. Now, this doesn't really give you a good idea because it's a little bit circular, but let's go to the next slide. And there's two senses and meanings that I wanted to talk about, the notion of relational meaning or relational understanding and the other is the achievement of coherence. All right?
Those are the two aspects that are going to inform us today and that is really what we're seeking to do; that is, we want the students to understand relationships or find relationships amongst ideas and we want them to be able to integrate them into a coherent representation.
Now, this quotation which I'll read to you is one of my favorites. This comes from John Dewey in 1933 and he wrote, to grasp the meaning of a thing, an event or a situation is to see it in its relations to other things, to note how it operates or functions, what consequences follow from it, what causes it, what uses it can be put to. In contrast, we have called the brute thing, the thing without meaning to us, is something whose relations are not grasped.
This is a very important quote and it's been around for a long time, but I don't think it's been quite realized in the meaning that I work with, that is, the meaning of things really is in relation to other things and you're going to see that strongly as I develop my arguments today.
Now, the search for meaning establishes coherence. What do you mean by that? Well, when I was working on causality and I spent the better part of five years of my life reading on causation, that is historical, philosophical, biological, psychological sources before I actually got into using causal analysis of narratives, I found this quote and I like it as well. When we witness a series of events we do not experience them as isolated series, rather we experience them as a coherent whole. All right?
So, there is a relationship between this quotation and the one by John Dewey that is a relational notion that somehow or other through relating ideas we're going to integrate them. John Mackey, who is a philosopher who wrote on causation had a title of his book which I like, too, which is really big coherence, okay? Causation, the Center of the Universe. So, we're going to focus a lot on causal relations.
How are relations created and coherence achieved? Now, I'm going to outline a series of process statements which I'm going to translate into more detailed theory. When people try to understand events they do so by relating them to other known events and relationships are established through three means, and I'm going to show evidence for this, explanatory, associative or predictive inferences. Next. Back, I'm sorry.
The inferences are made during reading that unite the past, that is what we've read in the past or thought about in the past, with the present, this is whatever we're focusing on at that moment, and the predictions aim at the future and achieve coherence together.
So, I'm going to present what I call a theory of dynamic understanding which is an adaptation of these ideas I just outlined. There's a lot of current work being done on so-called situation models and a lot of these situation models have the reader creating some kind of representation of things that are happening. It could be visual in a sense of the real world kind of situation, but it can be a little more abstract and my adaptation of these models is represented in the following series of statements.
First of all, when we're reading about agents readers create representations of them and sort of follow them in space and time. So, I said readers track agents in space and time which is part of the spatial models that people have talked about. And what's important to me is that you monitor their concerns, that is you're interested in what affects them and the goals that they have and why they're doing what they're doing.
So, readers update knowledge about agents as new information accrues over time. And so, people I met today are coming back. I've updated my memory representations of several people I've known in the past. I'm also creating new representations which I'm going to use in the future. Readers establish meaning and coherence through relations among ideas, through explanation, association and prediction.
Now, we're going to do a demonstration of dynamic understanding with you. We're going to show you a story one sentence at a time. I can't obtain permission from all of you so you'll have to give me an exemption on this, your participation, and I will not use your data in any way, but I want to use it mainly as a demonstration experiment.
So, you're going to see a story presented one sentence at a time. I would like you to try to understand each sentence in the context of the story and at the end of the story I will ask you about your understanding. We've got to pace these now. Okay.
Okay, let me ask you a question. Show the question, the next one. Can somebody give me an answer of why she knit that sweater, why she wants to? She enjoys knitting, okay. Where did you get that from?
SPEAKER: It was in one of the sentences.
DR. TRABASSO: Yeah, right. You retained some text in order to explain that. It was a carryover from the previous one. All right. Okay, anybody else? Yeah.
SPEAKER: That person definitely needed a sweater.
DR. TRABASSO: She does?
SPEAKER: I said the person needed a sweater.
DR. TRABASSO: Oh, the person needs a nice new sweater. So, you think maybe it's possible she's going to give it to her mother?
SPEAKER: Possibly.
DR. TRABASSO: Possibly, okay.
SPEAKER: She saw her friend doing it.
DR. TRABASSO: I'm sorry?
SPEAKER: She saw her friend doing it and...
DR. TRABASSO: Yeah, you'll see some interesting protocols on what that friend means to her. Okay, we're going to do another story. This is a very similar story except for one or two changes. This is like an experiment, right? We've got a variable in here, all right? So, this is Betty 2 and both Betty 1 and Betty 2 are going to become quite important. Let's go through the same process, okay? You can go a little faster now, up to a point.
Okay, now I'd like some more volunteers. Why in this situation does Betty want to or decided to knit a sweater, anybody?
SPEAKER: Because she wanted to have it for her mother's birthday.
DR. TRABASSO: That is a failed goal outcome, that's correct. Yes, and?
SPEAKER: Because she decided to knit the sweater for her mother.
DR. TRABASSO: Okay, so she had the goal to give her mother a present. That goal failed and now we see a new goal being generated to knit a sweater which would satisfy, help move towards the satisfaction of the other goal, right? Okay, good. Anybody else?
(No response.)
DR. TRABASSO: What I'd like you to understand here is the major difference in understanding that it took place by changing only a few sentences. In a lot of these kinds of stories I can change only one sentence, either success or failure. Success or failure of what? Success or failure of the person's concern, of their goals.
So, if you're tracking Betty and updating your memory of her as you go through, you're learning quite a bit about her and you're probably thinking and adding things to your representation of her, as well, but the important thing about this example is whether or not you're monitoring her goals and then using them to explain or understand her actions and the other goals that she may have. This is what you did. Thank you very much for that very nice inciteful set of inferences.
Okay, now, my purpose is to develop and test this theory of dynamic understanding and what I'm doing, what I'm concerned with is how well I can use the theory to account for narrative comprehension data. Now, although I'm talking about narrative comprehension, at the end of the talk I will show you several extensions of this work into non-narrative or expository text, as they say here, and I'll give you examples of that. So, it's not, this is not restricted entirely to just stories that kids are going to read in elementary grades. Rather it's applicable to any situation in which there are agents, like institutions or people upward. It's relevant to biography, to history, to social studies and so forth. And that's one of my final remarks, but I'll bring it up front in order to point out that we're not just talking about understanding stories, we're really talking about agency.
Okay, now, in this theory we're going to have readers reason about how and why agents experience state actions and state changes over the course of the narrative. okay. So, we're talking about dynamic understanding. This is going to be over the entire course of the narrative, what's happening, and I'm going to argue for explanation based understanding. That is, the readers in a sense are using what they know to understand and explain, for the most part, what's coming in. All right? And they're using not just only knowledge of the text, but they're going to use information from other thoughts that are activated by what they read and they may set up predictions and some of these predictions actually come true in the text, themselves, but for the most part it's explanation based. And explanations are done by making causal inferences.
So, I said it depends on upon what information you access. It can be from the text, it can be from your world knowledge and it can be from prior thoughts you've had during the course of reading. And you have to relate what you activate or what you access to new information in the current focus. This is where memory comes in and makes a big difference.
So, when the reader explains what's happening using something say from the text and is focusing on some current sentence, then relationships between ideas are learned and they become integrated into memory. Now, as new information or conditions are encountered and they're explained and they're integrated, what is known about these entities, these agents and so on is updated and accessed to memory changes.
Okay, and so then the theory of dynamic understanding involves processes of causal inference, learning and integration. These processes when carried out will update and change what is in memory and they also change the accessibility of what is in memory.
So, we're going to test this. We're going to test all the implications I can think of with respect to all those statements I made. The first one is about causal inferences and comprehension. So, one question is when and where inference is made during reading. And part of the answer lies in doing discourse analysis. One of the main procedures that several of us use who are doing work with text is to analyze the potential inferences that could be made in the text.
So, in this research I'm going to report, then, we're interested in whether or not people make the inferences that we think they should make at various points in time during reading. What we we're going to do, I'm going to vary the structure of the text, and we did this in our studies, you actually experienced two texts. You had a sequential text, that was the first one where Betty was successful and gave her mother the present and then we had a hierarchal text in which Betty suffered frustration, a failed goal, had to generate a subgoal, carry that out, and there's more to the story than I gave you, but she does finally succeed in the end. Okay.
All right. Well, now, I need to go over here to talk about this. I wish I had a pointer, but I'll use this. All right, these are two structures that we have used to actually generate stories. See, we actually make up experimental stories this way. The Betty story starts out with this girl named Betty. This was a setting statement, Betty finds out her mother's birthday is coming soon, she wants to give her mother a present, she goes to a department store, picks out a purse and she gives it to her mother. Get this point, I'm sorry, this is the wrong story. In this story she goes to the department store and she finds out she can't buy what she wants, okay?
SPEAKER: Right.
DR. TRABASSO: So, she fails. Now, she sees her friend here knitting, and this is where you made those nice inferences, right? She wants to knit a sweater, too. And here we see a link between her desire to get a present for her mother, the failure of that desire here being determinants of this particular, possible determinants of why she is doing what she is doing here. She then takes out a pattern and she follows these instructions carefully, she knits the sweater and these activities here are motivated by a subgoal.
And one of the things I want you to think about is when we're down here and she's in the activity of carrying out that subgoal, are you thinking about this goal up here? The evidence suggests otherwise. When people are focused at subgoals, that's what they focus on.
So, she completes that sweater here. Now you return back to this goal here which is going to motivate various attempts and in this particular story I think it was only like she pressed the sweater, folded it, put it into a box and gave it to her mother. And that's enabled by her having finished the sweater and motivated by her goal to give her mother a present and that's a hierarchal story.
Now, what I want you to think about, also, is how connected this story is, all right? In this story we have Betty going through the same activities, but here she's able to find the purse, gives it to her mother, sees her friend, decides to knit a sweater, follows the pattern, makes the sweater, presses it, puts it into a box and puts the box away or something like that. She's happy.
So, this particular story is very linear. I ask you, which story is more coherent, the hierarchal story or the sequential story, which one is more coherent?
SPEAKERS: Hierarchal.
DR. TRABASSO: Why is that? More connection somebody said. Right, that's the idea. I'm going to show you that, in fact, is the case. Okay, now, one of the studies that I did with Joe Magliano and with So Yung Su at the University of Chicago was to use think out loud procedures. This is a technique which I recommend teachers use because you can find out an awful lot about instruction and whether you're being effective and how kids understand what they're reading. And I'll give you examples both on college students and on nine year olds, okay? So, it's not restricted only to University of Chicago students.
Okay, wait, right. Get the instructions. The instructions were to people, you're going to read a story one sentence at a time and try to understand the story in the, try to understand the sentence in the context of the story and then tell me about your understanding. Simple. You're going to see some interesting results, okay?
Let's look at the first, or the Betty Number 2, the goal failure story, and I'll show you the next slide. Okay, this sentence says Betty could not buy anything for her mother and then she says this is because everything was so expensive. What she's doing is carrying over and explaining. Go ahead. Betty felt sorry. This is because she wanted to give her mother something special as you can buy in a department store, but she didn't have the money and she felt that maybe she'd be shortchanging her mother without giving if she didn't give her something from the department store. Okay?
Now, look, it's a very long elaborated explanation for her emotions and an indication that the reason she is feeling sorry is because she wasn't able to do something she wanted to do which fits in with my wife's theory, Nancy Stein, theory of emotion and how goals which are not met lead to various kinds of emotional states. Next one.
Several days letter Betty saw her friend knitting. This is a very smart young lady here. She says, this is incidental that she saw her friend knitting, but this could put ideas in Betty's mind that she could learn how to knit and make something for her mother. So, that's prediction, okay? And also kind of explanation, but it's prediction primarily. Okay, next.
Betty decides to knit a sweater. Hmm, that's a very difficult gift to knit. So, again, it shows that Betty really cared for her mother. This was a thought that she had early on, she wanted to do something special for her mother since she really cared about her. Then she carries this theme all the way through the story and wanted to give her mother something special, which is the retrieval of a text, more than just a rinky-dink, rinka-tink thing children with no money sometimes end up giving.
So, now we want to contrast that, that protocol, with the sequential story on the same sentences. Betty found a pretty purse. Okay, she thought that her mother would really appreciate it. That's in addition to, that's coming, that's an association in our terminology, the pretty purse, and probably she is going to ask how much it was or buy it. That's prediction.
Betty bought her mother the purse to give to her for her birthday, I'm pretty sure. okay? Her mother is happy. You know, I imagine her mother was happy because she got the purse and Betty said, here, mom, this is for your birthday. That's an association. You get a lot of this kind of stuff we use, filling in, associating. Okay next.
Several days later she sees her friend knitting. Okay, not much to say on that, interesting. Next. Betty decided to knit a sweater. She probably got the idea from seeing her friend. She thought, wow, that's really pretty what my friend is making, I'll make one, too. All right, somebody offered that and I thought it was kind of an explanation for why she wanted to knit it. Okay.
So, now, there are two mechanisms here I want to talk about, accessing memory and making inferences. In the next slide I give you a little theory about that. Oh, keep going, next one. Here we are, you're reading this sentence right here. You've already read all these, we're up to this point. This is where you are and we'll call that the focal sentence.
So, the question is, what's going on during this time? Well, we found several things happening. There are these processes of explaining, predicting and associating, but there's also a lot of memory involved. So, for example, way back here might have been a sentence such as she wanted to make something for her mother and maybe this is the sentence where she sees her friend knitting and then someone says, oh, I think she'll probably want to make something for her mother. Okay?
And what they're doing is retrieving out of, we'll call it long term memory for the moment, it's certainly far back in time and they retrieve this and make a prediction and the prediction is generated off the sentence. When they elaborate and add information we'll call that association and sometimes, as someone did also in my demonstration, we simply carry over a sentence and use it to explain. Okay?
So, these can come from, then, text sources being retrieved or information which is added to the situation by the person through association and can then become a source of explanation later on in time. And I'll show you some evidence for that. Okay.
These are data based upon, I think, eight college students, 2,800 and something thoughts which we had to classify and work with. Maybe more than that, but that was quite a large number. These are based upon 16 stories. Each person read, I think eight versions, half of which were sequential and half were hierarchal and so we had both kinds, but the important thing is to look at the sources of what was used for explanation, association and prediction and look at the percentages of each of those three types.
The light blue is explanation and you'll see that predominates. So, when it's something new that they're thinking about, about nearly 60 percent of the time it's going to be an explanation, but a large percentage of new activations are associations. That's the black on the left side and then they make predictions about 10, 15 percent of the time. That's where we're bringing in new information to the text.
If they're using the text itself, which is the middle bar, then you've got 90 percent of the time it's explanatory. This is what I meant by comprehension being explanation based, and then it's a very small percentage, about 5 each, are associations and predictions. Predictions don't happen much. People like to be sure. All right? We live in the past as somebody said earlier today.
The prior activations are thoughts they had before. How did they use them? Well, they explain a lot. About 82 percent are explanations, these are previous thoughts. In other words, they elaborated the text and then they use it to explain. And then about, I don't know, 7 or 8 percent are associations and another 7 percent or so are predictions. Overall I think predictions are around 11 percent, associations are about 20 something percent and explanations around two-thirds of the time. So, this is what I meant by people are going to use what they know either from the text or from their world knowledge to try to understand and explain what's going on. Okay.
Now, do young readers think like college students? The answer is evidence from thinking out loud. Let's look at the protocol. So, here we have, this is a story about John, this is a typical sibling rivalry conflict about becoming King and the father is the King and so John works hard to make his father happy.
Now, we went further in this particular example, let me give you categories of what they're talking about. They paraphrase, that is they simply restate what's there, but it goes somewhat differently. They explain and sometimes they make predictions. Okay? In this case, go back, no, the next one, next. No, go back, back, back, you're going forward. Back, back, back one more. Okay.
So, John works hard to make his father happy and the kid explains that, well, because his father was the King and he wanted to be King, all right, when he grew up. Next one. But his father gave the throne to his first son instead of John. Well, maybe since he's the first son he gave it to him instead of John, maybe he was more important or something. This is a third grader, a nine year old protocol. This is not, I didn't pick these out like they're somehow different or special, they are representative. Okay, next slide.
These are the data. So, I'm comparing University of Chicago students, which is an unfair comparison by any means, but they're in black and you can see the shape of their function here. That is explanation predominates, paraphrase is second, but it's not a close second and then you've got predictions and associations in the middle. The pattern is the same for the third graders. They paraphrase a lot. They're still viewing with what's there and I think this helps their memory and, in fact, we have strong evidence that paraphrasing what you're reading does make it more available later on in time. So, I examined what's the role of paraphrasing? It's to maintain memory.
Well, they also explain a great deal. So, they're explaining about one-third of the time, they're predicting more than the college students are, which is fine, and they're associating a little more on average relative to what they do which means that they're trying to bring in what they know to help understand the text. Okay.
So, the bottom line is there's a close resemblance between third grade readers, which I think is an important time with respect to comprehension, and college students. And this procedure of thinking out loud we found very easy to do with these 24 third graders, with two different kinds of text, of course. Go ahead.
Well, what's the effect of doing this explanation and prediction? In our theory the idea is that you integrate information and you're going to improve your memory. So, is there a relationship between third grader reader explanations and predictions during reading and memory for the text? And the answer is, next slide, there are 24 people in this category, all right? The important thing is the correlation at the top which is .58. It's a positive relationship between the average number of explanations and predictions, which is on the ordinate and on the abscissa, I'm sorry, on the ordinate we have the number of sentences recalled, the average number of sentences recalled, and there's a strong relationship between memory and integration of the information either thorough explanation and/or prediction.
So, those kids who did a lot of integration and a lot of prediction remember the text the best. That's a pretty strong correlation with only 24 children. So, I'm happy with that and that's good evidence for that point. Next.
Now, we're going to go into a little more complicated modeling process so we're going to do a simulation of dynamic understanding. I'm going to carry you through this mess and it's not as complicated as you think, but it illustrates the whole approach. You start with the text here, you do a discourse analysis. This is something I would argue for as being quite important. And every time I've done it against other people's studies, I do a lot of simulation with other people's work, when I've done it I learned more about their experiments than they knew, all right, because they don't usually analyze their text that carefully. Sometimes I agree with them and sometimes I don't which is rather interesting and gets you into controversy and arguments with your friends.
Okay, anyway, discourse analysis is necessary. This generates a network, like the Betty one we talked about. So, I could take Betty's story, the complete story, analyze it, and I'll get a structure like this. Now, this structure is actually built up by the reader one unit at a time. So, what we do is we have a connectionist model right here, right here, in between here which is going to take in one unit at a time plus its relationships. So, I'm going to feed in the first node, which is the first sentence or the first clause, and the model of output and matrix connection strengths and end by a matrix. So, if we put in the second one there's a connection there, we'll get a two by two matrix of connection strength. We put in the third node and we get a three by three matrix and so on.
What's output here, now, in each cell of the matrix is a connection strength between any two ideas. This tells me how accessible is one idea from the other. So, if I know they're reading Idea J here and I want to know what's it like if they're achieving Idea I, then that index will tell me that.
So, we have to have the appropriate measure. Now, I'm going to use sometimes the connection strength between two ideas. If it's during reading and I want to know whether something's available I'll use this measure between ideas, but sometimes, like if you're judging the importance of the main idea of the whole story, then you want to know the average connection strength. In other words, how available is that particular idea to all the other ideas. So, I have two different ideas that I work with a lot in this model.
The test of the model is how well it compares with the actual data. Now, we have a shopping list of methodology that I use. I've got five kinds of methodology. Discourse analysis is quite qualitative, all right? But it's fairly rigorous and that gives us an analytical representation. However, these representations, I can vary the discourse in the experiment and manipulate these representations and get different connection strengths as a result. It's also quantitative because I'm putting it into a connectionist model which gives me exact measures of some index proxy for how accessible something is. It's also correlation in the sense I'm going to correlate these measures over here, okay? So, it's a whole shopping list of methods straight across. so, I don't know what to write down on my grant.
SPEAKER: All of the above.
DR. TRABASSO: All of the above, yeah, okay. All right, any questions on this model? You don't have to understand the details of it. The basic idea is to analyze the text, all right? We looked through all the relationships. Now, the reader is going to find them one step at a time. We're going to simulate that by how we feed the model, but the success of this enterprise depends upon how good the discourse analysis is and what we put into the model. In fact, that's going to carry more of the variants in the model, itself, but the model is a nice way of quantifying what's going on.
And there are several other people who have done this kind of process of analyzing text. They work at different levels than I do and they may use different kinds of input, but the idea is basically the same. This particular model is a variant of one used by Walter Kinch.
SPEAKER: Did you see any distinction between behavioral data from the various text structures?
DR. TRABASSO: Sure. We're going, I'm going to answer that question in detail in a minute, okay?
SPEAKER: Okay.
DR. TRABASSO: Yeah.
SPEAKER: I assume your findings on relationships in this model using the same constructs you defined earlier of explanation...
DR. TRABASSO: It's primarily, this is primarily causal explanations.
SPEAKER: Okay.
DR. TRABASSO: What we're doing is identifying the potential causes that people could infer. Whether they do that it's an empirical question, right? In other words to think a lot of protocol is one kind of evidence. So, I have to predict that. So, it's our analysis, but the reason we're successful with this is because we all share similar naive theories of causation and how agents operate and act and so on. Okay, next.
Okay, I'm going to talk about monitoring the fate of goals, what happens to goals over time, if they're satisfied or not and how easy is it to access these goals to explain other goals or actions or outcomes during comprehension. Next.
So, we're going to look at the proportion of people who, when they read those Betty stories and other stories like it, referred to the main goal, the goal, the birthday goal for example, and used, I think, a lot of protocols. Let's see the data, okay.
Now, on the left side what you see are the data for the sequential story and what's there on the abscissa are the sentences that go over the course of the story, the various goals and attempts. And what's plotted is the probability of reference to that goal. When a goal succeeds early what happens is they don't refer to it any more, although you tried to with respect to the sweater, I know. Well, a few people do that, but not too many, okay? It drops out, they get into other issues like she's in competition with her friend, stuff like that. We had eight different stories, so it's not just the Betty story, but that's a good example.
So, what happens when a goal succeeds is it becomes less and less available in time. And that's actually built into this model anyway; that is, distance in a network leads to less accessibility. On the right-hand side is the hierarchal story where you had goal failure and then goal reinstatement. So, the beginning is a little bit like the on the left, it's very available, it's used to explain why she goes to the department store and why she's sorry and so forth, but then when the subgoal comes in and it starts to determine what she's doing, there are some references to the subordinate goal, but not as many. And, then, once she completes the sweater it comes back in again as an explanation. So, it's a recovery.
So, the fate of goals either is to be lost and become unavailable, as soon as it's over with, it's over with, but if it's unsatisfied you're probably going to be using it to monitor what the characters are doing. Okay, hold it. I'm going to show you in the next slide our simulation of these data, okay, and you tell me which are the data and which are model. Go ahead.
Okay. The data are the dark lines, and I don't know if you can see them, but the dotted lines are the model which means we've got a very, very good simulation of these data. Well, what's being simulated? What I'm looking at is, this is the first attempt here, and what's being used here is the connection strength between the first attempt and the main goal. And since they're close in time and was just used it's very available.
But the second goal, the subgoal down here, is now further away, less available and its connection strength is going down and down here it's pretty much flat. There's some little perturbations in here, but the connection strengths are quite low. So, it's an exercise between that main goal and each attempt and each goal and each attempt and so on. That's what's being used here to predict these data. Here the goal was available, you know, high connection strength, it drops down, recovers and that's the function.
So, I have several examples like this where the shape of the function can be beautifully mimicked. The actual measure, itself, means another theory tacked on in terms of how you retrieve information, but basically it's a pretty good simulation. Yes?
SPEAKER: The scales are...
DR. TRABASSO: Yeah, I know, yeah, they are and I can show, but they're very different. I mean, the thing is there are two arguments you can make. One is you can put them on the same scale, and I've done that, and you will see if I put it on the same scale the shape is there, but this one, this model will be a little higher like this. Okay, but this one fits beautifully. Okay? And I've replicated this on three different experiments by Rabinski and Lutz, Lutz and Rabinski, and there I put them on the same scales and it was quite excellent all the way through.
In some cases we underpredict and in some cases we overpredict, but these are not exact measurement of the proportions. In other words, there's going to be some translation between connection strength and the actual behavioral data. This model doesn't do that. Okay?
You're right in pointing it out, but the shape is okay. In this particular one I used a program which gives me a bivalued ordinate and it fits, you know, it actually gives you the best fit it can under the circumstances. And so, it turns out that the shapes are quite good. So, if you think of it in terms of mimicking the qualitative aspects of the functions it does a very good job. It's very hard to know whether you're underestimating or overestimating, but if you put them on the same scale you get some idea. And it would be important, especially if they don't make all the inference you think they should make, right?
SPEAKER: You're not overestimating, you can't like, if you're overestimating in the second case then you're not consistently overestimating in the case where you do have comparing or vice versa.
DR. TRABASSO: Well...
SPEAKER: Do you think that's the case, when you have preliminary...
DR. TRABASSO: I think, you know, a good study to do would be one in which you have different kinds of readers, some who are more likely to make inferences than others and then look at the protocols and see whether or not you over, you should overestimate far more for the poorer readers than you do for the good readers, okay? Something like that. All right.
Let's look at story recall. Now, this one is, this is recalling the whole story, what percentage of sentences do you remember. the idea is that the more connections you have it's going to lead to greater connection strength, it's going to lead to greater accessibility and behaviorally a higher percentage of recall. So, show that next slide.
Okay, this is a correlation between story recall and connection strength. The connection strengths are running along the ordinate, no, the abscissa. Why do I keep getting that wrong? The abscissa and, that is along the base or the connection strengths and you'll see the correlation there is about .51. So, that's pretty reasonable. This is now a complete story, percentage of sentences remembered per story. Next one.
Here we're looking at story coherence. Here we had, I think, 24 people read the stories and then rated them for coherence. They had the same idea, more connections, greater mean connection strength. Now I'm using a different measure, mean connection strength. I used that in the previous one, as well. That is, the overall accessibility of a sentence is indexed by all of its connections to everything else, which is reflected in the model, and the greater that is, then the greater the accessibility and the higher the ratings of story coherence.
Now, what I want to show you is something in these stories. Up here are all the failure stories. Remember I said which one was more coherent and most of your agreed, okay? All these F's up here are the failure stories. These are the sequential stories down here. And the correlation is pretty. good. Our experience is 51 percent so the correlation is .7 or so. Not bad.
Finally, is this the last study? I think so. This is a study on memory. People read stories and for half the stories they have to do immediate recall and then two days later they do delayed recall and we're going to compare connection strength of individual sentences against recall. This is a rather interesting finding and it's totally consistent with Raynard and Reyna's theory of just understanding and fuzzy trace theory memory; that is we found that if you do immediate recall it's independent of connection strength. it doesn't matter. Immediately the whole story is available. You wait two days and what's left are the things that are only highly connected. All right? Which gives me some confidence that we're dealing with meaning and just not with verbatim recall. Next slide.
This is the retention function or, I'm sorry, this is the loss function. So, what gets lost? The stuff that's got the weakest strength. There's a very nice relationship, I've forgotten what the correlation is, it's quite high, between mean connection, strength and retention. So, the things which are left, this is the difference between the immediate minus delay is another way of looking at the same data.
All right. So, I've marched you through some experiments in which we had experiment and control groups so-to-speak; that is we had differences in text which would either promote causal inferences or not. And the conclusion I want to make is that I think it is the case that when people read or listen to stories or watch movies or watch plays or talk to their friends who are giving them autobiographical comments where people are telling personal stories, the people who are listening or reading about it track the agents in space and time. And they do monitor concerns. In other words, they're looking at people's goals, they explain states and actions of agents, they integrate and update information in memory and they have to access memory to do these things, which is a rather interesting kind of circular interaction between the processes; that is by integrating and updating it changes excessively and it's either going to make it easier or harder to explain, predict or elaborate. Okay, next one.
So, now what I've done with you today is to give you a feeling for the approach and some success we've had with thinking aloud and with memory and coherence and so on. I've extended this work to a large number of studies, not just my own, but other people's and literature. I think I have something like 35 different studies now that I've addressed using this model and its analysis.
The first one is on judgment which is a process, I think, used often to study comprehension. It usually operates at the sentence or clause level and you can have people do different things. You can have them look at pairs of sentences and judge causal relatedness, how strongly related are two things. And the model predicts distance between ideas in the network. The further apart in the network, the less causally related. The closer together the connections, then they're more strongly related.
I've developed a new technique called Goodness of Fit. A person goes through, they read the story one sentence at a time, they have to judge, how well does this fit into what I know so far? A very nice technique, quite simple, it can be used with young kids.
And then I analyzed six Chinese folk tales, and these are extremely long, 160 clauses in some cases and we looked at importance judgments of those. This is like the main idea and showed a linear relationship between number of connections and judgments of importance. So, when you're judging the importance of something you're relating it to other things in the story. That's what that says. Okay.
Someone, oh, yeah, today Grover Whitehurst said, who is she? Well, who is she is the anaphoric reference. It turns out that searching for references, either agents or spatial locations or objects, I can simulate recognition time, I propose to simulate recognition time. For example, Gordon Balor has a number of studies where people memorize a map and then he tells a story and the character goes to the laboratory checking out various rooms and making sure that they're clean because the visiting community is coming and so on. And they look at spatial locations and distances, but there's a story and so I can actually simulate the accessibility of Room A given he's in Room Z, all right, or Room B given he's in Room D. And it turns out that there's a nice linear relationship between spatial distance and connection strength. Oh, go back one.
With respect to agents phenomenal reference. When you're tracking somebody, there's a nice book called Narrative Understanding by Catherine Emmett, and I recommend this book for those of you who are interested in narrative. Catherine Emmett, she's a Linguist at the University of Edinboro, and I did a review on this book for a journal and I was very impressed with it. In fact, I think she, along with a few others, gave me the idea for a lot of this work. She said, you don't need to do a search for a pronoun because the entities you're tracking are there. All they need to do is become, either they're accessible because you just talked about them recently or you have to find other ways to find them, but it isn't necessarily a search for a person through a pronoun in the sense of some kind of active search. Rather it's a matter of accessibility.
So, if John and Mary make a date to have dinner together and then the story goes on and talks about John at work or Mary at play, or whatever it does, and then the next sentence is that evening with the candlelight, by the candlelight, that may be enough to bring back, oh, yeah, they're having dinner, oh, yeah, and then you know it's John and Mary. Okay? In other words, the reference to the people who are there isn't necessarily done by some kind of active search, but it could be done by associations and activation through things that are, other things that are present, but they're tied to these entities. Okay, next.
Okay. So, other processes during reading that I'm also interested in are reading time, how much time it takes to read sentences. We can look at integration or success of sentences, how easy is it to integrate one sentence into a representation given you just read something else. We can look at recognition, that is prior sentence information, how much time it takes to do that. This is called priming and I've done a lot of applications of that. We can look at generation of inferences to fill in gaps. In other words, I take a little story like Cathy came into Chicago, she went to work. One day she passed out and they took her to the hospital. Well, how do you go from her going to work to her passing out?
So, I have a study right now which I'm planning to do in which people generate inferences that fill in the gap. And I can model that process, as well, of how many steps it takes from stories that have been written already in literature for that particular occasion and it turns out there's about two or three steps and you can get there. So, we can actually simulate that kind of information, recognition. Okay, fine, let's move on.
Finally, one of the most interesting applications that I've been looking at recently is decision making. There's a phenomenon known as hindsight bias. That is, I give you information, I ask you to make a prediction and then I tell you the outcome and I say, okay, disregard that outcome, make a new prediction. And it turns out people always are biased towards what they knew and what they were told.
This model is a great model for updating your information and accounting for hindsight bias. And I'm working on a project where I have seven different stories, I'm sorry, seven different versions of one story, there are actually two different kinds. These come from Reed Hasting and a group in Colorado and one of them is on gold mining where three professors are about to go into a venture of mining gold and they have a series of attempts and then finally either succeed or fail and this is being manipulated.
Basically, the idea is if you update information it's very hard to disregard it. Okay? In other words, once you learn something and you integrate it into what you know it's going to play a role and influence your decision making.
On counterfactual thinking, this is a question of accessing alternative causes. So, I can tell you a story about Mr. Jones by Conaman Traversky, who leaves his office one day, drives home and is killed and his family members are sitting around wondering, once they find out about this they're thinking if only he had taken a different route, if only he hadn't stopped at the intersection, if only the boy who hit him wasn't on drugs and so on.
I can account for these data by looking at the accessibility of alternative causes. So, if you know that he could have gone home by the lakeshore rather than following the main highway, if you know that, then it makes that kind of reasoning much more easy to do. In other words, if he had only gone home by the lakeshore he wouldn't have been killed. Okay.
All right. So, that's my shopping list of studies that I'm working on and I hope that you see it as a wide range of convergent validity with respect to methods of studying comprehension. These are all typical kinds of procedures used. Some people don't recognize them as such like the counterfactual reasoning or the hindsight, but they are other ways of looking at comprehension.
Let's talk about some of the implications here. So, the first implication, a general one, is I think we actually ought to teach and foster the use of naive theories of goal directed action. In other words, the kind of theories that we're using to actually do our analysis is widely available, but not necessarily explicit.
So, you can use them to understand narratives or personal stories, fiction, biography and history as I pointed out earlier and I'll give you some techniques on that in a minute. And then in those situations wherever agency is central, literature, political science, history or social studies, you can think of countries as agents, okay? The United States, okay, attacks somebody, Afghanistan. So, the agent, in other words the United States is acting like an agent, but they're doing behavior. So, that behavior can be understood in terms of goals.
You can put down the goals and motivation behind the North and the South in the Civil War. There are lots of opportunities to use this kind of theory to specify conditions that you want the students to learn. Okay, next.
So, basically, we're going to be interested in the how and why of human behavior and this shows up in a number of different places. Explanation based teaching is extremely important. In my review, Mike Presley and Gerald Duffy were using explanation based teaching and teaching strategies. In other words, you want to give the students a reason why they're doing what they're doing and that's one way of thinking about it but, also, as Mickey Chee and others have worked on the science, the more explanation based understanding you promote, the better. It leads to better memory, it leads to better understanding, of course, and those two are going together.
So, I would be an advocate, a strong advocate of any kind of approach in teaching that's going to provide students with explanations as to why they're doing what they're doing, but even in the context of what they're doing, explaining what's going on. This fits in with John Dewey's statement about meaning.
With respect to text, I think we ought to be, we ought to expend a lot of effort. In 1973 I was a Chairman at that time of a Subcommittee on Comprehension. I'm trying to remember what the agency was back then.
DR. SWEET: This one? National Institutes of Education.
DR. TRABASSO: Maybe. Yeah, it was, NIE, right. So, we made a lot of recommendations about the study of text and we were, I think, quite successful in helping to get that agenda into being. And there was an incredible amount of work done in text in the '70s and '80s and there still are some very good people who are still actively doing discourse work. And one of the implications of a lot of that work, it's not just causally coherent, I think coherent text would be a better way to argue for it. That is text which are written are poorly done. They don't contain the conditions that you need to know in order to understand something.
So, there is a message here and it's unfortunate, but a lot of these texts that we analyze were done by commercial publishers who were not open to this. But it seems to me that there should be some interaction between those people who write texts and those people who know something about the discourse to help improve comprehensibility of text.
We can have explicit instruction, which is one of the things we did look at in the Reading Panel, on causal reasoning, why and how in asking oneself questions. One of the best predictors of success in the instruction comprehension is the generation of questions, question generation, asking why and how and answering it.
Now, when you think about procedures that are actually used in some of these methods of instruction, for example in the reciprocal teaching method, the teacher demonstrates by thinking out loud, asking questions, generating predictions, indicating, well, I don't understand this word so I'm going to have to look it up or I'm going to have to go back and see if I can find the meaning for it and so on.
So, if you think about the procedures that you actually used that way instructionally, I would advocate the use of think out loud procedures as a way of finding out what the student knows as they're reading. And you can interrupt and say, okay, or give them, you know, you're training them to do this, it's not hard as you saw in the evidence I gave you with respect to nine year olds and that's about the right age. I think you want to start doing this around third grade.
So, think out loud procedures are a very good way of finding out what the reader knows at this point in time and just say, tell me how you understand this, tell me what you think is going on here. That's basically it, not much more than that.
Counterfactual reasoning is interesting especially with respect to historical text. And Brink and Chuck Perfetti and others have actually used this technique where you read a historical passage and you say, well, what if Theodore Roosevelt did not send that letter to the Ambassador in Panama, what would have happened? That's a counterfactual. But what it does, it forces the student to think about alternative possibilities and to generate reasons based upon what they understand that were historically present. And they have to then undo things, which is not easy to do, but it's a very nice way of teaching thinking during understanding.
Now, I want to end, I think, no, a little more, that's it, that's it. Okay, I want to end with one other point. One of the things I was impressed with when we did our review with respect to instruction in comprehension is that all of these techniques which are done to teach reading comprehension are, in fact, done to teach content areas.
And I would like to advocate, I think, the use of strategic comprehension instruction in content areas as well as teaching in isolation or teaching it as a reading problem. I wouldn't want to teach it as a reading problem, I'd rather have it done in situations where they're reading anyway.
If you think about what goes on in content areas, what teachers are doing are providing a lot of knowledge for the student so they can learn to read, or read, I'm sorry, so they can learn from what they're reading. And so, it's like when you learn to use a manual. Like I learned to use PowerPoint, I learned PowerPoint in the last six months or so. I was much better off to read the manual on PowerPoint after I had done it, okay, and worked with the program itself than I was reading it from scratch. Why was that? Because I had a lot of knowledge against which I could now understand the text.
But I think that's, the content area is something which I hope that you'll put out and at least you encourage with respect to teaching reading comprehension; that is look at the procedures and techniques that are going to improve understanding in course content, as well, rather than isolation. Thank you. Okay, questions?
We can open up more generally. I'm not sure what your agenda was, but I think there was some announcements that we would talk about more than just my talk, okay, although I'm willing to talk about my talk. Yes?
SPEAKER: I'm just curious what...
DR. SWEET: You need to go to the mike. You've got to go to the mike.
SPEAKER: Oh, I'm sorry, that's right. I'm just curious what approaches you feel this discredits, or to be a little more polite, compared to what?
DR. TRABASSO: I haven't thought about it that way and I'm not sure I can answer your question directly. This is a view of the reader as a very active participant in the process. This is a view of the reader that emphasizes the active use of what they know and also the building of what they know.
There is, and I wouldn't say I'm opposed to this because I think it plays an important role and I don't think it discredits it, but there is also a dominant view which I have to deal with all the time and I get some tough reviews because of this or I get people that simply disregard what I do, and that is that a lot of learning is just simply passive. You get a lot of things thrown at you, you build up these associations and then we just do a latent semantic analysis and that captures it and you don't need to do all this analysis that you're doing on the text.
So, there's this kind of passive view which is manifested in something like latent semantic analysis, but it's also manifested in people like Macooner Makiff who looks at the reader as a lazy person, basically text washes over them and somehow or other gets associated, okay?
So, I guess those are my pet peeves with respect to, I mean, that's attention, it's attention. I don't think this discredits them because there are passive processes going on and there's a lot of passive activation that I don't account for. I'm working at the sentence or clause level. I'm working at the level where there's a definite state or an action, okay? So, Tom is hungry, not yet, but Tom is hungry would be a state. Tom is relieved, Tom was tense, but now Tom is relieved would be a state change. Tom had a goal of trying to give a decent presentation at the urging of friends of his and worked hard on his presentation and those are actions, okay?
So, in other words, I'm working at the level of state actions and state changes. So, clauses are natural. One predicate and a bunch of other things going with it, okay, one main predicate. On the other hand there are people who work at lower levels than I do, word level, other levels. That's clearly going on.
So, I don't want to say that my data are against them. They're not because my theory is very incomplete, there's so much else i don't capture. What I'm able to do quite well is to say, look, people are going to look for reasons why something is happening, can we identify those reasons? Yep, we can and we do a pretty good job. That's basically the bottom line.
SPEAKER: I have a question. What if a Member of Congress were to happen to ask you. why do we need to know all of this theoretical stuff about the nature of comprehension? I mean, what matters is that we do what's effective, what matters is what works. What value added is explanatory or predictive research to that bottom line notion of what works or not?
DR. TRABASSO: Well, getting to what works requires having this theory to begin with. So, in other words, the theory, what the theory tries to do is make explicit what you think are the processes that go on while people are reading and trying to understand. And once you do that, there are a whole bunch of implications that follow.
What works is actually teaching kids to understand through explanation or getting them to explain themselves. We already have that partly in practice because people ask why questions all the time, why and how, but it seems to me that we now have very strong evidence supporting the idea that this is a good way to go. So, it's definitely evidence based, not scientifically based. It doesn't take away from the fact this is going on, it just gives us a better understanding of them and I think it's going to, it gives you a basis for actually deciding where you're going to answer those questions. You wouldn't know otherwise, if you hadn't done some analysis of the text a little bit. What's appropriate, what's the appropriate question to ask, where do you ask it and when are you going to ask it, who are you going to ask it of, those kinds of questions.
Some of these techniques also are not standard in the literature on assessing comprehension. So, for example, like thinking out loud was a practice, but it was embedded in a larger framework. If you look at what's going on in those situations you see people doing exactly what I said they're doing, they're associating, they're explaining and so on. It gives us better understanding of our teaching procedures than we would have had otherwise.
And there are further implications from all this analysis that need to be explored. As I pointed out there are other applications to help us understand other phenomena. This is a rather unifying theory. This cuts through a lot of stuff and puts together, you know, a lot of things that people would not have seen, relationships have been forged. I think it's one of the goals of science and education, that is to get some kind of coherence with respect to, and a wider range of things as to how we understand them. That's some of my answers, but there, I mean, there are people who are actively using these techniques to teach teachers teach better. So, it's not like I discovered something entirely new.
SPEAKER: I'm just wondering about the kids that you used, did you have somebody tell you that they could read or that they couldn't or did you find them...
DR. TRABASSO: These were third graders and so they were...
SPEAKER: But did you find some children who didn't do these things?
DR. TRABASSO: We didn't select that way. I mean, one of the things that's really under-researched, I think, is the interaction between, it's an old idea, it's not new, but it's the treatment by, you know, with the interaction; that is, which kind of procedures are most appropriate for which level of reader and so we didn't do that.
We did look at individual differences, however. We found that there were some children who made no inferences at all and they didn't have a very good memory, either. We had other children who made lots of inferences and they had really good memories. So, that's what I showed you evidence for. So, as an assessment procedure we had a pretty good assessment, but we didn't select readers that way.
Now, I'm working with Jack Stenner who does work on lexiles and I'm working on a project with him on analyzing the goals of NAP and other national assessment goals and so on in terms of standards for what a reader can do. And one of his claims, and I endorse this actually, is a confound between the reader ability and the text they interact with.
So, we have to have a good way of assessing text that's independent of reader and his lexile method is one of those techniques. Actually, I used his procedure in some experiments and I got a little bit enamored with it. So, we are actually going through the criteria of different kinds of readers and if you break the criteria down it turns out there's statements about what the reader of this level should do.
So, the reader ability, what that breaks down in terms of the kind of text and what you do with the text and it's totally confounded with what people think are your good readers and bad readers or if they're poor readers.
So, hopefully, one of the outcomes in the future is this business of recognizing the contribution of text and choosing text which are appropriate for your readers at different levels of ability because otherwise you're going to frustrate some kids because the text, itself, is simply too difficult and beyond them.
And one of the criticisms of the research that we looked at was that it would take poor readers and give them more advanced text and train them on that, but why not train them on text that's more appropriate so that they learn the strategies and then transfer it to the more difficult text? So, the model is one where the text, itself, is something you're up against. It's the reader against the text and if the reader can do well on this level of text, then that's where that reader is at that point in time, the criteria of performance success.
So, if you want to have success in maybe teaching certain kinds of reading strategies, then I think you want to work with text that are accessible to these kids rather than ones which are not accessible would be one of the messages. So, lots of, you know, in other words, there's a whole set of implications that your question would raise with respect to individual differences in reading ability. And there are assessment issues, but there are also going be, then, training issues.
And people who did a lot of work on strategies and teaching comprehension worked with those who are regarded as the low average readers or poor readers. We did not adjust that, per se, although we did find evidence of both kinds.
SPEAKER: I have two questions. The first one directly links to what you were just talking about, it's a technical question. You were working with third graders...
DR. TRABASSO: Yes.
SPEAKER: ...and they read coherent and less coherent text.
DR. TRABASSO: Right.
SPEAKER: How did you measure comprehension recall, is that what I saw?
DR. TRABASSO: We had two, we had two think out loud protocols which we analyzed in some detail sentence by sentence.
SPEAKER: Did you do a recall after that?
DR. TRABASSO: Yes.
SPEAKER: Okay. So, you had...
DR. TRABASSO: And we also, yeah, yeah. The recall and we did a, we had an intervention design. So, we had like, you know, you read, we intervene, get you to think out loud and then we give you a transfer and you read again by yourself. And it turns out our intervention facilitated comprehension, as well, as measured by recall and as measured by the fidelity of the treatment. That is, did they actually do a lot more of the stuff? Yeah, they did a lot more...
SPEAKER: Self-explanation?
DR. TRABASSO: Right.
SPEAKER: Yeah.
DR. TRABASSO: So, you could think of this as a kind of, I mean, that particular study could be written up largely as an intervention study where you actually used those tough notes and I, you know, it's something I like to do sometimes.
SPEAKER: So, there you're working with causal coherence, right?
DR. TRABASSO: Yes, that's right. I haven't done it with other forms of coherence, although those stories are coherent, they're cohesive, locally cohesive, as well.
SPEAKER: Okay. So, that's what I was going to ask is...
DR. TRABASSO: So, that's sort of like, that marches together. Yeah, one of the things you want to think about, and you know this as well as I do, I'm not going to be teaching you anything new, but I think for this audience is that when we deal with text we have a lot of things working together, a lot of clear variation in the text. And so it's hard to separate all these things out. That's the reason why we did multiple regression with respect to text because of this complicated natural occurring thing.
But there are many ways in which you can assist them to understand and, again, you know, yourself as a good example of a person who is very sensitive to local cohesion; that is carrying an idea over or carrying over a character or carrying over a concept so that there's links between sentences in some way so that the student knows what's being maintained in some way in memory.
And that is a different level of analysis than what I do. I don't look for common arguments or argument overlap or common analysis. I've done comparisons. Generally speaking, if you have linear text and any kind of argument of any analysis of argument overlap it's going to be almost perfectly correlated with causality. However, if you're getting into those big networks then causality is going to win out.
SPEAKER: Well, you certainly guessed my next question so I'll continue to answer to the question that I just didn't ask. So, if argument overlap can't account for causal coherence when you get into the larger networks...
DR. TRABASSO: That's right.
SPEAKER: ...which...
DR. TRABASSO: It's more complicated.
SPEAKER: ...which is what LSA solely accounts for...
DR. TRABASSO: Which one?
SPEAKER: Which one, which LSA?
DR. TRABASSO: LSA accounts for a lot of...
SPEAKER: Oh, just semantic coherence.
DR. TRABASSO: Right, but it's...
SPEAKER: Latent semantic analysis.
SPEAKER: Latent semantic analysis is basically an analysis where you compare from one sentence to the next whether or not they are, the semantic coherence between them and also between that sentence and other...
DR. TRABASSO: Is it really only in terms of successive sentences? I thought it was...
SPEAKER: No, I hadn't finished. And also between that sentence and the paragraph and then that sentence and the whole text and then you average them together. DR. TRABASSO: Sure, sure. Right.
SPEAKER: And it doesn't do a very good job when you're comparing different texts like George's text on the psalmodists and a history text, it doesn't account for differences, it only works when you're looking at texts that are coherent from one person to the next, but it is just semantic coherence?
DR. TRABASSO: Well, it's word, word relationships of some kind, right?
SPEAKER: Right.
DR. TRABASSO: So, you can call that semantic if you want to, it's okay, but don't forget, it's just basically knowing that, if I know something about the heart and I know that there are veins and arteries so I've got some kind of cosign that's going to relate to veins and arteries, right?
SPEAKER: Right.
DR. TRABASSO: Okay. So, that's what it does, but I don't think it gets at the use of the text the way in which I'm talking about; that is, answering questions for example. It does predict something about memory, right? But I don't think it can deal with things like, well, explain to me why the heart does what it does. If you analyze that you can get the words out that maybe go together, but what constitutes a good explanation? I don't think that particular analysis will do a good job.
SPEAKER: Well, it might predict when an explanation would be necessary in a text, but that...
DR. TRABASSO: Sure, yeah.
SPEAKER: What my question is, is the LSA is certainly weak in terms of doing that, but what do you think could we capture in a text that would allow us to capture the qualities that, what can we say, how can we look at a text and say automatically, well, that's more causally coherent. Are there any indicators in the text that we can use for causal coherence?
DR. TRABASSO: Well, causals are not necessarily marked explicitly because a causal, a causal event, you're going to have temporal succession. So, if A is the cause of B, A will be causally prior to B, but they're not necessarily marked explicitly. That's one of the problems with a theory that wants to go for explicit marking.
The only time you get causes marked is when you reverse things. You say, you know, Tom went to the store because he was hungry. Now, I could say Tom was hungry and so he went to the store, okay? So, in that case you're not going to get an explicit marking. I don't think there's any easy way to do that analysis.
Now, there are, we're working on that problem, we're doing, we have a narrative coding system that we're working on which tries to identify causal relationships, but for the most part it's a judgment call. So, that's a real problem. So, this technique doesn't export well because it requires so much work in terms of doing the analysis, but it wouldn't be very hard, I think, to, you know, if you're going to give teacher's assistance you have to identify, you know, these are certain conditions that are necessary for this, make sure you ask questions about these things.
But I'm not disagreeing with, you know, in saying that, but I think it's good for looking at word to word relationships and maybe necessary vocabulary for this text against the body of text that happens to be on the same subject and I also believe, you know, that there is all these paths of processes of learning associations that go on that we have to take advantage of.
So, like for example, one of the benefits of reading a lot is just that, the more you read, the more these kinds of associations, you're going to have the kind of base that LSA measures. So, a lot of exposure to text, but I'm arguing for a much more active intervention, much more, or getting readers to be much more active themselves than they would be otherwise. Okay? I guess that's it.
DR. SWEET: So, again thanks to Tom and thanks to all of you for making the effort to come today and the OERI staff and I will be around after we conclude so that if there are any individual questions you wish to have answered, we'll do our very best. Thanks again.
(WHEREUPON, the Program for Research on Reading Comprehension Conference was concluded at 4:01 p.m.)
Pre-Application Meeting Videos and Transcripts
This page last modified May 28, 2002 (jer)