Although research has revealed ways of improving the accuracy of metacognitive monitoring, most of this research has involved monitoring of learning in associative learning tasks (e.g. learning paired associates). For instance, Nelson and Dunlosky (1991) found that delaying judgments of learning produced high levels of monitoring accuracy. Shaughnessy and Zechmeister (1992) showed that practice tests produced a substantial improvement in monitoring accuracy in associative learning tasks. However, many of the factors that have improved monitoring accuracy in asso- ciative learning tasks have not produced the same effects on accuracy of comprehension monitoring (called metacom- prehension accuracy). For instance, Maki (1998a) failed to replicate the delayed judgment of learning effect for a reading task. Furthermore, Maki and Serra (1992) showed that practice tests had only a modest effect on metacompre- hension accuracy. Thus, more research is needed to identify techniques to improve metacomprehension accuracy.
One method used to measure metacomprehension accu- racy is to have participants read several texts and then judge how well they will perform on a test over the content of each one. After all texts have been read and judged, par- ticipants take a test of comprehension for each text. Meta- comprehension accuracy is then estimated by correlating each participant’s judgments with his or her own test per- formance. Higher correlations indicate that the person can more accurately discriminate between texts that are more versus less understood.
Although metacomprehension accuracy is often quite poor (for a review, see Maki, 1998b), several techniques have recently been discovered that can substantially improve accuracy. Namely, accuracy is better (a) after rereading than after a single reading (e.g. see Dunlosky & Rawson, 2005, for caveats and review), (b) when keywords are generated for each text, but only if generation is delayed after reading (Thiede, Dunlosky, Griffin, & Wiley, 2005), and likewise, (c) when summaries are written, but only after a delay (Thiede & Anderson, 2003). In the latter case of summarization, which is the focus of the current research, metacomprehension accuracy reached a level of +.57 when summaries were written after all texts had been read (delayed-summarization) but was only +.26 when summaries were written immediately after each text had been read (immediate summarization).
Thiede and Anderson (2003) established summarization as a technique for improving metacomprehension accu- racy. Moreover, they speculated as to the possible cause of the delayed-summarization effect, using construction- integration model of text comprehension (Kintsch, 1988). According to this model, while reading, readers construct several levels of representation for a text: a lexical or sur- face level, a textbase level, and a situation model level. The lexical level contains the surface features of the text. The textbase level is constructed as segments of the surface text are parsed into propositions, and as links between text propositions are formed based on argument overlap and other text-explicit factors. Deeper understanding of the text
is constructed at the level of the situation model, which involves connecting text information with the reader’s prior knowledge. Thiede and Anderson (2003) suggested that writing summaries after a delay focused readers on their situation model, and this produced the boost in accu- racy (see also Rawson, Dunlosky, & Thiede, 2000 for a similar explanation of why rereading improved metacom- prehension accuracy). However, it is important to note that Thiede and Anderson (2003) did not evaluate possible hypotheses proposed to explain the delayed-summarization effect – in large part because their between-subjects design was insensitive to difference in summary length and quality. A primary purpose of the present investigation was to show the delayed-summarization effect using a within-subjects design and, more important theoretically, then conduct content analyses of summaries to evaluate two hypotheses proposed to explain the effect.
One hypothesis proposed to account for this delayed-sum- marization effect is the accessibility hypothesis, which states that metacognitive judgments are based on the amount of information accessed from memory (Baker & Dunlosky, 2006; Benjamin, Bjork, & Schwartz, 1998; Koriat, 1993; Morris, 1990; Serra & Dunlosky, 2005). According to this hypothesis, accessibility of information at the time of sum- marization provides cues to be used to judge subsequent test performance. For instance, if a person can retrieve a lot about a text, he or she will judge the text as better under- stood; whereas, if the person can retrieve very little about a text, he or she will judge the text as less understood. It is important to note that according to this hypothesis, judg- ments are based on the amount retrieved from memory and not the content of that which is retrieved (i.e. details or gist).
When summaries are written immediately after reading, information about a text has not had a chance to fade from working memory; therefore, a person has access to infor- mation for writing a summary even for texts that were not well understood. As a result, summaries of less under- stood and more understood texts are more similar when written immediately after reading than after a delay. There- fore, immediate summarization may produce a more homogeneous set of cues, which makes it more difficult to discriminate between less understood and more under- stood texts, thereby reducing metacomprehension accu- racy. By contrast, when writing summaries after a delay, information has had a chance to fade from memory; there- fore, the person has access to relatively little for texts that were not understood and more for texts that were under- stood. As a result, delayed-summarization may produce a less homogeneous set of cues, which makes it easier to dis- criminate between less understood and more understood texts, thereby improving metacomprehension accuracy.
A key prediction derived from the accessibility hypothe- sis is that metacomprehension judgments will correlate with the total number of idea units contained in summaries. As the amount produced is more homogeneous in immediate summaries than delayed summaries, it follows that the cor- relation between metacomprehension judgments and the
112 M.C.M. Anderson, K.W. Thiede / Acta Psychologica 128 (2008) 110–118
total number of idea units contained in summaries will be weaker for summaries written immediately after reading than after a delay – due to attenuation of variance in amount produced by an individual. Given that the total number of ideas is correlated with test performance, this would account for the difference in metacomprehension accuracy for immediate and delayed-summarization conditions.
An alternative to the accessibility hypothesis is the situ- ation model hypothesis, which states that metacomprehen- sion accuracy will increase as the judgment of comprehension for a text is based on information relevant to one’s situation model for the text. According to the sit- uation model hypothesis, because an appropriate situation model is important for test performance (McNamara, Kin- tsch, Songer, & Kintsch, 1996), metacomprehension judg- ments based on information relevant to the situation model for a text will support higher levels of metacompre- hension accuracy.
The situation model hypothesis has been proposed to account for the findings of a number of studies. In particu- lar, Rawson et al. (2000) suggested that metacomprehension accuracy was greater for participants who read texts twice (versus once) because those participants who were given an opportunity to reread could allocate resources not only to developing a better situation model but also to monitor- ing their situation model. This produced the boost in accu- racy. The situation model hypothesis has also been proposed to account for boost in accuracy associated with generating summaries (Thiede & Anderson, 2003) and key- words (Thiede et al., 2005) prior to judging comprehension.
In contrast to the accessibility hypothesis, this hypothe- sis suggests that accuracy is driven less by the total number of ideas in the summaries and more by the number of ideas produced that are relevant to the situation model. A pre- diction derived from this hypothesis is that the correlation between metacomprehension judgments and the number of gist idea units will be weaker for summaries written imme- diately after reading than after a delay. Given that the number of gist ideas produced is correlated with test performance, this would account for the difference in metacomprehension accuracy for immediate and delayed- summarization conditions.
It is important to note that the situation model hypoth- esis makes predictions specifically about metacomprehen- sion accuracy (versus metamemory, which addresses metacognitions related to memory of the details contained in texts, see Wiley, Griffin, & Thiede, 2005, for a discussion of the difference between metacomprehension and metame- mory for text learning). Thus, in the present experiment, we assessed comprehension in a way consistent with compre- hension research (e.g. Graesser & Bertus, 1998; Royer, Carlo, Dufresne, & Mestre, 1996), namely, using inference questions. This practice is consistent with the early research in metacomprehension – Glenberg and Epstein (1985) and Weaver (1990) used only inference tests in order to tap comprehension as defined by comprehension theory.