Read carefully, because this is important! How value-driven strategies impact sentence memory

Abstract

Little is understood about how people strategically process and remember important but complex information, such as sentences. In the current study, we asked whether people can effectively prioritize memory for sentences as a function of their relative importance (operationalized as a reward point value) and whether they do so, in part, by changing their sentence processing strategies when value information is available in advance. We adapted the value-directed remembering paradigm (Castel, Psychol Learn Motiv 48:225–270, 2007) for sentences that varied in constraint and predictability. Each sentence was associated with a high or low value for subsequent free recall (whole sentence) and recognition (sentence-final words) tests. Value information appeared after or before each sentence as a between-subject manipulation. Regardless of condition, we observed that high-value sentences were recalled more often than low-value sentences, showing that people can strategically prioritize their encoding of sentences. However, memory patterns differed depending on when value information was available. Recall for high-value sentences that ended unexpectedly (and therefore violated one’s predictions) was reduced in the Before compared to the After condition. Before condition participants also showed a greater tendency to false alarm to lures (words that were the predicted – but not obtained – ending) from strongly constraining sentences. These observations suggest that when people try to prioritize sentence-level information that they know is valuable, the reading strategies they employ may paradoxically lead to worse memory.

A string of 12 random numbers is not easy to remember – but people would try very hard if they were told it was the next lottery number. This ability to exert strategic control over memory for important information is a critical life skill and has been studied extensively through the value directed remembering (VDR) paradigm (Castel, 2007; Castel et al., 2002). In this paradigm, participants read a list of study material, commonly a list of unrelated words, and each item on the list is associated with a point value indicating the number of points earned for successful recall. Because participants are given more words than is feasible to remember, they benefit from prioritizing higher value information – i.e., words paired with higher point values. While an emerging body of work suggests that there are some dopamine-driven impacts of value information that accrue relatively automatically (see Knowlton & Castel 2022, for a review), this paper focuses on how metacognitive control of strategy engagement may impact downstream memory. Through variations of the VDR paradigm, studies have found that people, both young and old, exert control to prioritize important information as measured by improved recall performance for high value items (Castel et al., 2002, 2007). People do this by not only strategically allocating more time to study the higher value information (Ariel et al., 2009; Castel et al., 2013) but also by selectively engaging different kinds of rehearsal strategies, such as using imagery mediators, elaboration (e.g., generating sentences), or the engagement of relational processing (Cohen et al., 2014; Ariel et al., 2015); indeed, the memory boost for high value items often disappears when participants are made to engage in the same rehearsal strategy for all items (Hennessee et al., 2019; Festini et al., 2013). People are also found to be strategic during retrieval operations, taking advantage of the primary or recency effect when retrieving high-value words (Murphy & Castel, 2022; Stefanidi et al., 2018).

In daily life, the task of retaining important information is further complicated by the fact that the kinds of things that people need to try to remember are rarely just sets of random numbers or disconnected words. More commonly – as, for example, in the case of a student reading course materials or a patient getting instructions from a doctor – people are trying to encode more complex concepts or scenarios, which are then coming into the system via more complex stimuli, like texts and conversations. It is not clear that the types of memory strategies that people can use to prioritize relatively simple information will generalize effectively even to the level of single sentences. Sentences differ more substantially than do individual words in factors such as their length, level of complexity, demands on relational or inferential processing, and so forth. As such, sentence recall is also rarely verbatim and is more likely to be based on conceptual representations (Potter & Lombardi, 1990), even when semantic support is minimal (Alloway, 2007). For example, people are more likely to falsely recognize a synonym of a word that was studied in a sentence (e.g., cemetery after studying graveyard) than a new compound word created from surface level parts of words that had been encountered in those sentences (e.g., blackbird after reading blackmail and jailbird) – a pattern that is reversed when participants study single words in a list (Matzen & Benjamin, 2009, 2013). The information carried by a sentence is also coming in over time, meaning that people may have to choose their processing strategies before they have critical information about memory-relevant properties of the input.

Indeed, even for simple stimuli, the relationship between the timing of value information and its impact on memory is complex. In contrast to some findings that young adults successfully recall high-value single words when the value cue is presented after the study item (Hayes et al., 2013), studies that directly manipulate when value information is available have found that it is only or primarily when value information is presented before the to-be-encoded item that it has an impact on cued recall (Soderstrom & McCabe, 2011) and recognition (Villaseñor et al., 2021) of single words. Similarly, Leippe et al. (1978) found that participants were better at identifying a confederate-thief who ‘stole’ an item of high monetary value when they were told about the value of the item before the theft compared to after the theft. Findings like these suggest that the timing of when value information is encountered may impact how successfully or efficiently people allocate resources to strategize and process even simple information. The impact of timing on how value information can be used is likely to be amplified when people process more complex information like sentences that unfold dynamically and can vary in how predictable they may be. Moreover, whereas studies with simple stimuli have emphasized the utility of having value information in advance, the literature on language processing (reviewed below) raises critical questions about the memory consequences of the kind of more “active” processing strategies people may adopt when they anticipate that an upcoming sentence will be important. Therefore, the present study aims to address these important questions: Can the kinds of value-based effects that have been documented for simple stimuli also be observed for complex ones like sentences? And, if so, how does the timing of the value cue (i.e., whether provided before or after the information) affect these patterns, and what can that tell us about the strategies people may be using to prioritize information in the context of sentence processing?

A large literature has demonstrated that people process sentences incrementally and use multiple strategies to do so (see review in Federmeier 2022). Information from a sentence naturally accrues, as evidenced by the N400 word position effect, in which the amplitude of the N400, an ERP component linked to semantic access, is gradually reduced as sentential context builds (Van Petten & Kutas, 1990, 1991). In addition to this passive accumulation of information, individuals can engage more active mechanisms during sentence processing, including the use of prediction to pre-activate both meaning- and form-based information about likely upcoming words, concepts, and language structures (e.g., Altmann & Kamide 1999; Federmeier & Kutas 1999; Wicha et al. 2004; Van Berkum et al. 2005; DeLong et al. 2005).

As reviewed in more detail below, the tendency to engage prediction has been shown to vary across individuals, tasks, and items – i.e., the use of prediction seems to be at least partially strategic. Moreover, a growing body of literature attests that prediction not only affects the ease and efficacy of in-the-moment comprehension but also has important consequences for down-stream memory. Given that people can adopt different modes when comprehending, and that this matters for memory, a key question is how people might adapt their sentence processing strategies when trying to prioritize information and whether they are able to effectively choose strategies that promote accurate retention of the information that they value.

Individuals differ in their tendency and/or ability to engage in prediction during sentence processing. For example, predictive processing effects are reduced among healthy older adults (Wlotko et al., 2012; Wlotko & Federmeier, 2012) and adults of all ages with lower literacy (Ng et al. 2018; Steen-Baker et al. 2017). Even among highly literate young adults, for whom prediction effects are well-attested, the tendency to engage prediction is modulated by task demands (Brothers et al., 2017, 2019; Ness & Meltzer-Asscher, 2021; Wlotko & Federmeier, 2015). For example, across two experiments, Brothers et al. (2017) demonstrated the strategic nature of predictive processing by either encouraging prediction through explicit instruction to predict, or discouraging it by including many unpredictable items in the stimulus set. When participants were instructed to predict, they obtained more benefit from the sentence context, as measured by larger reductions in N400 amplitudes to expected words, compared to when they were simply reading for comprehension. When, instead, participants were discouraged from predicting, expectancy effects in self-paced reading times were correspondingly diminished. Critically, the tendency to engage prediction also varies at time scales smaller than the individual or the task. Examination of trial-by-trial variability in sentence processing by both younger and older adults (Payne & Federmeier, 2017; Jongman & Federmeier, 2022) has revealed that prediction is differentially engaged across items, even within the same individual and the same task.

It is not surprising that people may tend to predict when they can, as prediction has a number of well-documented benefits for online comprehension. Prediction may help to mitigate against noise or uncertainty in the input (see discussion in Kuperberg & Jaeger 2016). Even when incoming information is relatively clear and unambiguous, prediction has been shown to augment people’s ability to make use of the full range of available context information and to ease the processing of expected information at multiple levels (e.g., Wlotko et al. 2012; Frisson et al. 2005; Altmann & Mirković 2009). However, these positive effects of prediction on comprehension may come with some associated costs for downstream memory. Prediction violations may be recognized (Federmeier et al., 2007) and recalled (McFalls & Schwanenflugel, 2002) more often than confirmed predictions, suggesting some learning benefit (e.g., Dell & Chang 2014), whereas confirmed predictions may be processed through a comprehension system in ‘verification mode’ (Van Berkum, 2010), which results in more shallow encoding and a less robust representation of the predicted word (Rommers & Federmeier, 2018b). Additionally, disconfirmed predictions may not be fully suppressed in memory, introducing interference: Predicted but disconfirmed (and therefore never seen) sentence endings modulate the N400 as if the word had been encountered previously, resulting in a ‘pseudo-repetition effect’ (Rommers & Federmeier, 2018a). Hubbard et al. (2019) also found that recognition of these predicted-but-not-encountered lures elicited a higher false alarm rate compared to new words, further supporting the idea that predictive processing may have drawbacks for memory performance. One thing to note is that most of the extant work looking at the impact of prediction on memory has focused on participants’ memory of (usually sentence-final) individual words from different kinds of sentences. What impact prediction as a processing strategy may have on memory for the sentence as a whole is still unclear.

Current study

Building on prior research showing that people have some strategic control over the mechanisms they engage in as sentences unfold, including the possibility of predicting, and that predictive processing can have downstream consequences for memory, the current study adapts the VDR paradigm to examine whether varying the importance and predictability of the information they need to remember will impact people’s memory patterns for such information. Participants, whose goal is to maximize their score, read six sets of 16 sentences presented word-by-word. Sentences vary in how strongly they converge on a single likely ending (i.e., their constraint), as well as in whether they end with the most likely completion or, instead, with an unexpected but plausible completion. Each sentence is paired with a point value that indicates the number of points earned for successfully remembering the sentence (or part of it) on subsequent memory tests. After each study block, participants complete a free recall task, trying to remember as many sentences as possible as accurately as they can. In addition, after every third free-recall task, participants complete a recognition test. Critical items for the recognition test include old words, which were either predictable or unpredictable sentence endings, as well as new ‘lure’ words – words that were predictable in a given sentence frame but were never actually seen, because the sentence was completed with an unexpected ending instead.

The goals of the current study are two-fold. First, we want to confirm that the VDR paradigm can be used to measure memory performance when adapted with sentences, as to our knowledge the current study is the first attempt to examine how value impacts sentence memory. Therefore, in one condition (the “After condition”), we present the value cue after the last word of each sentence. As in single word studies using the VDR paradigm, participants in this condition can, once they know the value of the sentence they just read, use a range of strategies to try to prioritize higher valued information, including additional rehearsal and various types of elaboration. We predict that participants will be sensitive to value information and prioritize recalling high-value sentences, leading to value-based effects in recall performance. How predictability of the sentence ending will affect recall memory for sentences is less clear. Although some studies report superior memory for unexpected material (O’Brien & Myers, 1985; Corley et al., 2007), consistent with the idea that the error signal generated from unpredictable input may lead to learning (e.g., Dell & Chang 2014), others report that unexpected (Röer et al., 2019) or irrelevant (Foss & Cairns, 1970) words can disrupt serial recall and that grammatical but nonsensical sentences (Holmes & Murray, 1974) or identical sentence frames ending with a less preferred synonym (Brewer, 1975) are less easily remembered compared to predictable sentences.

Although recall performance is sensitive to value in single word studies using the VDR paradigm, value-based manipulations have been shown to have little to no impact on recognition memory measures (Castel et al., 2007; Wong et al., 2019). In the current study, our main interest in the recognition memory data focuses on how participants remember predictable words, unpredictable words, and lures as a function of sentence constraint. We expect recognition behavior in the After condition to replicate previously reported patterns from studies that similarly instructed participants to read sentences for a subsequent recognition memory test (Hubbard et al., 2019). That work found that participants falsely recognized predicted-but-not-seen lures more often than completely new words; however, the tendency to false alarm to the lures was not reliably modulated by sentence constraint. The recognition data from the After condition will also serve as a baseline to probe for possible changes in how people use value information depending on when that information is available to them, which is manipulated in our second condition as described next.

The second goal of the study is to examine the extent to which participants adapt their online comprehension strategies to try to prioritize information for later memory. To that end, in a second condition (the “Before condition”), we provide value information before each sentence begins. The two conditions are otherwise identical. Participants in the Before condition can thus employ all the same strategies as those in the After condition but, in addition, can potentially use value information to change how they take in the sentence information to begin with. We hypothesize that, when given information that an upcoming sentence is high value, participants may try to exert more top-down control, perhaps including increased engagement in prediction. Paradoxically, given the kinds of downstream memory effects that have been observed in conjunction with predictive processing, this may actually have negative consequences for (at least some aspects of) memory. Thus, although we expect participants in the Before condition, like their After condition counterparts, to show value-based performance and recall more high value sentences than low value sentences, we further predict that participants in the Before condition (compared to the After condition) may show worse memory for high value sentences in a manner that is modulated by predictability.

For recognition memory, we again do not expect to see value-based effects, so our focus is on the impact of constraint and predictability. If participants predict more in the Before condition compared to the After condition, we expect to see a greater luring effect, especially for the strongly constraining sentences that support stronger predictions. Additionally, increased prediction in the Before condition could result in reduced recognition for predictable words (Rommers & Federmeier, 2018b) and increased memory for unpredictable words (Dell & Chang, 2014), a contrast that again would be magnified in strong constraint sentences compared to weak constraint sentences.

Method

Participants

Participants were recruited through Amazon’s Mechanical Turk (MTurk), using MTurk’s native worker requirements to ensure that participants were located within the US and had an approval rate greater than 90%. Demographic information was not collected but we assume our participants to be on average 34 years old, based on recent studies that examined the demographics of large MTurk samples in the US (Casey et al., 2017; Burnham et al., 2018). Workers were asked to complete the task in one sitting in a quiet environment with a desktop or laptop computer and were paid $10 for attentive completion of the task. Our a priori sample size was set to 24 participants per condition, following the maximal sample size reported in Castel et al. (2002). Participants were recruited until a counterbalanced sample set of 48 was collected. One additional participant was recruited but not used due to the participant reporting technical issues during the experiment, resulting in 48 participants included in the final analysis.

Stimuli

Study Phase

The experimental stimuli consisted of 96 sentences that were a subset of the sentences used in Federmeier et al. (2007), with some sentences modified to ensure that recognition test items only appeared once throughout the entire study. All sentences were normed for cloze probability (refer to Federmeier et al. 2007, for details) and contained on average 9.76 words (range 4–19). Each sentence was completed with the word most expected in that sentence frame (Mean cloze probability = 0.55) or an unexpected but plausible ending that had a cloze probability close to 0 (Max cloze probability = 0.11).

Half of the sentences were drawn from a set designated as ‘strong constraint’ in which the cloze probability of the most commonly completed word was 0.68 or higher and the remaining half from a ‘weak constraint’ set in which the cloze probability was 0.42 or lower. The lexical characteristics of sentence endings are presented in Table 1 and example stimuli are provided in Table 2. Sentence endings averaged 5 to 6 letters in length and were fairly concrete, imageable and familiar, and of similar frequency range. Following the VDR paradigm (Castel et al., 2002), each sentence with each ending was associated with a high value (i.e., 10 points) or a low value (i.e., 3 points) across participants. Thus, for recall performance, the conditions were expected and unexpected sentences of high or low value. For recognition, we further divided the test items by constraint (strong/weak) to assess whether we replicated the luring effect reported in prior work (Hubbard et al., 2019).

Table 1.

Lexical properties of sentence endings

Category	Frequency	Concreteness	Imageability	Familiarity	Word Length
Expected	4.73 (4.79)	507.43 (107.07)	525.41 (37.87)	567.41 (103.09)	5.11 (1.38)
Unexpected	4.52 (4.48)	503.10 (117.41)	522.30 (45.95)	558.05 (91.44)	5.90 (2.01)

Note. Values represent means across items with standard deviations presented in parentheses. Frequency values are log transformed and obtained from the norms of Kucera and Francis (1967). Concreteness, imageability, familiarity values are obtained from the MRC Psycholinguistic database (Coltheart, 1981)

Table 2.

Example sentence stimuli and recognition test items

Sentence Frame	Expectedness	Ending	Cloze Probability	Recognition Test Item	Recognition Test Category
The little puppy grew up to be a huge	E	dog	97% [SC]	dog	Sentence Final
	U	responsibility	2.2%		Lure
While the national anthem is played, everyone is expected to	E	stand	76.4% [SC]	national	Sentence Medial
	U	behave	0%
She thought she had seen a	E	ghost	68.6% [SC]	department	New
	U	ship	0%
The officials tried to evacuate everyone from the area before the	E	explosion	41.2% [WC]	parade	New
	U	parade	0%		Sentence Final
The young boy asked for another	E	cookie	26.5% [WC]	young	Sentence Medial
	U	drink	8.8%
Seth was uneasy because he had never liked to	E	lie	11.4% [WC]	lie	Sentence Final
	U	ski	0%		Lure

Note. SC: strong constraint; WC: weak constraint; E: expected; U: unexpected. Across participants, a single sentence was paired with both a high and low value

Sentences were counterbalanced for expectedness and value, resulting in four unique lists. For each list, the position of the value cue was manipulated between subjects; half of the participants always saw the value cue before the pre-trial fixation cross (Before condition) and the remaining half of participants always saw the value cue after the last word of the sentence (After condition). In other words, across participants, the same sentence frame appeared with either its expected ending or unexpected ending and was paired with either a low value or high value that was presented before or after the sentence.

Recognition Test

The test stimuli were comprised of 192 words: (1) 48 Old sentence-final words (Sentence Final), (2) 48 Old mid-sentence words (Sentence Medial), (3) 72 New words that did not appear in any of the experimental sentences and (4) 24 Lure words, which were the expected (highest cloze probability) ending of a sentence frame that had, instead, ended unexpectedly (and thus these words were never viewed by the participant); see Table 2 for examples. Old test items appeared only once in the entire set of sentence stimuli. The lexical properties of the test items across the four categories are depicted in Table 3. Sentence Final, New (non-Lure) and Lure items, which were the critical items used in analyses, were fairly concrete, imageable, familiar and of similar frequency, although the average length of New items was slightly longer than Sentence Final or Lure items.

Table 3.

Lexical properties of recognition test items

Category	Frequency	Concreteness	Imageability	Familiarity	Word Length
Sentence Medial	4.68 (4.84)	404.33 (111.92)	450.55 (80.13)	542.25 (34.02)	6.46 (1.92)
Sentence Final	4.67 (4.60)	510.79 (112.79)	525.66 (97.96)	558.94 (41.07)	5.54 (1.77)
New	4.86 (5.10)	480.10 (121.83)	500.91 (101.28)	554.90 (40.20)	6.04 (1.94)
Lure	4.66 (4.71)	511.74 (107.60)	520.47 (106.44)	562.09 (36.22)	5.13 (1.45)

Note. Values represent means across items with standard deviations presented in parentheses. Frequency values are log transformed and obtained from the norms of Kucera and Francis (1967). Concreteness, imageability, familiarity values are obtained from the MRC Psycholinguistic database (Coltheart, 1981)

The same test items were used for all participants, such that a Lure item for one participant was a Sentence Final item for a different participant. For example, all participants were asked if they remembered reading the word “escape,” which would be a Sentence Final item for participants who read “The prisoners planned their escape” but a Lure item for those who read “The prisoners planned their party.” This ensured that, across participants, Lure and Sentence Final items were the same words and came from the same sentence frames. Half of the items were old and half were new for each participant, and the proportion of test items from each category was identical across participants: Sentence Medial (25%), Sentence Final (25%), New (37.5%), Lure (12.5%).

Experimental procedure

Once MTurk workers consented to participating in the study, they were redirected to a survey page hosted on Qualtrics. Participants were informed that they would be reading sentences that were paired with a numerical value and that the values reflected the importance of the sentences for subsequent memory tests. Participants were instructed that their job was to get as many points as possible by remembering the sentences they read to the best of their ability. In particular, they were told that they would read a total of six sets of 16 sentences and that they should try to remember the entire sentences as best as they could because, after each set, they would complete a memory test in which they would be asked to type out the sentences they just read and, after every three sets, they would complete a memory test in which they would be asked to mark whether they saw certain words appear in the sets of sentences they had just read. The instructions included a request not to use external aids to note-take during the study and emphasized that participants were not expected to remember all of the sentences, as the study’s interest was in what participants could remember. Once participants started the study, they were unable to return to a previous point. A progress bar at the top of the window was provided as a visualization of their progress. On average, participants took 45 min to complete the entire experiment.

Participants completed six study blocks, consisting of 16 experimental sentences presented per block, counterbalanced for constraint, expectedness and value. After each study block, participants completed a free recall test. Every third recall test was followed by a yes/no recognition test.

Study Phase

As depicted in Fig. 1, in a single study block, 16 sentences were presented one word at a time for 200 ms each with an interstimulus interval (ISI) of 300 ms. Each sentence was preceded by a fixation cross that was presented for 200 ms. Depending on the condition, a number value was presented for three seconds either before this fixation cross (Before condition) or after the last word in the sentence (After condition). The value was always accompanied by the words ‘Next sentence:’ or ‘Previous sentence:’ depending on what was appropriate to the condition. To aid discrimination, the value and the accompanying words appeared in orange (for low value sentences) or in blue (for high value sentences). Sentences were presented with a three second inter-sentence interval.

Fig. 1.

Example sequence of a single study block in the After (a) and Before (b) condition. Note. (a) An example sequence of a single study block in the After condition. (b) An example sequence of a single study block in the Before condition. For both conditions, participants studied a study block that consisted of 16 sentences and completed a free recall test. After every third study block, participants completed a free recall and yes/no recognition test

To prevent participants from interfering with stimuli presentation, the sentences were presented in a pre-recorded video with the control bar removed. Participants were not able to pause or replay the video and were given the free recall test immediately after the last sentence of the block was presented.

Free Recall

After each study block, participants were given a recall test (6 times total). Participants were asked to type out the sentences they just read in any order they could remember. They were encouraged to write down complete sentences but told that incomplete sentences or phrases were acceptable if they could not remember the whole sentence. Participants could take as long as they needed before proceeding to the next study block.

Recognition Test

After the third recall test, participants were given a recognition test (twice total). They were tested on the 48 sentences they had just read. Participants were shown one test item at a time in alphabetical order and asked whether or not they remembered seeing that word in any of the previous sentences they had read so far. They indicated their response by clicking either ‘Yes (seen word)’ or ‘No (new word).’ There were 96 questions in a single recognition test. Participants could take as long as they needed on each question. Participants’ recognition score (see below) was provided at the end of the recognition test, as well as encouragement to take a break as necessary before proceeding to the second half of the experiment.

Coding and scoring

Recall

Recall responses were coded offline by the first author and an independent coder who was blind to the study hypothesis. Intercoder reliability was 94.65%. Responses that the two coders disagreed on were resolved through discussion between the first and second author. Because such coding processes were necessary for recall data, participants were not given separate feedback regarding their recall performance during the experiment. For descriptive purposes, participants’ responses were categorized into six different categories depending on the quality of the response compared to the target sentence. ‘Verbatim’ referred to responses that matched the target sentence without any changes to content words. ‘Almost verbatim’ referred to responses that conveyed the main message of the study sentence but with changes to the surface structure. ‘Gist’ responses conveyed the main message but lacked detail compared to ‘Almost verbatim’ responses (e.g., a prepositional phrase might have been dropped). ‘Fragment’ responses only captured a part of the target sentence and missed the main message, and ‘Single words’ indicated cases in which participants could only recall a single, identifiable content word from the target sentence. Responses that could not be reasonably linked to any of the target sentences or conveyed an inaccurate message compared to the study sentence were marked as ‘Unknown/Incorrect.’ A breakdown of the response types for each condition can be found in Appendix. Because we were interested in assessing memory for the primary message conveyed by the sentence, the main analyses were conducted by aggregating across responses coded as Verbatim, Almost verbatim, and Gist (i.e., that represented accurate memory for the primary message conveyed by the sentence).

Recognition

During the experiment, participants were provided with a value-conditionalized recognition score, calculated using Qualtrics’ native ‘Scoring’ tool. Participants were given points corresponding to the associated sentence value for correct hits to old items (i.e., Sentence Final or Sentence Medial items). For example, a participant who correctly said ‘yes’ to a critical word that appeared in a sentence paired with a high value (i.e., 10) was awarded 10 points. For offline analyses of recognition data, we calculated the proportion of “Yes” responses to Sentence Final, Sentence Medial, New, and Lure Items.

Data Analysis

For statistical analysis, trial-level behavioral responses from recognition data (1 = judged ‘Seen’ and 0 = judged ‘Not seen’) and recall data (1 = recalled and 0 = not recalled) from the Before and After condition were each entered into separate mixed effects logistic regression models in R version 4.0.3. (R Core Team, 2021), using the lmer function package lme4 (Bates et al., 2014). The categorical predictors were contrast coded to allow factor coefficients to be interpreted as main effects. The random effects structure for all models was always the maximal random effects structure that converged, with random slopes that explained the least amount of variance being removed in cases of non-convergence (Barr et al., 2013). Correlations between random effects were not calculated to ease convergence. 95% confidence intervals for parameter estimates were computed using parametric bootstrapping (1000 iterations), using function confint(). The specific fixed and random effects structures for each model are reported in the Results section below.

Results

Recognition

Overall performance comparison between conditions

Figure 2 plots the proportion of “Yes” responses for Sentence Medial and Sentence Final items (for which “Yes” was a correct response) as well as New and Lure items (for which “Yes” was an incorrect response) for each condition. Comparing the hit rate for the critical Sentence Final items to New items, the average d’ for the After condition was 1.59 with a range of 0.14–2.89 whereas the average d’ for the Before condition was 1.46 with a range of 0.64–2.36. The average d’ did not differ across condition as revealed by a Welch’s t-test (Before: M = 1.46, SD = 0.51; After: M = 1.59, SD = 0.84), t(38.21) = −0.68, p = 0.50. On average, then, participants in both conditions were successful in discriminating previously seen items (i.e., Sentence Final) from never seen items (i.e. New) to a similar extent.

Fig. 2.

Recognition memory accuracy. Note. Proportion of “Yes” (seen word) responses by recognition test item type for the After condition (left) and Before condition (right). Error bars reflect standard error around the mean

False alarms

False alarms were broken down into three categories of New, Lures from a Strongly Constraining sentence frame (SC Lure) and Lures from a Weakly Constraining sentence frame (WC Lure) (Fig. 3). Trial-level behavioral responses were entered into a mixed effects logistic regression model. The model included condition (After, Before) and category (New, SC Lure, WC Lure) as fixed effects and random intercepts for participants and items but no by-participant or by-item slopes. This model for false alarms revealed a significant difference between New items and Lure items (β = −0.55, 95%CI [−0.86, −0.22) but no difference between SC and WC Lures (β = 0.43, 95%CI [−0.10, 0.91]. Additionally, the model showed a significant interaction between condition and lure type (SC Lure, WC Lure), β = −0.70, 95%CI [−1.25, −0.14]. A visual inspection of the interaction in Fig. 4 suggests that it arises from the Before condition participants being more likely to false alarm to SC Lures. No other predictors were significant as their 95%CI contained 0 as a plausible value. Thus, as in Hubbard et al. (2019), participants in both conditions lured to new items that were related to – and possibly could have been predicted in – sentence frames that had been studied, irrespective of sentential constraint. However, only participants in the Before condition showed an effect of sentence constraint, being more likely to falsely recognize a word they had not actually seen if the word was one that would have been highly predictable in a sentence they had read compared to one that would not have been as predictable in its sentence context. This enhanced sensitivity to sentential constraint is consistent with the hypothesis that participants in the Before condition are more likely to engage in prediction – and hence more likely to false alarm to a word that was likely to have been strongly predicted.

Fig. 3.

False alarms to unstudied items. Note. SC: Strong constraint. WC: Weak constraint. Proportion of “Yes” (seen word) responses for items that were not studied (i.e., false alarms) plotted for the After condition (left) and Before condition (right). Error bars reflect standard error around the mean

Fig. 4.

Two-way interaction between constraint and condition for false alarms. Note. SC: Strong constraint. WC: Weak constraint. Predicted probability of “Yes” (seen word) responses for items that were not studied (i.e., false alarms) plotted for the After condition (left) and Before condition (right). Error bars reflect standard error around the mean

Hits

To assess whether recognition memory was modulated by value, hit rates for Sentence Final items were compared by the value that had been assigned to the carrier sentence for each condition. In the After condition, the hit rate for words from low-value sentences did not significantly differ from those in high-value sentences, t(23) = 0.80, p = 0.43, suggesting that, as for memory for single words studied in lists (Castel et al., 2007; Wong et al., 2019), value does not have a large impact on recognition memory for single words studied in sentences. Similarly, the hit rate for words from low-value items also did not significantly differ from its high-value counterparts in the Before condition, t(23) < − 0.00, p = 1, suggesting that even when participants knew upcoming information was a low value item, they did not completely ignore it.

The next model assessed the hit rate of sentence-final words by modeling responses for Sentence Final items with condition (After, Before) constraint (strong constraint, weak constraint), expectedness (expected, unexpected) and the resulting interaction of the three as fixed effects. The random effects structure included a random intercept for participants and items. We found a significant interaction between constraint and expectedness, β = −0.53, 95%CI [−1.04, −0.05]. No other fixed effects were significant, as evidenced by confidence intervals that included zero as a possible value. This suggests that irrespective of when the value was presented, SC words compared to WC words differed in their rate of recognition depending on expectedness. As can be seen in Fig. 5, the interaction arose because, especially in the strong constraint sentences that rendered expected words more predictable, participants showed reduced encoding of expected items but enhanced memory for prediction violations. While a visual inspection of the three-way interaction between condition, constraint, and expectedness (see Fig. 6) suggests this interaction may be mostly driven by the Before condition compared to the After condition, the interaction was not found to be significant, β = 0.47, 95%CI [−0.28, 1.23].

Fig. 5.

Hits to studied items. Note. SC: Strong constraint. WC: Weak constraint. Proportion of “Yes” (seen word) responses for studied words (i.e., hits) that were either expected or unexpected endings of strong or weak constraint sentences plotted for the After condition (left) and Before condition (right). Error bars reflect standard error around the mean

Fig. 6.

Two-way interaction between constraint and expectedness for hits. Note. SC: Strong constraint. WC: Weak constraint. Predicted probability of “Yes” (seen word) responses for studied words (i.e., hits) that were either expected or unexpected endings of strong or weak constraint sentences plotted across conditions. Error bars reflect standard error around the mean

Recall

Average recall rate across conditions

To examine the impact of value on participants’ ability to recall sentence-level information, analyses on recall data were limited to responses that contained accurate message-level information – i.e., responses coded as verbatim, almost verbatim, or gist. This measure excluded responses that were single words or that were inaccurate recollections of the message. On average, participants in the After condition recalled roughly a third of the study sentences (M = 31.08, SD = 13.17) whereas Before condition participants recalled slightly less than a third of the study material (Before: M = 28.33, SD = 13.54); this difference was not significant t(46) = 0.71, p =.48. This suggests that knowing whether or not the upcoming sentence was important did not boost memory for sentences overall.

To examine the impact of value and the timing of its availability on sentence memory, participants’ sentence recall performance was analyzed by constructing a logistic mixed-effects regression model that included fixed effects for condition (After, Before), value (high, low), sentence expectedness (expected, unexpected), and the resulting three-way interaction, as well as all lower order interactions. The random effects structure included a random intercept for participants and item, random by-participant slopes for the fixed effects of value and expectedness and random by-item slopes for the fixed effects and all interactions except for the interaction between value and condition. This model revealed a main effect of value (β=0.52, 95%CI[0.28, 0.81]), which was qualified by a significant three-way interaction between condition, value and expectedness (β=0.57, 95%CI[0.04, 1.11]). A visual inspection of the interaction (Fig. 7) suggested that the relationship between condition and value differed across the two levels of expectedness, and so two additional models were constructed and are reported below to separately examine the impact of condition and value on expected and unexpected sentences, respectively. None of the remaining main effects or interaction terms were significant as the confidence intervals for the remaining predictors included zero as a plausible estimate.

Fig. 7.

Three-way interaction between condition, value and expectedness for recalled responses. Note. SC: Strong constraint. WC: Weak constraint. Predicted probability of recall for high or low value sentences that ended with an expected (left) or unexpected (right) ending for the After condition and Before condition. Error bars reflect standard error around the mean

Sentence recall by expectedness

To examine how having value information in advance might affect encoding strategies, we directly compared the impact of value and condition on recall rates for sentences by expectedness in two separate models. The model assessing recall of sentences ending with an expected word included fixed effects for value (high, low), condition (Before, After) and the interaction of the two and converged with the maximal random effects structure, which included random intercepts for both items and participants, random by-participant slopes for the main effect of value, and by-item slopes for the fixed effects. This model revealed only a main effect of value (β = 0.52, 95%CI[0.24, 0.80]), indicating that high-value sentences were recalled more often than low-value sentences. The 95% confidence intervals for the remaining estimates included zero, reflecting that no other effects were significant.

The model for sentences ending with an unexpected word included fixed effects for value, condition and the interaction of the two. The random effects structure included random intercepts for both items and participants and random by-participant and by-item slopes for the main effect of value. This model also revealed a main effect of value (β = 0.51, 95%CI[0.21, 0.79]), but also revealed an interaction between value and condition (β = − 0.68, 95%CI [−1.19, −0.20]), as the difference in recall accuracy for high and low value sentences was greater in the After condition compared to the Before condition. The effect of condition was not significant (β = −0.14, 95%CI [−0.61, 0.31]). Thus, although participants in the Before condition (compared to the After condition) were made aware of item importance prior to encoding, they showed worse performance for high value sentences that ended incongruently with their expectations; this pattern can be seen in Fig. 7. This data pattern aligns with the prediction that Before condition participants are more likely to engage prediction-based strategies for high value sentences, inducing interference during recall when those predictions are violated.

Discussion

Although much is known about how people can guide their memory for important but simple information like single words (Castel et al., 2007) or word pairs (Griffin et al., 2019; Elliott et al., 2020), less is understood about how people strategically encode and retain important information that takes on more complex forms. To address this question, we adapted the VDR paradigm developed by Castel et al. (2002) for use with sentences. After studying sets of sentences that were counterbalanced for importance (i.e., assigned high value or low value), sentential constraint (i.e., predictability of the sentence final word was strong or weak), and predictability (i.e., ended with the expected or an unexpected word), participants completed free recall tasks and yes/no recognition tests. A point value that indicated the importance of each item was always presented either before (Before condition) or after (After condition) each sentence.

We first asked whether people can use value-based information to strategically prioritize their memory for sentences and the words they contain. We found effects of value on patterns of sentence recall that aligned with prior work using single words: Sentences paired with high values were recalled more often than sentences paired with low values. This overall effect of value was observed irrespective of when participants had access to value information and for sentences that both did and did not end as expected. Our findings align with a breadth of prior research showing that memory is enhanced for important information, including not only high value words (Castel et al., 2013; Cohen et al., 2014) but also visual scenes associated with rewards (Adcock et al., 2006) and items of subjective importance (Lin et al., 2015; Friedman et al., 2015; Murphy & Castel, 2021a, b; Murphy et al., 2022). Here, we extend this area of work by showing that value information can also be used to guide memory for more complex and temporally-extended stimuli, like sentences.

The second goal of the present study was to examine in more detail how memory for a sentence might be impacted by the encoding strategies that people could differentially employ depending on whether or not they knew the value of the study item before reading it. In particular, Before condition participants, by virtue of getting value information before reading the study item, could choose to engage top-down processing strategies, including prediction, depending on the associated value of the to-be-read sentence. Consistent with our hypotheses, the present study revealed multiple pieces of converging evidence across recall and recognition measures that suggest that having value information available in advance may encourage people to be more likely to engage in prediction during online sentence processing and that doing so may actually have a negative impact on downstream memory.

First, we observed that participants in the Before condition showed worse recall performance for high value sentences that ended unpredictably compared to their After condition counterparts. Although prior research has shown that less predictable sentences are less easily remembered (Holmes & Murray, 1974; Brewer, 1975), this pattern in the current data is particularly striking because the study items were counterbalanced for value and predictability. In other words, participants in the two conditions read exactly the same sentences, with the only difference being when participants obtained the value cue, and, correspondingly, what encoding strategies they could engage in. Second, in both conditions, we replicated the luring effect found in Hubbard et al. (2019), such that participants showed increased false alarms to lures compared to non-lure new items; however, importantly, a constraint effect on luring behavior was only observed in the Before condition. Specifically, when participants were given the value cue prior to studying items (and thus, presumably, the opportunity to selectively engage in active prediction), they were significantly more likely to falsely recognize lures from strongly constraining sentences – sentence frames that supported stronger predictions compared to weakly constraining sentences. We also observed that expected endings from strongly constraining sentence frames showed reduced recognition, whereas recognition for prediction violations from those same sentence frames increased. This difference seemed to be driven primarily by the Before condition, although this trend was not found to be significant. This pattern is similar to that observed in Federmeier et al. (2007), who also recorded ERPs and found evidence (in the form of an anterior positivity to prediction violations) that participants were engaging prediction during sentence processing. Together, these patterns support the view that people may tend to engage in prediction-based strategies during comprehension of important information and that such behavior may have unintended costs for downstream memory.

Taken together, these findings suggest that considerations of when value information is available may have consequences for people’s processing of more complex information like sentences. It is striking that we found that some aspects of memory were more benefited when value information was presented after a sentence than when people had that information earlier. This is in contrast to previous findings for relatively simple information like single words or items, which suggested that valuable information might be better remembered when people know the value before encoding (Soderstrom & McCabe, 2011; Villaseñor et al., 2021; Leippe et al., 1978). Indeed, our data point to potential benefits of providing value information after encoding rather than before encoding when the items to be remembered are more complex. This can have implications for daily-life scenarios. For example, reminders like “The blue pills should only be taken on Tuesdays. It’s important to remember this.” may be more effective than “It’s important to remember that the blue pills should only be taken on Tuesdays.”

The data also point to interesting differences in the consequences of these strategies for memory for the whole sentence compared to the sentence-final words. When value information was provided in advance, high value sentences containing prediction violations were less likely to be recalled as a whole. However, the prediction violating words themselves were more likely to be recognized. Integrating our findings with the larger literature suggests that this pattern is unlikely to be driven just by task (recall versus recognition) differences. McFalls & Schwanenflugel (2002) had participants recall sentence-final words after reading sentences that were controlled for their expectedness and constraint. Their finding that prediction violations from strongly constraining sentences were recalled most often complements our (recognition-based) data in supporting that a memory benefit may exist for sentence-final words that go against the accrued context. On the other hand, Fisher & Craik (1980) tested participants’ recognition memory of whole sentences and reported that sentences that were predictable (in this case, by virtue of containing more semantic associates) were recognized more often than less predictable sentences – a pattern analogous to our sentence-level recall effects in suggesting that reduced predictability disadvantages memory. Thus, taken in conjunction with previous literature, the present study provides support for the idea that the impact of prediction violations on memory is different for whole sentences and single words. In particular, unexpected (Röer et al., 2019) or irrelevant (Foss & Cairns, 1970) words may disrupt encoding and thus reduce memory for the sentence as a whole but still generate an error signal that aids learning for the prediction violation itself (e.g., Dell & Chang 2014).

Although the present study successfully adapted the VDR paradigm to use with sentences and examined what may be the downstream consequences of different encoding strategies people engaged in for important information, the conclusions we can draw about the precise nature of those encoding strategies is limited. Here, we used behavioral data that is collected after encoding to try to make inferences about the strategies in which participants may have engaged during encoding, and, although the After condition and Before condition comparison reveals aspects of the consequences of that (different) processing, we cannot directly answer what kinds of strategies participants employed in the moment. For example, one might wonder whether patterns of memory performance in the Before condition, which we have hypothesized are linked to the use of predictive strategies, could instead reflect the cost of divided attention during encoding, in order to both maintain the value information and take in the sentence. However, although the negative impact of divided attention during encoding for memory has been well-documented (Craik et al., 1996; Fernandes & Moscovitch, 2000), we think it is unlikely to be the driving force for the different patterns observed in the Before versus After conditions here. First, participants in the Before condition did not show an overall memory impairment compared to those in the After condition, as would be typically observed under divided attention conditions. Before condition participants recalled a similar number of sentences and were equally capable of discriminating old words from new words compared to their After condition counterparts. Instead, presenting the value before study impacted the distribution of participants’ memory for words and sentences. Second, even if the Before condition manipulation did introduce some additional task load, memory for important information seems to be relatively spared under conditions of divided attention. For example, Middlebrooks et al. (2017) used a variety of divided attention tasks and showed that people still selectively retained memory for high value items. When participants engaged in a second task, be it actively checking for a sequential string of odd digits or passively listening to music, they recalled an overall smaller number of words compared to when their attention was not divided, but their memory for high value items did not differ across the different study conditions. Thus, we think different sentence processing strategies are a more likely source for the patterns we observed.

In conclusion, the present study successfully adapted the VDR paradigm to use with sentences, revealing that people can use value information to shape their encoding of even complex information that comes in over time. We further showed that people may adapt their value-based strategies as a function of when value information is available to them. Based on findings in the sentence processing literature, we suggest that this pattern could reflect the increased use of active prediction strategies in the Before condition. Intriguingly, however, if participants did engage in prediction strategies, these strategies did not improve memory overall and, indeed, created memory disadvantages for some types of important information, such as sentences that ended unpredictably. Thus, when people are given information that a sentence that they are about to read is important, they may tend to adopt modes of processing that can be advantageous for comprehension in the moment, but that may then not be optimal for later memory. Having established that people are sensitive to value information for sentences and that when that information is available impacts memory, future work using measures such as eye-tracking and event-related potentials that can be collected concurrently with sentence encoding may help to more specifically target which encoding strategies are being used under different conditions. In the long term, knowing more about how people adapt their language processing strategies as a function of the importance and likely predictability of what they are trying to comprehend can provide guidance about how text might be structured to best promote comprehension and later recall.

Appendix: Distribution of recall responses

Fig. 8.

Note. Average proportion of recalled responses categorized by quality