The impact of reading at rapid rates on inference generation
Abstract
Aims
Speed reading is advertised as a way to increase reading speed without any loss in comprehension. However, research on speed reading has indicated that comprehension suffers as reading speed increases. We were specifically interested in how processes of inference generation were affected by speed reading.
Methods
We examined how reading speed influenced inference generation in typical readers, trained speed readers and participants trained to skim read passages. Passages either strongly or weakly promoted a bridging or predictive inference. After reading, participants performed a lexical decision task on either a nonword, neutral or inference‐related word.
Results
Typical readers responded to strong and weak inference words faster than neutral words. There were no statistical differences in reaction time between inference‐related and neutral words for speed and skim readers.
Conclusions
These findings provide no substantive evidence that the appropriate inferences are generated when reading at rapid speeds. Thus, speed reading may be detrimental to normal integrative comprehension processes.
Highlights
What is already known about this topic
- Speed reading is advertised as an effective method to reduce reading time.
- Claims about how speed reading works are inconsistent with reading research.
- Research suggests comprehension suffers when reading at rapid rates.
- The reading strategies used by speed readers have been less explored.
What this paper adds
- This paper explicitly investigates the comprehension process of inference generation in speed, skim and typical readers. Inference generation is important for the creation of a coherent memory.
- This paper also investigates whether speed reading is an acceptable reading strategy for certain passages (i.e. very predictable passages).
- This paper found no evidence of correct inference generation for speed readers in any situation.
Implications for practice and/or policy
- Speed reading appears to be an unacceptable reading strategy regardless of the predictability of a passage.
- Speed reading should not be promoted as an effective reading method either academically or for leisure reading.
Speed reading courses are advertised as a way to increase reading speed without any loss in reading comprehension. These courses were originally offered as a workshop or series of classes (e.g. Evelyn Wood Reading Dynamics – http://www.ewrd.com/ewrd/index.asp) but with advances in technology, speed reading training can be accessed through DVD lessons, audio CDs, different computer programs and apps (e.g. Berg, 2010, www.spritzinc.com, www.spreeder.com, etc.). Recently, a number of stories have appeared in the popular media about new speed reading apps. These apps claim they can help people read at speeds of up to 1,000 words per minute without having to move their eyes (e.g. Khazan, 2014). Although there is evidence that speed reading techniques are not effective strategies based on what is known about the physical limitations of reading and the limited capacity of the human information processing system (Rayner, 1998), there is little recent research on these ‘new’ speed reading techniques and how they affect comprehension. Most speed reading courses claim these benefits can be attained using two main techniques. First, readers are instructed to train their eyes to take in more information each time they fixate on the text (e.g. five to seven words at a time or even an entire sentence at a time). The logic behind this idea is that the brain does not process as much information as it can on each fixation. Thus, if readers look at more information at once, then the brain will process information closer to its optimal processing capacity. Contrary to this claim, research shows that the perceptual window through which readers can process information is roughly 10 characters (e.g. 1 to 2 words; Rayner & Pollestek, 1998). Second, speed reading courses instruct readers to inhibit sub‐vocalisation (i.e. inner speech), which is the act of silently repeating words while reading. Speed reading courses claim that sub‐vocalisation is an unnecessary and time‐consuming process. However, sub‐vocalisation allows the reader to process words (Baddeley & Hitch, 1974) and improves passage comprehension by facilitating integrative processes (Slowiaczek & Clifton, 1980).
The research that has directly examined reading comprehension in speed readers supports the conclusion that speed reading is not effective (Walton, 1957; McLaughlin, 1969; Just et al., 1982; Carver, 1985). Just et al. (1982) compared reading comprehension among speed readers, typical readers, and typical readers asked to skim the text. Speed readers answered a similar number of general comprehension questions (i.e. gist questions) correctly as typical readers. However, when asked about details of the text, speed readers were unable to answer questions about the material if they had not directly fixated on the region where the information was found. In contrast, typical readers correctly answered questions about details. Furthermore, typical readers asked to skim the text demonstrated eye movement patterns and comprehension results very similar to speed readers. Additionally, Calef et al. (1999) tested comprehension of readers before and after a speed‐reading course. They found that although the speed reading group increased their reading speed, speed readers correctly answered fewer comprehension questions after the course than before (contrasted with a control group who correctly answered slightly more comprehension questions). Likewise, Carver (1984) found that presenting words at a rapid pace (1,000 words per minute) reduced comprehension to approximately zero, even for easy, predictable passages. Based on these results, he proposed his Rauding theory, which suggests that comprehension decreases linearly as speed increases, contrary to claims of speed reading training.
More recently, Miyata et al. (2012) examined individuals with no speed reading experience, some speed reading experience, and extensive speed reading experience and found a significant negative correlation between reading speed and comprehension; those with speed reading experience read faster but had lower comprehension scores. Furthermore, Schotter, Tran, and Rayner (2014) have shown that speed reading techniques that eliminate eye movements (such those with rapid serial visual presentation) by presenting words briefly and sequentially one at a time, without allowing regressions back to previous words, also affect comprehension negatively.
Thus, research shows that speed reading does not result in optimal comprehension of a passage. However, to the best of our knowledge, little research has examined how speed reading affects specific processes during reading comprehension, such as processes related to the generation of inferences. Inference generation is an important process in reading comprehension because inferences help readers connect ideas in a passage (Singer & Halldorson, 1996; McKoon & Ratcliff, 1992), help explain casual, temporal and spatial events (Graesser et al., 1994), and help readers fill in missing information (McNamara et al., 1996). Thus, it is essential to generate inferences while reading to create a coherent memory of a passage. Furthermore, memories that contain inferences are typically less resistant to forgetting across time (Kintsch, 1998; Kintsch et al., 1990). The purpose of this research is to extend prior findings on speed reading and reading comprehension by examining inference generation when reading at rapid rates.
Although there are several types of inferences (Klin, 1995; Keefe & McDaniel, 1993; Peracchi & O'Brien, 2004; Poynor & Morris, 2003), this research focuses on causal inferences generated during reading. A causal inference is generated to connect a cause or an antecedent to a consequence during reading. These inferences are typically generated rapidly and easily. There are two major types of causal inferences: bridging and predictive. Bridging inferences, which are often referred to as backward inferences, are inferences that connect ideas in a passage. Take the following sentences as an example:
- The actress stood on the edge of the 14th floor ledge and suddenly fell to the ground.
- Her orphaned daughter was compensated for the accident.
When readers finish sentence (2), they must generate the inference, the actress died, to understand the causal relationship in the passage. Thus, generating bridging inferences is necessary for a complete understanding of a passage. While reading for comprehension, it has been well established that readers will generate bridging inferences, aiding in their comprehension of passages (Fincher‐Kiefer, 1995; Millis & Graesser, 1994; St. George et al., 1997; Virtue et al., 2006a).
On the other hand, predictive inferences, which are often referred to as forward inferences, are not necessary for comprehension but form expectations about what will happen. For example:
- Todd decided to take his shoes off and wade in the water.
- With his next step, he didn't notice a piece of broken glass under his foot.
When readers finish sentence (2), they may generate the inference, cut his foot, although it is not necessary to comprehend the passage. A number of studies have shown that readers generate predictive inferences (O'Brien et al., 1988; Murray et al., 1993; Keefe & McDaniel, 1993).
For both types of inferences, the ease of generating the inference depends on causal constraint, which refers to how strongly the context of a passage suggests a possible event. Thus, a strongly constrained passage suggests an event or word is extremely likely and readers tend to easily generate the correct inference whereas in a weakly constrained passage, an event or word is less likely and it is more difficult to generate the correct inference (van den Broek & Huang, 1995; Cook et al., 2001; Linderholm, 2002; Virtue et al., 2006a; Virtue et al., 2006b).
Although inferences are crucial for a coherent memory representation, the inferences generated do not always reflect correct information, which subsequently results in an incorrect memory representation. This can occur for a variety of reasons. For example, an incorrect memory representation could result from a lack of effort (McNamara et al., 1996), individual differences in cognitive abilities (Virtue et al., 2006a; Long & Chong, 2001) or faulty semantic and syntactic processing (Barton & Sanford, 1993; Christianson et al., 2001; Erickson & Mattson, 1981; Ferreria et al., 2001).
Speed reading may also lead to an incorrect memory representation by generating incorrect inferences. It is possible that speed readers are skimming across content words, subjectively arranging them to make sense. In other words, readers might be generating inferences based on the skimmed words and the reader's subjective view of the contents of the passage. This strategy may work in predictable or highly constrained passages (e.g. strongly constrained bridging); however, as passages become less predictable and constrained (e.g. weakly constrained predictive), skimming is less effective for comprehension, and it is reasonable to assume that fewer correct inferences will be generated.
The primary goal of this research is to investigate whether correct inferences are generated when reading at rapid rates versus typical reading. To test this, typical readers, speed readers and readers trained to skim passages read many passages for comprehension. In these passages, the final sentence was manipulated to either illicit a bridging or predictive inference. Additionally, the constraint of these inferences was manipulated to either be strong (i.e. very predictable) or weak (i.e. less predictable). To test inference generation, a lexical decision task was performed after each passage. Some words were related to the inference that should have been generated. If those formally trained in a speed reading course are skimming passages and subjectively interpreting the content, then correct inference generation should become more difficult as the content becomes less predictable and they should behave similarly to individuals who were simply skimming the passages. Therefore, we predicted that typical readers should generate inferences regardless of constraint, resulting in faster response times to inference‐related words than neutral words. However, speed and skim readers should generate fewer correct inferences in weakly constrained conditions, resulting in no reliable difference in response times to weakly constrained words relative to neutral words. Additionally, if speed readers are skimming passages, then there should be no evidence of differences between speed and skim readers in reading time and responses to words related to inferences.
Method
Participants
Ninety students enrolled at a college in the southeastern United States participated for course credit or for payment. Thirty students were trained in a speed reading course and received 10 dollars an hour for their participation. Thirty students were trained to skim passages and made up the skimming group. And 30 students read typically and made up the control group. This group sample size was chosen because it was the mean sample size in Calef et al.'s (1999) study comparing speed readers (25 participants) and typical readers (34 participants) and is the largest sample of trained speed readers found in our literature review. Students in the skimming group and the control group were also paid 10 dollars an hour or received course credit. All participants were proficient in English. Following the experiment, participants were debriefed regarding the experiment and informed that while speed reading training can improve reading speed, research has shown it can hurt comprehension.
Speed reading training
Subjects in the speed reading group completed between 8 and 12 training sessions of eyeQ by Infinite Mind, a nationally marketed, computerised speed reading training course which claims to significantly increase reading speed while also improving comprehension. According to the developers, this is accomplished by training eyes to move and see faster and by teaching readers to overcome bad reading habits such as subvocalisation and rereading. Three subjects completed the training program in a tutoring centre at Nova Southeastern University in preparation for graduate examinations. The remaining subjects completed their training over three training days during a two week period in a lab at Nova Southeastern University. Each training day consisted of three training sessions, for a total of nine sessions. According to the developers of the course, training in the course is considered complete after eight sessions. Each training session of the program takes approximately 8 minutes to complete and consists of many ‘high speed imaging activities’ using both images and text. For instance, in one activity, subjects are asked to track and fixate on objects as they rapidly appear on a computer screen at different locations and with varying speeds. In another activity, subjects are asked to read/scan scrolling passages, which were presented at different speeds. Subject reading time was measured before and after each training session. Subjects were also given a 5‐minute break between each training session to rest their eyes, as recommended by the developers. Additionally, at the end of each training day, subjects completed two other training exercises designed to increase their reading speed, according to the course. One exercise consisted of quickly scanning lines of a passage, while the other consisted of speed reading a passage and answering comprehension questions about material in the passage.
Skimmer training
Subjects in the skimming group were trained in the lab to skim passages at approximately 550 words per minute. This was accomplished by presenting subjects with four 550 word training passages, which they were asked to read in 1 minute. After reading each passage, subjects adjusted their reading speed based on how far they had gotten in the passage, to reach the target pace. All subjects achieved the target reading time before reading four passages. Once they reached the target reading time, they were asked to read that way during the experiment.
Materials
Passages
During the experimental testing session, participants read 150 passages, which were approximately 80 words each. All passages were four sentences and were read sentence by sentence with reading time recorded. Sixty of the passages, which were adapted from Virtue et al. (2006a), prompted either a bridging or predictive inference. In a bridging inference passage, the inference was necessary for comprehension. In a predictive inference passage, although there was a weak implication, predicted outcome was not as clear. Of the 60 inference passages, 30 prompted bridging and 30 prompted predictive inferences. For the passages that prompted an inference, the first two sentences introduced the scenario and the characters and the last two sentences in the passage consisted of an implicit critical sentence (1) and a target sentence (2). For example, (1) ‘Business people scattered to get indoors as to not ruin their work clothes’, followed by (2) ‘The once dry sidewalks now were all wet’. The 30 inference passages of each type were either strong or weak causal constraint passages. In a strong causal constraint passage, the implicit critical sentence implied that something would happen, but when readers reached the target sentence, they either had to generate a bridging inference to continue to comprehend the passage (e.g. rain) or generate a predictive inference about what was going to happen. In a weak constraint passage, the implicit critical sentence weakly implied something would happen. Thus of the 60 passages, 15 were strong‐bridging inference passages, 15 were weak‐bridging inference passages, 15 were strong‐predictive inference passages and 15 were weak‐predictive inference passages. Each passage was counterbalanced between passage type and constraint type, resulting in four versions of the experiment (See the Appendix for sample passages). Ninety passages were considered filler passages and did not prompt any particular type of inference.
Lexical decision task
This task involved deciding as quickly and as accurately as possible whether a string of letters presented on the screen is a word. Following each passage, readers were presented with either a nonword (e.g. rord), a word that is related to the inference promoted by the passage they just read (e.g. see sample passage in the Appendix followed by the word: rain) or a word that was not related to the inference promoted by the passage (e.g. root). The purpose of the lexical decision task was to determine if the inference implied by the passage was activated in the reader's memory. Responding to the inference‐related words faster than the nonrelated words is evidence that the correct inference has been activated. This method has been validated through several experiments (Millis & Graesser, 1994; Virtue et al., 2006a). Sixty of the words used in this task were related to an inference prompted by the passages. Thirty were related to passages that prompted a bridging inference (15 strongly and 15 weakly constrained), and 30 were related to passages that prompted a predictive inference (15 strongly and 15 weakly constrained). All inference‐related words followed either a bridging or predictive inference passage. Virtue et al. (2006a) normed the passages and found that target inference words were mentioned significantly more often as part of causal explanations for strongly constrained versions than weakly constrained versions. Additionally, for both strongly and weakly constrained passages, average responses were significantly greater than would have been expected by chance (see Virtue et al., 2006a for details). An additional 15 words were used that were not related to an inference prompted by a passage (neutral words). These words followed filler passages and were collected to compare to inference‐related words. As seen in Table 1, all words in the lexical decision task were matched on orthographic neighbourhood size and bigram frequency (Medler & Binder, 2005), as well as age of acquisition, concreteness, familiarity, word frequency, imageability, length and number of syllables (Wilson, 1988). To combat a preference bias, 75 pronounceable nonwords (i.e. English phonemes were used) were used and followed the additional filler passages. Nonwords matched normal words on length and number of syllables.
| Inference words | Filler words | Pairwise analyes | ||
|---|---|---|---|---|
| Measure | M (SD) | M (SD) | t‐value (df) | p‐value |
| Age of acquisition (years) | 5.03 (1.36) | 4.74 (.30) | −.47 (59) | .64 |
| Bigram frequency | 1423.48 (1083.44) | 1362.40 (1141.90) | −.2 (73) | .84 |
| Concreteness | 463.85 (77.87) | 462.44 (58.16) | −.05 (53) | .96 |
| Familiarity | 558.35 (34.62) | 555.50 (26.15) | −.25 (54) | .81 |
| Frequency | 50.07 (60.68) | 28.38 (32.77) | −1.37 (73) | .17 |
| Imageability | 512.91 (60.13) | 497 (70.44) | −.74 (54) | .47 |
| Length | 4.46 (.97) | 4.38 (1.20) | −.29 (73) | .78 |
| Number of syllables | 1.08 (.28) | 1.13 (.34) | .49 (73) | .63 |
| Ortho neighbourhood size | 8.19 (5.62) | 8.81 (4.90) | .41 (73) | .69 |
Comprehension questions
Twenty‐two yes/no comprehension questions were asked after the lexical decision task for randomly chosen filler passages. These questions were statements that were either explicitly stated in the prior text or contained incorrect information from the prior text. They were used to ensure passages were read for meaning.
Questionnaires
Participants also completed a speed reading test, a language history questionnaire and a vocabulary quiz. The speed reading test calculated how many words participants read silently per minute, as part of the eyeQ software package. The vocabulary quiz contained 30 words that were normed by Raney et al. (2000) for a college population at a diverse university. This test is significantly correlated with reading ability (r = .52). The language history questionnaire was used to assess the languages that participants speak, read and comprehend and the relative proficiency in each language. This questionnaire was needed to ensure that the participants were proficient with the English language and only participants that averaged higher than 8 (out of 10) in speaking, reading and comprehending English were included in the sample.
Procedure
All participants completed a test to compute their reading speed before beginning the experiment. After the test, speed reading subjects were instructed to read for comprehension while using the training they had learned during speed reading; skimming subjects were instructed to read for comprehension while skimming at the speed they had been trained; subjects in the control group were instructed to read normally for comprehension. Participants were told that they would read many passages on a computer one sentence at a time for comprehension while their reading time was measured. Additionally, they were told that after the fourth sentence of every passage, they would perform a lexical decision task in which they would decide if a string of letters was a word or nonword. They were also told that after one‐third of the passages, they would answer a comprehension question. The experiment began with two practice passages and two lexical decision tasks. Sentences were presented horizontally left justified and vertically centred. Participants pressed the space bar to advance to the next sentence. This caused the previous sentence to disappear and a new sentence to appear. This process was continued until the entire passage was read. Then, a fixation cross, which was vertically and horizontally centred, appeared on the screen for 1000 ms, and then participants were presented with a string of letters. Participants pressed the 1 key on a keyboard if the string of letters was a word and the 2 key if the string of letters was not a word. After the lexical decision task, for one‐third of the passages, a yes/no comprehension statement appeared. Participants pressed the Y key if the statement was true and the N key if the statement was false.
Once they finished the computerised portion of the experiment, each subject completed the vocabulary quiz and the language history questionnaire.
Vocabulary scores
We compared vocabulary scores of readers across the three between‐group conditions (i.e. normal, speed and skim). Because vocabulary is strongly correlated with reading ability (Raney et al., 2000; Stanovich, 1986; Stanovich, 2000), it serves as a proxy of reading ability. The number of correct answers in the vocabulary quiz was tallied to create a vocabulary score for each subject. To determine if there were any noticeable differences in vocabulary as a function of reading group, a one‐factor ANOVA with reading group as the between‐subject independent variable and vocabulary score as the dependent variable was computed. The main effect was marginally significant, F (2, 87) = 3.09, p = .05, with speed readers having marginally better vocabulary scores. Table 2 contains descriptive statistics for the reading groups. Variability in our sample is comparable to previous research samples (Raney et al., 2002).
| Typical readers | Speed readers | Skim readers | ||||
|---|---|---|---|---|---|---|
| Measure | M | SD | M | SD | M | SD |
| Vocabulary score | 17.4 | 3.5 | 20.1 | 5.31 | 19.0 | 4.20 |
| Average word reading time | 279.81 | 63.88 | 178.58 | 65.96 | 149.90 | 45.41 |
| Comprehension score | 16.4 | 2.13 | 15.9 | 2.02 | 14.1 | 2.95 |
- Note. The vocabulary quiz is from Raney et al. (2000) and consisted of 30 items. Average word reading time is measured in milliseconds. Twenty‐two comprehension questions were asked.
Results
Speed reading training
An alpha level of .05 was used to determine the significance for all analyses. Because three of our speed readers were trained using the same program at a tutoring centre on campus, we do not have training data for them. However, to determine if speed readers trained in the laboratory (N = 27) increased their reading speed, seven reading time scores were collected during each of the three training days. To examine whether subjects increased their reading speed during the training to acceptable speed reading levels, a 3 (training day: one, two and three) × 7 (reading time trials: 1–7) RM ANOVA was computed with average number of words read per minute as the dependent measure. There was a significant main effect of training day, F (2, 54) = 18.13, p < .001, η2 = .20, and reading time trials, F (6, 162) = 10.18, p < .001, η2 = .07, indicating reading time increased significantly over training day and trial. Linear regression analyses were used to test whether training day and trial predicted reading speed. Results indicated that the two predictors explained 10% of the variance [R2 = .10, F (2, 584) = 30.69, p < .001]. Both training day (β = .27, p < .001) and reading time trial (β = .15, p < .001) significantly predicted reading time. As can be seen in Figure 1, subjects significantly increased the number of words they read per minute both over reading time trials during one day and over speed reading training days. All subjects increased their reading time by more than 145 words/minute, and half of the subjects increased their reading speed by more than 300 words/minute. These analyses show that the speed readers significantly increased their reading speeds.
Experiment sentence reading time
Reading time for each sentence was collected and divided by the total number of words per sentence. To investigate differences in average word reading time across all passages, a one‐factor ANOVA with reading group as the independent measure and average word reading time as the dependent measure was computed. Reading group served as a between‐subject measure in this analysis and all remaining analyses. There was a significant main effect of reading group, F1 (2, 88) = 39.97, p < .001 η2 = .48. Tukey's honestly significant difference test revealed that speed readers and skim readers' average word reading time was significantly faster than typical readers, p < .001. As Table 2 shows, speed and skim readers read 101 and 130 ms faster per word, on average, than typical readers, respectively. Although, skim readers read, on average, 29 ms faster per word than speed readers, this difference failed to reach acceptable levels of statistical significance, p = .16. Furthermore, to ensure that speed readers trained in the lab and speed readers trained at the tutoring centre on campus were comparable on reading speed, we specifically examined average word reading time of these groups. Lab trained speed readers read at a comparable speed (178.51 ms/word) to those trained in the tutoring centre (179.23 ms/word), t (27) = .02, p = .99.
Comprehension questions
To ensure participants were putting forth effort, mean comprehension scores (Table 2) for each group were compared to chance performance (11 out of 22 yes/no comprehension questions). All groups performed significantly better than chance (all ps < .001) and scored 79% or higher on the comprehension questions, indicating that participants were taking the task seriously. Furthermore, speed readers trained in the lab performed similarly on the comprehension questions (15.92) to those trained in a tutoring centre (15.67), t (27) = −.2, p = .84.
Lexical decision task
Lexical decision response time and accuracy for the target words were collected and analysed. Accuracy scores were created by averaging lexical decision response scores for each subject as a function of inference condition and constraint (see Table 3 for means). The analyses reported here include outliers. The same analyses were conducted with outliers three standard deviations above and below the mean in each condition Winsorized (4% of response times in the speed reading group, 2% of response times in the skim group and none in the typical group), and the results did not change. In all the analyses reported, F1 refers to tests based on participant variability, and F2 refers to tests based on item variability. For consistency, results are reported as significant when both participant variability and item variability analyses have p‐values less .05. Furthermore, all posthoc comparisons were preplanned comparisons.
| Normal | Speed | Skim | ||||
|---|---|---|---|---|---|---|
| Measure | RT | ACC | RT | ACC | RT | ACC |
| Bridging inference | ||||||
| Strong | 735 (201) | 1.0 (0.02) | 806 (316) | 0.99 (0.02) | 874 (335) | 0.99 (0.03) |
| Weak | 790 (202) | 0.99 (0.03) | 833 (297) | 0.99 (0.02) | 843 (262) | 0.99 (0.03) |
| Predictive Inference | ||||||
| Strong | 758 (213) | 1.0 (0.02) | 826 (360) | 0.98 (0.03) | 890 (394) | 0.99 (0.03) |
| Weak | 805 (222) | 1.0 (0.02) | 849 (303) | 0.99 (0.02) | 858 (271) | 0.98 (0.04) |
| Neutral | 839 (233) | 0.79 (0.02) | 821 (258) | 0.81 (0.06) | 879 (266) | 0.79 (0.05) |
| Nonword | 902 (216) | 0.96 (0.05) | 971 (403) | 0.91 (0.19) | 1116 (470) | 0.89 (0.12) |
To investigate how reading rates influenced accuracy of the lexical decision task, a 3 (reading group: speed, skim and typical) × 5 (word type: bridging strong, bridging weak, predictive strong, predictive weak, neutral) ANOVA was computed with accuracy as the dependent measure. Nonwords were not included in the analyses because they cannot be matched to actual words on all the lexical characteristics target words were matched upon. None of the comparisons were statistically different. Thus, accuracy on the lexical decision was not significantly affected by reading group or word type.
Lexical decision response time
Incorrect responses on the lexical decision task were omitted from response time analyses. To investigate how reading rates influenced generation of inferences, a 3 (reading group: speed, skim and typical) × 5 (word type: bridging strong, bridging weak, predictive strong, predictive weak, neutral) ANOVA was computed with reaction time to words as the dependent measure. Nonwords were not included in the analyses because they cannot be matched to actual words; however, as Table 3 shows, all reading groups responded to nonwords slower than words. There was no significant main effect of word type, F1 (4, 348) = 2.24, p = .07 η2 = 0.06; F2 (5, 432) = 72.38, p < .001. There was also no significant main effect of reading group [F1 (2, 87) = .73, p = .48, η2 = .02; F2 (2, 432) = 143.06, p < .001]. However, there was a significant word by reading group interaction, F1 (8, 348) = 2.23, p < .05, η2 = 0.05; F2 (10, 432) = 4.43, p < .001. As seen in Figure 2, tests of simple effects revealed a significant main effect of word type for typical readers F (4, 116) = 7.61, p < .001, η2 = 0.27. Pairwise comparison showed that typical readers responded to neutral words slower than all inference‐related words, p < .05, except words in the predictive weak condition. This suggests typical readers were generating the inferences prompted by bridging and strongly predictive inference passages. Additionally, inference‐related words that followed strongly constrained predictive inference passages were responded to faster than inference‐related words that followed weakly constrained bridging and predictive inference passages, ps < .05. No other comparisons were statistically different. In contrast, there was no main effect of word type for either speed readers, F (4,116) = 1.07, p = .38, η2 = 0.04, or skim readers F < 1, η2 = 0.02. Thus, there is no evidence that speed or skim readers generated the appropriate inferences prompted by bridging or predictive inference passages.
Facilitation scores
To further explore the role of constraint, proportional facilitation scores were computed for inference‐related words. To create these scores, response time differences between neutral words and the four inference word conditions were computed and then divided by the neutral word reaction time. Thus, proportional facilitation scores illustrate the amount of facilitation received on inference‐related words compared to neutral words relative to each reader's baseline reading speed. A 2 (passage type: bridging or predictive) × 2 (constraint: strong or weak) × 3 (reading group: speed, normal or skim) ANOVA was computed with proportional facilitation scores (response time to neutral words minus response time to inference words/response time to neutral words) as the dependent measure. There was a marginal effect of constraint, F (1, 87) = 3.8, p = .06, η2 = 0.03, such that strongly constrained words showed marginally more proportion of facilitation than weakly constrained words. There was a significant main effect of reading group, F (2, 87) = 3.69, p < .05, η2 = 0.23, such that typical readers illustrated greater facilitation than speed readers, p < .01 and marginally more facilitation than skim readers, p = .06. There was no statistical difference in facilitation scores for either speed or skim readers. These effects were qualified by constraint by group interaction, F (2, 87) = 4.33, p < .05, η2 = 0.06. Tests of simple effects were computed at each level of constraint comparing the three groups. As Table 4 shows, in the strongly constrained condition, there was a significant effect of group, F (2, 87) = 4.82, p < .01. In the weakly constrained condition, the effect of group was not significant, F (2, 87) = 2.78, p = .07. In the strong constraint condition, posthoc LSD tests indicated typical readers showed greater facilitation scores than both speed readers and skim readings (ps < .01). There were no statistical differences in facilitation scores between speed and skim readers in either constraint condition. Once again, there was little evidence that speed or skim readers were generating the correct inferences. It is also important to note that the proportional facilitation scores for speed and skim readers are close to 0 (see Table 4), which provides further evidence that the correct inferences do not appear to be active in speed or skim readers memory.
| Weakly constrained | Strongly constrained | |||
|---|---|---|---|---|
| Reading group | M | SE | M | SE |
| Normal | 0.04 | 0.02 | 0.10 | 0.03 |
| Speed | −0.03 | 0.02 | 0.01 | 0.03 |
| Skim | 0.03 | 0.02 | 0.00 | 0.03 |
Discussion
The primary goal of this research was to investigate whether correct inferences are generated when reading at rapid rates versus typical reading. As predicted, typical readers responded to inference‐related words faster than neutral words. This suggests that typical readers generated the appropriate inferences prompted by bridging and predictive passages. Furthermore, the effect was larger for strongly constrained passages than weakly constrained passages and occurred in all conditions except predictive passages that were weakly constrained. Overall, typical readers were more likely to generate appropriate inferences when a passage was more predictable. Generation of these inferences is important for a coherent mental representation, particularly because predictive inferences play a role in situation model development (Kintsch, 1998; Schmalhofer et al., 2002). However, contrary to typical readers, there was no evidence that speed and skim readers generated the appropriate inferences prompted by bridging and predictive inference passages, as there were no statistical reaction time differences between inference‐related words and neutral words between these two groups. Additionally, there was no statistical difference in sentence reading time and lexical decision reaction time between speed and skim readers. Thus, there was no evidence that speed and skim readers engaged in different strategies while reading. Speed reading training significantly sped‐up reading time, but, it appears that this speedup was at the cost of processes important for the development of a coherent memory representation (i.e. correct inference generation). This is consistent with Carver's (1984) Rauding theory which proposes that increases in reading speed decrease reading comprehension. Interestingly, speed readers did not appear to be generating correct inferences even though this group had marginally higher vocabulary scores (and thus potentially higher reading comprehension) than both the typical and skim readers.
The finding that speed and skim readers did not produce a reliable difference in reaction times between inference‐related words and neutral words is partially consistent with predictions. Although it was predicted that no facilitation would occur for the weakly constrained passages because they were less predictable, we did expect facilitation to occur for strongly constrained passages, which are more predictable. However, there was no evidence that speed and skim readers generated the appropriate inferences in either more or less predictable situations. There are several explanations for this null finding. First, it is possible that the lexical decision technique was not sensitive enough to capture inference generation. Because typical readers' reaction times fluctuated as a function of constraint and because prior research (e.g. Potts et al., 1988; Virtue et al., 2006a) had similar results using this method, this explanation seems unlikely. Second, it is possible that speed and skim readers generated the necessary inferences after the lexical decision task. That is, there might be a delay when generating inferences at rapid rates. Given that the lexical decision task occurred after reading the final sentence of the passage and that the generation of bridging inferences in typical reading occurs during reading and not after (O'Brien et al., 1988; Virtue et al., 2006a), this explanation also seems unlikely.
Another explanation for the finding that speed and skim readers did not produce a reliable difference in reaction times between inference‐related words and neutral words is that neither are generating inferences during reading. This also seems unlikely. Comprehension scores were also collected for speed and skim readers. Comprehension across all groups (speed, skim and typical readers) was significantly better than would be expected from chance performance. Because comprehension questions only tested explicitly stated information (and not information related to potential inferences), it is reasonable to assume that speed and skim readers were building at least some form of a textbase representation. Although these representations mainly reflect what was explicitly mentioned, some inference generation is necessary, even if it is simply background information that is quickly and easily available (McKoon & Ratcliff, 1992). If speed and skim readers are in fact not building a representation beyond some of what was explicitly mentioned in the text, they might be engaging in comprehension strategies similar to good‐enough representations. For example, Ferreira and Patson (2007) argue the comprehension processing system might interpret the local structure of a sentence (i.e. adjacent words) but the system fails to interpret or becomes ‘lazy’ when constructing the global structure, which ultimately leads to a misrepresentation. While this misrepresentation reflects some information about the text, the global meaning is often lost. Although the current research did not directly test this assumption, given that speed and skim readers did comprehend some explicit information about the passages, but failed to show evidence of more global comprehension (i.e. inference generation), it is reasonable to speculate that they might be creating good‐enough representations. Future research should explore the level of representations (surface form representation, textbase, situation model) of speed and skim readers to determine what kind of passage information is stored in memory.
Another possibility is that speed and skim readers are not generating the appropriate inferences even in very predictable situations. Reaction times did not fluctuate as a function of constraint for these readers. Additionally, facilitation scores, which demonstrate the amount or proportional speedup on inference‐related words in comparison to neutral words, are relatively small. Proportion scores after reading strongly constrained passages illustrate this point. The facilitation scores for speed and skim readers were 0.01 and 0.00 respectively, which indicates a 1% and 0% speedup on inference‐related words, respectively. This is quite a difference from the typical readers' facilitation score, which was .10 and indicates a 10% speedup on inference‐related words. Thus, it appears that little to no speed up occurred for inference‐related words in speed and skim readers. Consequently, there is no evidence that any correct inferences are activated in the speed and skim readers' memories. Thus, it may be the case that correct inferences were not generated even in predictable situations. If this is the case, then there are no situations in which reading at rapid rates is an acceptable strategy for a coherent and accurate memory representation.
The absence of a statistical difference between speed and skim readers in reading time and lexical decision reaction times offers some evidence that speed and skim readers are not using different inference generation strategies. Thus, speed reading courses do not appear to ‘teach’ readers any special strategies that facilitate comprehension. Taken together with the lack of an effect as a function of constraint in lexical decision reaction time for speed and skim readers, it appears that speed reading is detrimental to building a correct memory representation in all situations.
This research does not provide evidence of correct generation of inferences when reading at rapid rates. However, this research does not explore the actual representation that speed and skim readers are creating. Thus, it is unclear what is included in speed and skim readers' memory for texts read at rapid rates. Future research could potentially explore this by measuring the kinds of memory representations (surface form representation, textbase, situation model) speed readers have. Nonetheless, it appears that the correct inferences are not included in these memories for text read at rapid rates. Another potential limitation of the study is it cannot determine whether inference generation was occurring online, during reading. It would be worthwhile to use measures of eye tracking to get a better idea of the time course of comprehension processes in speed readers.
To summarise, this research found that typical readers generate the correct inferences and that the ease of inference generation fluctuates as a function of constraint. On the other hand, speed and skim readers show no evidence of correct inference generation and no evidence of strategy differences between the two groups. Although research illustrates why speed reading claims cannot work based on the physical characteristics of the eye and reading processes (e.g. Rayner & Pollestek, 1998; Slowiaczek and Clifton, 1980), this research illuminates differences between normal and speed reading strategies, namely that of inference generation. In other words, this research explores what speed readers are doing while reading at rapid rates. Additionally, because there was no evidence of inference generation for speed readers in either less or more predictable situations, reading at rapid rates appears to be a poor strategy regardless of predictability of the passage. This research, like previous research conducted on speed reading, illustrates the downside that comes from rapid reading. Namely, those who believe they have been trained to read at rapid rates with no loss in comprehension may have false confidence in their own comprehension of text. In fact, it appears they are actually experiencing a speed‐comprehension trade‐off which could have serious implications wherever speed reading training is encouraged, whether it be with lawyers, business people or even students.
Appendix
Sample Passages by Inference Type and Constraint
Bridging Inference, Strong Constraint.
The weatherman said that the day was going to be sunny and clear.
But, there were many ominous clouds forming in the sky.
Business people scattered to get in doors as to not ruin their work clothes.
The once dry sidewalks now were all wet.
Bridging Inference, Weak Constraint.
The weatherman said that the day was going to be sunny and clear.
But, there were many ominous clouds forming in the sky.
Business people scattered to get in doors as to not ruin their work clothes.
Everything outside was now damp.
Predictive Inference, Strong Constraint.
The work week for adults was about to start.
The weatherman said that the day was going to be sunny and clear.
But, there were many ominous clouds forming in the sky.
Business people scattered to get indoors as to not ruin their work clothes.
Predictive Inference, Weak Constraint.
The work week for adults was about to start.
The weatherman said that the day was going to be sunny and clear.
But, there were many ominous clouds forming in the sky.
Business people scattered to get indoors.
Target inference word in all conditions: RAIN.
Biographies
W. Matthew Collins is an Associate Professor in the Department of Psychology and Neuroscience at Nova Southeastern University. His research interests include the representation of language in memory, text comprehension and the interaction between language and memory.
Frances Daniel is an Associate Professor in the Department of Psychology at Indiana University Northwest. Her research interests include language processes, text comprehension and bilingualism.
