The Effects of Different Sleep Stages on Language Learning Tasks in Young Adults
Dublin Core
Title
The Effects of Different Sleep Stages on Language Learning Tasks in Young Adults
Creator
Carly Power
Date
2021
Description
In order to learn a language, one must practice multiple tasks, including speech segmentation and generalisation. Segmenting speech allows for the identification of words and learning the meaning as well as syntactic role of those words within phrases and sentences. Novel generalisation requires generalising over the structure of a new language not yet experienced. Frost and Monaghan (2016) showed that participants were able to use the same statistical information at the same time to complete both language tasks. They suggest that segmentation and grammatical generalisation are dependent on similar statistical processing mechanisms. The role of sleep for learning to segment and generalise language is still unclear. Sleep affects memory consolidation, which is necessary for learning a novel language. This refers to the amount of sleep individuals get within their sleep cycle, yet it is unknown whether the duration of separate sleep stages has an effect. The declarative/procedural (DP) model by Ullman (2004) on learning provides distinctions in DP memory that associate with slow-wave sleep (SWS) and rapid-eye movement (REM) sleep respectively. SWS has a role in declarative memory processes, including memory for words and grammar. Rapid-eye movement (REM) sleep has a role in procedural memory processes, involving motor skills and coordination. Sleep spindle density should also be considered, as spindles are involved in offline information processing and information transfer. It was found that increased SWS and stage 2 spindle density have a positive effect on speech segmentation compared to generalisation.
Subject
Language learning, novel generalisation, REM, sleep, sleep spindle density, sleep stages, speech segmentation, SWS
Source
Participants
The original experiment was completed by 54 participants, 8 males and 46 females, with an age range of 18-24-years-old (mean age = 18.52). All participants reported being native-English speakers, with no history of auditory, speech or language disorders known. All participants either received university course credit or £20 for completing the experiment. Observations may be excluded for the first linear mixed-effects model. Exclusions may come from participants in the sleep group who did not sleep during the permitted time. This is because the first analysis aims to compare sleep vs. wake. The same participants’ data will be kept for the other linear mixed-effects models which aims to compare duration of sleep stages. This research received ethical approval by Dr Padraic Monaghan and Lancaster University’s Psychology Department on 22/04/2021.
Design
This study had a between-participants design with two conditions: sleep vs. wake between training and testing, and test type. There were two test types of speech segmentation and novel generalisation. Participants were randomly allocated to the sleep or wake conditions, and split evenly, meaning 27 participants slept and 27 remained awake. This study had access to PSG data for 18 of the participants in the sleep group. All participants received both test types. All participants were provided with an information sheet and gave written consent before the study commenced.
Materials
Stimuli
Using the Festival speech synthesiser (Taylor et al., 1998), speech stimuli were created that were based on similar stimuli used by Peña et al. (2002). This artificial training language contained monosyllabic items, of which there were nine (pu, ki, be, du, ta, ga, li, ra, fo), used to form three different non-adjacent pairings with three possible X items in-between (A1X1–3C1, A2X1–3C2, and A3X1–3C3) (Frost & Monaghan, 2016). Using Peña et al.’s (2002) study, A and C items contained plosive phonemes (pu, ki, be, du, ta, go) and X items contained continuants (li, ra, fo). All AXC item strings had a duration of approximately 700ms. Any preferences – for dependencies not due to the statistical structure of the sequences – were controlled for by generating eight versions of the language. Each version had randomly assigned syllables to A and C items, and the same X items were used in all versions. These versions of the language were counterbalanced across both task types. When testing for novel generalisation, three additional syllables were used with continuant phonemes (ve, zo, thi) (Frost & Monaghan, 2016). Research on the similarities in phonological properties of non-adjacent dependent syllables has shown that these similarities show support for acquisition of such nonadjacencies (Newport & Aslin, 2004). Nonetheless, other research has found that they are not essential for language learning to occur (Onnis et al., 2004). Words in the same grammatical category tend to be coherent regarding phonological properties (Monaghan et al., 2007), so regardless of learning, this property of the artificial language used within this study is consistent with natural language, which allows for real-life implications.
Training
The speech stimuli were formed into a 10.5-minute-long continuous speech stream by stringing together the AXC words within the language. It was ensured that no Ai_Ci dependencies were repeated immediately after each other. The speech stream included 5s fades for the onset and offset of speech, which ensured that such a feature of speech could not be used as a language structure cue (Frost & Monaghan, 2016).
Testing
Segmentation: part-words were trisyllabic items that were heard in the training speech stream but overlapped word boundaries. As such, part-words comprised of either the last syllable of one word and the first two syllables of the next (CiAjX), or the last two syllables of one word and the first syllable of the next (XCiAj). For all nine AXC items, both part-word types were created. 18 test pairs were constructed which participants listened to, by matching each part-word with its corresponding word (for example, the A1X2C1 item was paired with the X2C1A2 part-word) (Frost & Monaghan, 2016).
Novel generalisation: nine forced choice tests included a rule-word which contained one of three novel syllables (ve, zo, thi) (AiNCi), where N is the novel syllable and a novel part-word. For each Ai_Ci dependency, each novel rule-word appeared once. Part-words were made of two syllables that were heard in the training task, in their respective positions, with the same novel syllable as in the rule-word sequence (Frost & Monaghan, 2016). This novel syllable could appear in any position (first NCiAj, second XNAi, or third CiAjN) and each novel syllable occurred once in each of these positions. Rule-word and part-word novelty presence controlled for the effect of the novel syllable, yet the novel generalisation task still tested for generalisation of the non-adjacent structure of items within speech (Frost & Monaghan, 2016). Randomisation of test-pairs in all conditions was ensured across all participants, including the position of the correct response in each test-pair, to reduce response bias. When listening to the test-pairs, items in each pair were separated by a 1s pause. All participants completed The Stanford Sleepiness Scale (SSS) (Hoddes et al., 1972). This was in order to note participant sleepiness before the period of sleep or wake. The SSS consists of one item on a scale of seven statements, within which participants were required to select one statement that best described their perceived level of sleepiness (Shahid et al., 2011) (see Appendix A). Participant responses in the testing task were excluded if 90% of responses were always “1” or “2”, or if responses alternated between “1” and “2”.
Procedure
The whole procedure lasted for a three-hour period. For the training task, all participants listened to the continuous stream of speech and were instructed to pay attention to the language and think of possible words it contains. After the training task was complete, participants were split into two groups for the sleep vs. wake condition. Half of the participants, the sleep group, were given an hour and 45 minutes to sleep. These participants slept at Lancaster University Psychology Department’s sleep lab, and their sleep was monitored using polysomnography (PSG). PSG and an Embla N7000 system can record the amount of time spent in each sleep stage, and sleep spindle density, with EEG sites: O1, O2, C3, C4, F3, and F4 referenced against M1 and M2. The other half of participants remained awake for the same duration, watching a non-verbal, emotionally neutral video with neutral music. The testing task was then given to all participants after the same amount of time, 15 minutes after the break period. All participants were then required to complete the testing forced choice tasks. Within each trial, participants listened to a test-pair of items and were instructed to select which item best matched the training language. A response of “1” for the first item or “2” for the second item on a computer keyboard was recorded. All participants listened to the speech using closed-cup headphones in a quiet room (Frost & Monaghan, 2016). To test speech segmentation, participants completed a forced choice task on preference for word/part-word comparisons. To test novel generalisation, participants completed a similar forced choice task for rule-word/part-word preference.
Data analysis
Analysis included mixed-effects models to allow for random participant and item variability. As all participants responded to both task types, therefore multiple items, the likelihood of correlations in responses from the same participant and to the same item increases. Generalised linear mixed-effects allow for a more flexible approach compared to ANOVA, that can handle missing data better, without significantly losing statistical power. Participant and item variation, the effects of sleep/wake, test type, and sleep stage duration were all considered. The interactions between sleep/wake and test type, and sleep stage duration and test type, were also considered in separate models.
The original experiment was completed by 54 participants, 8 males and 46 females, with an age range of 18-24-years-old (mean age = 18.52). All participants reported being native-English speakers, with no history of auditory, speech or language disorders known. All participants either received university course credit or £20 for completing the experiment. Observations may be excluded for the first linear mixed-effects model. Exclusions may come from participants in the sleep group who did not sleep during the permitted time. This is because the first analysis aims to compare sleep vs. wake. The same participants’ data will be kept for the other linear mixed-effects models which aims to compare duration of sleep stages. This research received ethical approval by Dr Padraic Monaghan and Lancaster University’s Psychology Department on 22/04/2021.
Design
This study had a between-participants design with two conditions: sleep vs. wake between training and testing, and test type. There were two test types of speech segmentation and novel generalisation. Participants were randomly allocated to the sleep or wake conditions, and split evenly, meaning 27 participants slept and 27 remained awake. This study had access to PSG data for 18 of the participants in the sleep group. All participants received both test types. All participants were provided with an information sheet and gave written consent before the study commenced.
Materials
Stimuli
Using the Festival speech synthesiser (Taylor et al., 1998), speech stimuli were created that were based on similar stimuli used by Peña et al. (2002). This artificial training language contained monosyllabic items, of which there were nine (pu, ki, be, du, ta, ga, li, ra, fo), used to form three different non-adjacent pairings with three possible X items in-between (A1X1–3C1, A2X1–3C2, and A3X1–3C3) (Frost & Monaghan, 2016). Using Peña et al.’s (2002) study, A and C items contained plosive phonemes (pu, ki, be, du, ta, go) and X items contained continuants (li, ra, fo). All AXC item strings had a duration of approximately 700ms. Any preferences – for dependencies not due to the statistical structure of the sequences – were controlled for by generating eight versions of the language. Each version had randomly assigned syllables to A and C items, and the same X items were used in all versions. These versions of the language were counterbalanced across both task types. When testing for novel generalisation, three additional syllables were used with continuant phonemes (ve, zo, thi) (Frost & Monaghan, 2016). Research on the similarities in phonological properties of non-adjacent dependent syllables has shown that these similarities show support for acquisition of such nonadjacencies (Newport & Aslin, 2004). Nonetheless, other research has found that they are not essential for language learning to occur (Onnis et al., 2004). Words in the same grammatical category tend to be coherent regarding phonological properties (Monaghan et al., 2007), so regardless of learning, this property of the artificial language used within this study is consistent with natural language, which allows for real-life implications.
Training
The speech stimuli were formed into a 10.5-minute-long continuous speech stream by stringing together the AXC words within the language. It was ensured that no Ai_Ci dependencies were repeated immediately after each other. The speech stream included 5s fades for the onset and offset of speech, which ensured that such a feature of speech could not be used as a language structure cue (Frost & Monaghan, 2016).
Testing
Segmentation: part-words were trisyllabic items that were heard in the training speech stream but overlapped word boundaries. As such, part-words comprised of either the last syllable of one word and the first two syllables of the next (CiAjX), or the last two syllables of one word and the first syllable of the next (XCiAj). For all nine AXC items, both part-word types were created. 18 test pairs were constructed which participants listened to, by matching each part-word with its corresponding word (for example, the A1X2C1 item was paired with the X2C1A2 part-word) (Frost & Monaghan, 2016).
Novel generalisation: nine forced choice tests included a rule-word which contained one of three novel syllables (ve, zo, thi) (AiNCi), where N is the novel syllable and a novel part-word. For each Ai_Ci dependency, each novel rule-word appeared once. Part-words were made of two syllables that were heard in the training task, in their respective positions, with the same novel syllable as in the rule-word sequence (Frost & Monaghan, 2016). This novel syllable could appear in any position (first NCiAj, second XNAi, or third CiAjN) and each novel syllable occurred once in each of these positions. Rule-word and part-word novelty presence controlled for the effect of the novel syllable, yet the novel generalisation task still tested for generalisation of the non-adjacent structure of items within speech (Frost & Monaghan, 2016). Randomisation of test-pairs in all conditions was ensured across all participants, including the position of the correct response in each test-pair, to reduce response bias. When listening to the test-pairs, items in each pair were separated by a 1s pause. All participants completed The Stanford Sleepiness Scale (SSS) (Hoddes et al., 1972). This was in order to note participant sleepiness before the period of sleep or wake. The SSS consists of one item on a scale of seven statements, within which participants were required to select one statement that best described their perceived level of sleepiness (Shahid et al., 2011) (see Appendix A). Participant responses in the testing task were excluded if 90% of responses were always “1” or “2”, or if responses alternated between “1” and “2”.
Procedure
The whole procedure lasted for a three-hour period. For the training task, all participants listened to the continuous stream of speech and were instructed to pay attention to the language and think of possible words it contains. After the training task was complete, participants were split into two groups for the sleep vs. wake condition. Half of the participants, the sleep group, were given an hour and 45 minutes to sleep. These participants slept at Lancaster University Psychology Department’s sleep lab, and their sleep was monitored using polysomnography (PSG). PSG and an Embla N7000 system can record the amount of time spent in each sleep stage, and sleep spindle density, with EEG sites: O1, O2, C3, C4, F3, and F4 referenced against M1 and M2. The other half of participants remained awake for the same duration, watching a non-verbal, emotionally neutral video with neutral music. The testing task was then given to all participants after the same amount of time, 15 minutes after the break period. All participants were then required to complete the testing forced choice tasks. Within each trial, participants listened to a test-pair of items and were instructed to select which item best matched the training language. A response of “1” for the first item or “2” for the second item on a computer keyboard was recorded. All participants listened to the speech using closed-cup headphones in a quiet room (Frost & Monaghan, 2016). To test speech segmentation, participants completed a forced choice task on preference for word/part-word comparisons. To test novel generalisation, participants completed a similar forced choice task for rule-word/part-word preference.
Data analysis
Analysis included mixed-effects models to allow for random participant and item variability. As all participants responded to both task types, therefore multiple items, the likelihood of correlations in responses from the same participant and to the same item increases. Generalised linear mixed-effects allow for a more flexible approach compared to ANOVA, that can handle missing data better, without significantly losing statistical power. Participant and item variation, the effects of sleep/wake, test type, and sleep stage duration were all considered. The interactions between sleep/wake and test type, and sleep stage duration and test type, were also considered in separate models.
Publisher
Lancaster University
Format
Data/Excel.csv
Data/Excel.xlsx
Analysis/r_file.R
Data/Excel.xlsx
Analysis/r_file.R
Identifier
Power2021
Contributor
Brad Hudson
Rights
Open
Relation
Secondary data analysis. Data were originally collected for the paper below, but they were not analysed by the authors.
Frost, R. L. A., & Monaghan, P. (2016). Simultaneous segmentation and generalisation of non-adjacent dependencies from continuous speech. Cognition, 147, 70- 74
Frost, R. L. A., & Monaghan, P. (2016). Simultaneous segmentation and generalisation of non-adjacent dependencies from continuous speech. Cognition, 147, 70- 74
Language
English
Type
Data
Coverage
LA1 4YF
LUSTRE
Supervisor
Prof. Padraic Monaghan
Project Level
MSc
Topic
Cognitive, developmental, neuropsychology
Sample Size
54
Statistical Analysis Type
Linear mixed effects modelling, correlation, sleep data analysis
Files
Collection
Citation
Carly Power, “The Effects of Different Sleep Stages on Language Learning Tasks in Young Adults,” LUSTRE, accessed May 2, 2024, https://www.johnntowse.com/LUSTRE/items/show/144.