A Comparison of the Effects of Soothing and Cacophonous Music on Implicit Learning and Transfer Victor Vuchic – vuchic@stanford.edu Brandon Suzuki – bsuzuki@stanford.edu 1 Abstract In the research for this paper we made two hypotheses. The first hypothesis was that the presence of musical consistency during learning and transfer assessment would improve implicit transfer results. We believed that subjects who both learned and attempted transfer while listening to the same music would perform better than those listening to different music during those tasks. The second hypothesis was that soothing classical music would help implicit learning and transfer to a greater extent than cacophonous music. We believed that soothing music would prevent the conscious mind from interfering with implicit learning without overly taxing cognitive load. We supposed that cacophonous music, which also served to distract conscious thought, would overly tax cognitive load resulting in decreased implicit learning transfer performance. Using Markov chains in the manner of Reber (1967) we trained subjects on examples of a particular grammar in the guise of a memory test while having them listen to either a soothing classical or cacophonous jazz music. Subjects were then informed of the grammatical nature of the training sentences and given the task of identifying grammatical words from a list while listening to one of the two music pieces used during the learning stage. Due to what we believed to be flaws in our post test method, the results did not show the expected implicit learning and transfer effect; therefore, we were unable to draw conclusions directly addressing our hypotheses. Introduction Studies have shown that environmental consistency aids transfer. In the case of music, a student who studies while listening to music is more likely to do better on the test of the material if listening to the same music during the test. Our study looks at whether implicit learning transfer works similarly to conscious learning transfer with respect to auditory environmental consistency. We also examined whether the type of auditory context provided had an effect on implicit learning transfer. At a more general level, findings in our study could have implications for the situative perspective relative to implicit learning both in terms of the impact of auditory environmental consistency and the effect of different types of auditory environments. The situative perspective considers the environment of the learner to be tightly linked to learning. In this study we are looking at a small subset of the environment, the auditory domain. Specifically we are examining two kinds of music. A finding of increased implicit learning transfer would show that the environment does affect implicit learning and therefore support the situative perspective. We used Markov chains based on Reber (1967) to train and test for implicit transfer. During the learning process subjects would implicitly learn the underlying grammar (same one used by Reber 1967) under the guise of a memory recall test. In order to test transfer, subjects would be informed of the grammatical consistency of the words they had just memorized and asked to identify grammatical sentences from a list of words containing equal amounts of previously memorized grammatical words, new 2 grammatical words, and subtly ungrammatical words. We intended to use the learning time and results of the post test as the basis for drawing conclusions relative to our hypotheses. General Description of Experiment Our independent variable was the audio accompaniment during the learning and post test stages. We chose audio accompaniment because it is something that we believe is relatively common in learning situations. People often have the radio on in the background or music playing while they do other activities such as read, study, drive, or cook. The audio accompaniment consisted of two musical pieces: one soothing and one cacophonous. We also included a control condition where no audio accompaniment was used in either stage as a baseline. In total we ran subjects in five conditions: Condition Number of Subjects 1 4 2 4 3 4 4 5 5 1 Table 1 – Study Structure Learning Stage Post Test Stage Soothing Soothing Cacophonous Cacophonous None Soothing Cacophonous Soothing Cacophonous None Our dependent variables include time to learn, repetitions of sentence sets during learning, and recognition of the three types of items in the post test: seen grammatical, new grammatical, and non-grammatical sentences. The time to learn and repetitions were chosen because they provide information about how the subject learned the implicit grammar (if there was an effect). The results of recognition of the three types of sentences in the post test are important because the “seen grammatical” items indicate how well students remembered sentences, the “new grammatical” items indicate the extent of implicit learning transfer, and the “non-grammatical” items provides information regarding the amount of negative-transfer. If subjects who have the same music for both stages (table 1: conditions 1 and 4) perform better on the post test this would imply that audio consistency in the environment improves implicit learning transfer. This finding would support our first hypothesis regarding consistent music and its positive effect on implicit learning transfer. On the other hand, if subjects in conditions with different audio accompaniment in each stage (table 1: conditions 2 and 3) did just as well or better, that would disprove our first hypothesis. If subjects listening to soothing music in both stages outperform subjects listening to cacophonous music in both stages, this would imply that soothing music increases implicit learning transfer relative to cacophonous music. This finding would support our 3 second hypothesis of soothing music increasing implicit learning transfer relative to cacophonous music. It would also suggest that it may be worth while to take care in choosing the kind of music a learner listens to while learning. Comparing the average post test results on new grammar sentences between conditions sharing the same learning stage audio accompaniment (table 1: conditions 1 and 2, 3 and 4), we can infer a particular audio accompaniment aided implicit learning if its average is higher than that of the other. Otherwise, if both results are the same, it would imply that the kind of audio accompaniment does not affect implicit learning. Comparing the average post test results on new grammar sentences between conditions sharing the same post test stage audio accompaniment (table 1: conditions 1 and 3, 2 and 4), we can infer a particular audio accompaniment aided transfer of implicit learning if its average is higher than that of the other. Otherwise, if both results are the same it would imply that the kind of audio accompaniment does not affect transfer of implicit learning. If all audio accompaniment conditions (table 1: conditions 1 – 4) have worse results in the post test than the control condition (table 1: condition 5), this would imply that audio accompaniment hinders implicit learning and transfer. Method The method for our study closely follows the method used by Reber (1967). Participants The study had 18 subjects. Participants were adults between the ages of 22 to 69, and were chosen based on availability and convenience. Materials The materials for the study included the creation of the 16 strings for the memorization exercises as well as an additional 16 strings for the post test. These strings are based on the Markov chains used in Reber’s study (see Fig 1). See appendix for actual strings used for memorization. 4 Figure 1. Markovian grammatical structure The following methods were used for generating the grammatical and non-grammatical strings: Grammatical The grammatical sentences were generated by randomly generating grammatical strings and adding any unique ones to an array until the maximum number possible was generated. We believe a transition tracking tree traversal method could also have been used which would not generate duplicates, but since the number of grammatical sentences we needed was small we went with the quicker implementation. Non-grammatical Aside from 4 random sentences, Reber (1967) used subtly different sentences for his nongrammatical post test entries. In an attempt to do the same we placed 3 constraints on the non-grammatical entries. 1. Same range of lengths as the grammatical items 2. The first and last letters are always consistent with the grammar 3. Only letters that belonged to the grammar were used In addition, for the soothing music we cut a 6 minute music clip of the Adante movement from Franz Schubert’s The Trout and looped it. For the cacophonous music we cut a 6 minute non-vocal section of John Coltrane’s OM and looped it. We then created the post-test which consisted of 39 strings divided equally between strings the subjects had memorized in the learning phase, strings that the subject had not seen but fit within the grammar, and strings that were not grammatically compliant. Finally, we developed a flash based program that would run on any PC to take the subject through the learning and post-test process and a PHP/HTML based framework to handle structure and file input/output. Design The design of our study was a classic 2 x 2 with a single control subject. 5 Learn Strings w/Classical Learn Strings w/Jazz Table 2 Study Design Post-Test w/Classical Condition 1 Condition 3 Post-Test w/Jazz Condition 2 Condition 4 We used the 2x2 (see Table 2) to attempt to isolate the causality of 2 potential phenomena in the results. Condition 1 (4 subjects) and Condition 4 (5 subjects) listened to the same music for both the learning phase and the post-test phase, while Condition 2 (4 subjects) and Condition 3 (4 subjects) changed music for each section. This allowed us to counter-balance the music the subjects listened to and isolate its effect on implicit learning. Our dependant measures were the number of repetitions per memorization exercise, total time per memorization exercise, and the measured implicit learning effect judged by the subject’s ability to correctly evaluate the strings on the post-test. Procedure Our procedure was for the subject to go through the study on a PC using a Flash based program we developed. The subject was told they are going to go through four sets of memorization exercises while listening to music. The subjects sat down at the PC, put on the headset, adjusted the music to a comfortable volume and pushed the start button. They were then shown 4 strings for 5 seconds apiece followed by a prompt to type in the strings that they just saw. They would then be told which were correct and which were not correct and if they made a mistake they would go through seeing all 4 strings again and reenter them. The subject would repeat this until they could accurately reproduce the 4 strings, and then they would continue to the next set of four strings. Once the subject finished the last set of strings, they were told that all of the strings they had memorized followed a grammatical structure and were formed based on a rule. At this point subjects in conditions 2 and 3 had the music change from classical to jazz or jazz to classical. For conditions 1 and 4 we simply restarted the song they were listening to. The subject then continued to the post test and was presented with 39 strings. They were asked to identify any of the strings they thought followed the rule. Once they finished selecting all of the strings, they clicked submit and the program terminated. We then followed up with a post study interview with the following three questions: 1. Were they looking for a grammatical structure to the strings in the learning phase and if so, did they see any? 2. How did they approach selecting the strings in the post test? 3. Did they know the music they were listening to? Coding The data we captured focused on two areas. First we captured how much time it took them to go through each section of the memorization phase and how many repetitions the subject had to go through for each memorization set. This gave us a 6 picture of any possible effects of the music on the memorization phase. The second area we coded was for the post test. We tracked how many strings the subject correctly identified that they had seen in the memorization phase, how many grammatically correct strings they identified correctly that they had not seen, and finally how many strings they correctly did not select that did not fit the rule (false positives). We combined these three categories into a score for total correct identifications out of 39 strings. Results The results did not provide any conclusive evidence to support or negate our hypothesis due to the fact that we did not get any clear implicit learning effect. The average scores (out of a possible 13) by categories for the post test were 6.7 (52%) correct identifications of strings seen in the memorization process, 2.8 (22%) correct identifications of new grammatically correct strings, and 12 (92%) correct identifications (by not selecting) of non-grammatical strings. The results of the total correct identifications showed that none of the categories performed as well as the control subject. Total correct identifications includes the number correctly selected as fitting the rule for both strings the subjects had seen and not seen, plus the number of non-grammatical strings that were correctly not selected as fitting the rule. There was a trend in that on average the conditions that changed music for the post test did better (22.5 out of 39) than the conditions that had the same music during the post test (20.3 out of 39) (see Fig 2). Total Correct IDs by Music Context for Learning and Post Test 30 25 20 15 Total Correct IDs 10 5 0 None-None Same Music Different Music Fig 2 Total Correct IDs by Musical Context The results did show some trends on how the music impacted the way people learned the material even though the results did not show any implicit learning effect from the post test. None of the subjects were able to identify grammatically correct strings that were not in the memorization tasks at a greater rate than chance. We did see 7 some effect on the learning patterns as well as a trend in false positives. The data shows us that on average, the subjects listening to classical music required less repetitions to memorize the strings than the control subject did. We also see that subjects listening to the cacophonous music during learning required more repetitions (see Fig 3) through the 4 strings to learn them than the classical music or the control subject. Average Repetitions by Music Category 25 20 15 Avg Repetitions 10 5 0 No Music Classical Cacophonous Figure 3. Average Repetitions by Music Category Even though the classical music listeners required fewer repetitions to memorize the strings, there was no difference or correlation in the amount of time it took any of the subjects to learn the material. Another area that showed an interesting dynamic was in the occurrence of false positives during the post test. The subjects that listened to cacophonous music during learning as well as the post test had a higher rate of false positives in that on average they had the highest rate of selecting strings that did not fit the grammatical structure. There were some more qualitative patterns that we found from interviewing our subjects afterwards. Four of them used grouping identification as a strategy in the memorization, for instance they would look for groups of letters that recur in the beginnings or ends of the strings. Also, subjects that listened to the cacophonous music during the learning stage tended to develop headaches by the end of the study. Discussion The below random results from our post test indicates a lack of an implicit learning transfer effect. We believe this finding is a result of a flawed method. By including seen grammatical items in the post test and showing all post test items on a single screen there was not enough incentive for subjects to risk selecting items beyond 8 the ones they explicitly remembered. Interviewing subjects after the post test many indicated they only specified as grammatical the sentences they explicitly remembered from during the learning and did not feel the need to take more risk. If subjects only selected those they explicitly remembered we would expect 100% for the nongrammatical, 0% for the transfer, and some middling percentage representing how well people remembered sentences from the learning stage for the seen grammatical sentences. This trend is visible in the data when looking at the average results for each post test sentence type. On average subjects got 92% of the non-grammatical items, 52% of the memorized items, and only 22% of the new grammatical items correct. We suspect the results showing that repetition count for subjects learning with soothing music was less than for the control and cacophonous music learning conditions is indicative of higher mental load associated with listening to cacophonous music than soothing music. We observed more frequent and stronger verbal expressions of frustration during the learning stage and interviewed subjects often spoke of the irritation and distraction caused by the cacophonous music. We propose the higher rate of negative-transfer by subjects in the cacophonouscacophonous condition (5) is caused by the higher mental load associated with the cacophonous music. Subjects in this condition indicated that by the end of the learning stage their head hurt and they just wanted to get done with the post test. We hypothesize that by the time they reached the post test they did not want to exert more mental effort and essentially didn’t care as much as other subjects so were willing to take more chances. This being the case we would expect a higher average result on new grammatical items as well which is the case in the data. An alternative interpretation for the lack of implicit transfer effect could be that implicit learning is hurt by soothing or cacophonous audio accompaniment. This is somewhat supported by the data since the single control subject got 31% of the new grammatical sentences right while the average for all audio accompanied subjects was 21%. However, since there was only one subject run in the control group and she was significantly older than the other subjects and had formal training in linguistics this data point is fairly weak. The significantly longer time taken by the subject during the post test and the interview also indicated that she had attempted a systematic identification of the grammar once alerted to it during the post test based on the items recalled from the learning stage, something no other subject reported doing. If we were to believe that the audio accompaniment caused the below random results for implicit learning rather than a flaw in the method as we suspect and mentioned above, it is possible to draw some other conclusions from the data. The finding that conditions with the same music during the learning and post test stages (table 1: conditions 1 and 4) did worse than those with differing music indicates that having musical consistency during the implicit learning and transfer of implicit learning stages may actually be detrimental. This goes directly against our first hypothesis. Given the fact that subjects performed below random, it is a story of less 9 detrimental as opposed to more beneficial. One possible, though we feel unlikely, explanation would be that the music interfered with implicit learning causing subjects to learn incorrect rules. Since the average post test results for new grammar sentences between conditions sharing the same post test stage music (table 1: 1 and 3, 2 and 4) are not statistically different, it suggests that the music type does not affect transfer of implicit learning. This is the opposite of our second hypothesis. However, we believe the previously mentioned method flaws are a more likely explanation. It could be argued that due to floor effects, post test new grammar item result comparisons are not valid. Since there have been many other tests which found implicit learning effects such as Reber (1967), we believe once the method has been fixed implicit learning transfer results would be high enough to avoid floor effect problems. Repetition count may not be a good measure of learning since it doesn’t account for learning while confirming memorized sentences and studying those that were right and wrong. As an example, a strategy might be used in which users don’t enter anything the first few repetitions of a set until they feel they have a better chance of getting them all right which may make repetitions seem higher than the average. A measure taking into account both time of each repetition and repetition count may be a more accurate measure of learning. General Discussion The two issues we chose to explore in this study are the effect of musical context on implicit learning transfer and the impact of the kind of music someone listens to on their ability to transfer implicit learning. The results from our study suggest that maintaining a consistent auditory context actually has a worse effect on implicit learning transfer than does listening to different kinds of music. This was demonstrated by the higher performance on the post test by the groups that listened to different music relative to the groups that listened to the same music. In relation to the second issue, we found that the kind of music one listens to does not have an effect on their ability to transfer. This was shown by the similar performance of all groups on the post test and their inability to transfer much of what they learned. While these results show some interesting dynamics, we think that they were influenced by flaws in the design of our study. Analyzing the information gathered in our post-study interviews, we believe that the redesign of some of the core components of the study could yield different results. The two areas that would benefit from design modifications are the control group selection and the post test exercise. We recommend using a larger and more carefully selected control group in order to provide more accurate baseline data to better assess the overall impact of music on implicit learning transfer. As for the post test, we recommend showing the subjects one string at a time and forcing them to say whether the string fits the grammar rule or not. Showing the words one at a time would make the subject explicitly address each string rather than just selecting a few 10 that they know and moving on. This was especially the case for subjects listening to cacophonous music during the post test, they simply wanted to finish the study as quickly as possible. We think that our study showed some interesting dynamics in relation to the impact of any type of music on implicit learning. The data does not show much of an effect on what kind of music you listen to, but it does show a potentially significant negative effect on how listening to music has a detrimental effect on implicit learning transfer. We think that this may be an interesting area to explore further by performing a similar study with people listening to music or not. A study such as this could impact the way educators manage sound and noise in the classroom. If any sound takes a significantly heavier cognitive load, then there should be more care taken on the auditory design of learning environments. In addition, parents could then substantiate telling their kids to turn there Ipods off when they study! References Reber, A. S. Implicit Learning of Artificial Grammers, Journal of Verbal Learning and Verbal Behavior 6 1967 Pp. 857-863 11 Appendix of Materials Appendix 1 Learned Grammar List for Pre-training Set 1 (20) VVS VXVPXVS TTXXVS VXVS Set 2 (22) TPPPTS TTS TTXVPXVS TPPTS Set 3 (25) TPTS VVPXXXVS VVPXVS TPPTXVS Set 4 (24) VXXVPS TTXXVPS VXXVS TPPPTXVS 12 Appendix 2 Post test results Post test results Control Condition 1 – Class, Class Condition 1 Condition 1 Condition 1 Condition 2 – Class, Jazz Condition 2 Condition 2 Condition 2 Condition 3 – Jazz, Class Condition 3 Condition 3 Condition 3 Condition 4 – Jazz, Jazz Condition 4 Condition 4 Condition 4 Condition 4 Average by condition Seengrammatical Grammatical Nongrammatical Variance by condition Seengrammatical Grammatical Nongrammatical Seengrammatical Grammatical Radmila Vuchic 9 4 13 26 cory lee david Jason Nicole 9 2 9 7 1 0 4 5 13 13 9 13 23 15 22 25 Bruce Jennifer Fan Laura 2 Lili Gill 5 4 13 22 4 8 8 4 5 1 13 12 13 21 25 22 Aneto grant suzuki Greg Warman Yeong Haur Kok 7 6 1 2 13 13 21 21 7 5 12 24 7 4 12 23 Debbie Dolores Richard Sava Saheli Singh Sherwin Average 8 7 6 5 0 0 5 13 13 18 20 19 6 5 6.666666667 4 2 2.833333333 11 12 12 21 19 21.5 Control ClassicalClassical Nongrammatical Total ClassicalCacophonous CacophonousClassical CacophonousCacophonus Overall 9 4 6.75 2.5 6.25 3.5 6.75 3 6.4 2.2 6.66667 2.83333 13 12 12.75 12.5 10.8 12 ClassicalCacophonous CacophonousClassical CacophonousCacophonus Control ClassicalClassical Overall 0 0 10.91666667 5.666666667 4.25 3 0.25 3.333333333 6.4 5.2 6.47917 3.06771 0 4 0.25 0.333333333 11.2 11.0469 13 Learning Timings (in seconds) Condition 0 1 1 1 1 2 2 2 2 3 3 3 3 4 4 4 4 4 Avg Learn Times 1 334 300 291 407 447 106 356 89 105 546 151 109 141 204 249 333 743 99 Radmila Cory David Jason Nicole Bruce Jennifer Laura Lili Aneto Grant Greg Yeong Debbie Sherwyn Dolores Richard Sava 2 476 169 417 304 285 313 212 275 188 226 443 338 184 311 342 413 181 288 3 479 349 422 292 372 267 270 438 198 146 361 353 268 269 214 485 235 391 4 762 301 415 115 277 255 610 400 400 230 921 517 227 509 269 344 346 500 Total Learning 2051 1119 1545 1118 1381 941 1448 1202 891 1148 1876 1317 820 1293 1074 1575 1505 1278 1 2 3 4 5 2051 1290.75 1120.5 1290.25 1345 Post Test 1079 240 93 186 155 46 90 185 183 119 345 147 439 106 56 88 129 61 14 Total 3130 1359 1638 1304 1536 987 1538 1387 1074 1267 2221 1464 1259 1399 1130 1663 1634 1339 Repetitions Section Radmila 1 4 2 5 3 5 4 8 Total repetitions 22 Cory David Jason Nicole Bruce Jennifer Laura Lili Aneto Grant Greg Yeong Debbie Sherwyn Dolores Richard Sava 4 3 6 6 2 7 1 2 9 2 2 2 3 4 6 13 2 3 3 5 3 5 3 3 2 3 6 4 2 3 5 5 3 5 6 5 5 6 5 5 5 3 3 5 5 4 4 4 9 5 8 5 5 3 4 5 11 6 7 5 19 7 5 8 6 7 8 10 18 16 19 19 17 26 15 14 20 32 18 13 18 19 27 29 25 No Music Average Repetitions 22 Classical Cacophonous 18 22.33333333 15