dr. patrick rebuschat - amazon simple storage service (s3 ... · pdf fileno difference between...
TRANSCRIPT
Dr. Patrick Rebuschat [email protected]
Overview 1. Definitions
2. Research paradigms
3. Artificial Grammar Learning
4. Implicit and Explicit Learning of L2 Syntax
Linguistics Circle talk on November 23:
John Williams (Cambridge)
What is implicit learning?
Example: • First language acquisition
• Learning process that results in unconscious (implicit) knowledge.
• Generally associated with incidental learning conditions.
What is explicit learning?
Classic example of explicit knowledge: • Pedagogical rules in language classroom
• Learning process that results in conscious (explicit) knowledge.
• Usually associated with intentional learning conditions.
Sequence learning: • serial reaction time task, matrix scanning task, etc. • Nissen & Bullemer (1987); Lewicki, Hill, & Bizot (1988).
Control of complex systems: • Broadbent (1977, 1978); Berry & Broadbent (1984).
Artificial Grammar Learning: • Reber (1967), etc.
Artificial Grammar Learning
Grammar: finite-state or phrase-structure grammar Lexicon: letters, tones, pseudowords, etc. Sample sequences • XXRM, VVTTRXRRM, VTRRRM • pel pel wadim rud, vot wadim tood
Testing phase: 1. Measure of learning:
Grammaticality judgments
2. Measure of awareness: Verbal reports, confidence ratings, etc.
AGL and Implicit learning Training phase: • Exposure to stimuli under incidental learning conditions.
• Subjects are not told that the stimuli are generated by a
complex rule-system.
• Typical training task: Memorization of letter strings.
AGL and Explicit learning Training phase: Option 1: • Tell subjects that there is a rule-system and that they
should discover the rules (= subjects have to acquire explicit knowledge themselves).
• Expose subjects to stimuli.
Option 2: • Provide subjects with rule descriptions • Expose subjects to stimuli.
Testing phase: • Same.
Common Findings Implicit learning
• Experimental group: 65% accuracy in classification task
• Control group (untrained): Chance performance
• Experimental subjects report no (relevant) knowledge of
rule-system, often believe they are guessing when making grammaticality judgments.
• The acquired knowledge is implicit (unconscious).
Common Findings Explicit learning
• Mixed results.
• Implicit vs. explicit: Reber (1976): Superior performance
for implicit learners over explicit learners (subject who were told about the existence of rules). Reber et al. (1980) report the opposite.
• Timing of rule presentation: Reber et al. (1980) suggest conveying explicit knowledge first works best. Mathews et al. (1989) suggest the opposite. Kim & Rebuschat (in prep)
Experiment 1: Rohrmeier, Rebuschat, & Cross (2011, Consciousness & Cognition) Lexicon: 9 tone units
Experiment 2: Rebuschat & Rohrmeier (in prep) Lexicon: 9 syllable units
Sample stimuli Experiment 1 Experiment 2 * *
Subjects • Music-Experiment: 60
– 30 experimental, 30 control (untrained) – Only half the subjects were musicians.
• Language-Experiment: • 15 experimental, 15 control (untrained)
Training phase (Experimental subjects only) • Music-Expt: Exposure to tone sequences • Lang-Expt: Exposure to syllable sequences • Training set: 51 sequences (3 x 17) • Distractor task: Counting units in each sequence
Testing phase (Experimental and control subjects) • Measure of learning: Familiarity judgment task • Measure of awareness: confidence ratings, verbal
reports
In the testing phase, subjects were exposed to three types of stimuli: 1. Old-gram:
Items generated by grammar, repeated from training.
2. New-ungram: Items that violate the grammar. Not previously used in expt.
3. New-gram: Items generated by grammar. Not previously used in expt.
Results
Clear learning effect in both experiments: ● Experimental groups significantly outperformed respective control groups. ● They also performed significantly above chance.
No difference between language and music groups: ● Experimental subjects in both conditions perform similarly across items. ● The nature of the input did not influence the pattern.
Subjects were unaware of rules used to generate stimuli and unaware of having acquired relevant knowledge.
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Old-gram New-gram New-ungram
Endorsement Rates in Familiarity Judgment Task
Mus-Exp Lang-Exp
AGs have been extensively used in language acquisition research.
Examples:
• Braine (1963, 1966), • Smith (1966, 1968, 1969), • Moeser & Bregman (1972), • Morgan & Newport (1981), • MacWhinney (1983), • Morgan, Aslin, & Newport
(1989), • Saffran (1997, 2001, 2002)
• Saffran, Aslin, & Newport (1996),
• Marcus et al. (1999), • Opitz & Friederici (2003, 2004), • Friederici, Steinhauer, &
Pfeiffer (2002) • Gómez & Gerken (1999, 2000) • etc. etc.
Why use artificial languages? Advantages • Control of previous knowledge • Control of input Disadvantage • Potential lack of generalizability • What does the learning of meaningless letter or
pseudoword sequences have to do with language?
Semi-artificial languages as more valid alternatives: Examples: Williams (1999, 2003, 2005), Williams & Kuribara (2007), Williams & Leung (2006), Rebuschat & Williams (2006, 2008, 2009)
Research questions
1. Is there implicit learning in the case of L2 acquisition? Or is there a “fundamental difference” between L1 and L2 acquisition?
2. How is implicit knowledge represented?
Method
Stimulus material Artificial language consisting of English words and
German syntax.
Linguistic focus: Three verb placement rules Each of these is related to a specific syntactic
pattern.
Today BOUGHT John the apple in the supermarket.
Some time ago CLAIMED Jim that his mother Melbourne LIKED.
After his wife a sandwich CRAVED, RUSHED Mike to Japan.
[V2]
[V2-VF]
[VF-V1]
Three syntactic patterns English words, German syntax
Subjects
• 30 NS of English with no background in German • Experimental condition (n = 15) • Control condition (n = 15)
Experimental group: • Exposed to artificial language under incidental
learning conditions. Control group: • Receives no exposure.
35
Exposure phase (experimental group only)
Listen to sentence
Judge semantic plausibility of
sentence
Listen to next sentence
120 sentences 40 [V2] pattern 40 [V2-VF] pattern 40 [VF-V1] pattern
60 semantically plausible items 60 semantically implausible items
meaning-focused task
Testing phase (experimental and control groups)
Listen to sentence
Judge grammaticality of
sentence
Report confidence level
Report basis for grammaticality
decision
37
60 new sentences 30 grammatical 30 ungrammatical
0 = guess 5 = somewhat confident 9 = very confident
V2, V2-VF, VF-V1 *V1, * V3, * V4, * VF, * VF-V2, * V1-VF
G = guess I = intuition M = memory R = rule knowledge
Results: Grammaticality judgments Overall accuracy: • Experimental group: 62% • Control group: 43%
Performance: • Experimental group >> untrained controls • Experimental group >> 50%
Clear learning effect in experimental group.
39
0
10
20
30
40
50
60
70
80
Grammatical Ungrammatical
En
do
rse
me
nt
rate
(%
)
Exp 3
Control
*
Exp
Performance across grammaticality Endorsement rates
40
Experimental group is significantly more likely than control group to endorse grammatical items.
0
10
20
30
40
50
60
70
80
Grammatical Ungrammatical
En
do
rse
me
nt
rate
(%
)
Exp 3
Control
*
Exp
Performance across grammaticality Endorsement rates
41
No significant difference between the two groups on ungrammatical items.
0
10
20
30
40
50
60
70
80
Grammatical Ungrammatical
En
do
rse
me
nt
rate
(%
)
Exp 3
Control
*
Exp
Performance across grammaticality Endorsement rates
42
Overall performance driven by correct endorsement of previously encountered patterns.
Performance across sentence types
Experimental subjects appear to know that: • V2 can occur in
simple sentences. • V3 is not an
option. • VF is a verb
position licensed by grammar.
• VF-V1 is licensed but *VF-V2 is not.
They do not display knowledge of the role of clause type and clause sequence.
0
10
20
30
40
50
60
70
80
90
100
*V1 V2 *V3 *V4 *VF V2-VF *V1-VF VF-V1 *VF-V2
En
do
rse
me
nt
rate
(%)
Sentence type
Exp 3 Control
43
Results: Confidence ratings
1. ZCC: Subjects were more confident in correct decisions than in incorrect ones. They were partially aware of having acquired knowledge.
2. Guessing Criterion: When subjects claimed to be guessing, their performance was at chance level. No unconscious (judgment) knowledge.
Experimental group: Accuracy and Proportions (in Percent) across Confidence Ratings
Guess Somewhat confident Very confident
Accuracy 53 60* 65*
Proportion 12 54 34
Significance from chance: * p < .05.
44
Results: Source attributions
1. When attributing grammaticality decisions to intuition, subjects perform significantly above chance. Evidence for both conscious and unconscious structural knowledge.
2. Native-speaker like intuition:
Knowing that a sentence is grammatical or ungrammatical, with high levels of confidence, but with basis on intuition.
Experimental group:
Accuracy and Proportions (in Percent) across Source Attributions
Guess Intuition Memory Rule
Accuracy 56 59* 57 65**
Proportion 10 32 15 43
Significance from chance: * p < .05. ** p < .001.
45
Results: Verbal reports • Subjects report noticing that verbs can occur at the
end.
• No verbalization of three rules described earlier.
• No examples of licensed sentences.
• No relevant conscious knowledge according to verbal reports.
• Vs. confidence ratings...
46
Experiment 1: Summary Adult learners are able to acquire L2 syntax: very rapidly (30 mins of exposure) without intending to without feedback while processing sentences for meaning.
Subjects were aware of having acquired knowledge, but they did not know what this knowledge was.
Phenomenology of intuition: Similar to native-speakers.
Experiment 1: Summary Adult learners are able to acquire L2 syntax: very rapidly (30 mins of exposure) without intending to without feedback while processing sentences for meaning.
Subjects were aware of having acquired knowledge, but they did not know what this knowledge was.
Phenomenology of intuition: Similar to native-speakers.
What do subjects learn? • Subjects were able to classify strings they had never
encountered before.
• Classification performance: Above-chance performance only on grammatical patterns.
49
Subjects acquired abstract linguistic knowledge as a result of exposure.
Exposure results in memory for syntactic patterns, not rules in the traditional sense.
Do subjects retain this knowledge? Grey, Williams, & Rebuschat (submitted): • Found retention in delayed post-test (1 week)
• Subjects performed just as well on the delayed post-
test as on the immediate post-test:
59% 59% - without any additional exposure.
50
Yes!
Methodological observation • Experiment confirms that reliance on verbal reports is
clearly inadequate in order to distinguish implicit and explicit knowledge.
• Analysis of the verbal reports showed that participants were unable to verbally describe the rules of the semi-artificial system. This would have supported the (wrong) assumption that learning in experiment 1 resulted predominantly in unconscious knowledge.
51
Methodological observation • We need to use more sophisticated methods to assess
the conscious or unconscious status of the acquired knowledge.
• Subjective measures of awareness useful tool.
• Rebuschat (forthcoming). Measuring implicit and explicit knowledge in second language research. Language Learning.
52
• Rebuschat & Williams (2009, 2011) found that subjects acquire relatively abstract knowledge while exposed to the artificial language.
• Subjects can judge sentences they have never encountered before.
How abstract is this knowledge?
Method
Stimulus material Same artificial language as expt. 1
Subjects 60 NS of English with no background in German Four experiments groups, n = 15 each.
Exposure phase: Plausibility judgment task
• All groups are presented with the same 120 sentences and asked to judge semantic plausibility.
Difference: • VV and VA groups are presented with written input, • AA and AV groups are presented with auditory input.
Group Exposure Testing
VV (n = 15)
Visual Visual
AA (n = 15)
Auditory Auditory
AV (n = 15)
Auditory Visual
VA (n = 15)
Visual Auditory
Testing phase: Grammaticality judgment task
• All groups are presented with the same 60 sentences and asked to judge whether or not they follow the artificial language.
• VV and AV groups judge written stimuli. • AA and VA groups judge auditory stimuli.
Group Exposure Testing
VV (n = 15)
Visual Visual
AA (n = 15)
Auditory Auditory
AV (n = 15)
Auditory Visual
VA (n = 15)
Visual Auditory
Modality change
Same modality
Results
• Significant learning effect in all groups.
• No significant differences between groups.
Changing modality from visual to
auditory or auditory to visual does not affect performance at test.
Results confirm that subjects build
up a knowledge base that is very abstract.
Group Accuracy
VV (n = 15)
62%
AA (n = 15)
62%
AV (n = 15)
59%
VA (n = 15)
59%
Thus far, we have seen that:
• adult subjects can acquire knowledge incidentally and without becoming aware of the knowledge they have acquired.
• the knowledge they acquire is quite abstract and consist of memory for syntactic patterns.
• Let’s look at explicit learning now.
Rebuschat & Williams (in prep) Compared incidental and intentional learning of artificial language
Subjects: • 30 NS of English, with no background in German
1. Incidental group (n = 15) 2. Intentional group (n = 15)
Exposure: • Both groups listen to the same 120 sentences. • Auditory exposure, as in expt 1.
Difference: Instructions on how to interact with the sentences • Incidental group:
• Plausibility judgments, not told about rule system
• Intentional group: • Rule-search task, told about existence of rule
system
Testing phase:
• Grammaticality judgments, 60 new sentences
• Same as previous experiments.
Listen to sentence
Judge grammaticality of
sentence
Report confidence level
Report basis for grammaticality
decision
Results: Grammaticality judgments
Overall accuracy: • Intentional group: 77% • Incidental group: 62% • Control group: 44%
• Intentional group
outperformed incidental group and untrained controls.
• Intentional group performed significantly above chance.
0
10
20
30
40
50
60
70
80
90
100
Mean Accuracy
Intentional (Exp 2) Incidental (Exp 1) Control
*
*
Improved performance under rule-search conditions.
65
Performance across grammaticality
• Significant difference between intentional and incidental group on grammatical and ungrammatical items.
• Classification performance is not categorical.
0
10
20
30
40
50
60
70
80
90
100
Gram Ungram
Intentional (Exp 2) Incidental (Exp 1)
*
*
66
Performance across sentences: Intentional vs. incidental subjects
Intentional group is • sig more likely to
endorse V2 sentences,
• sig less likely to endorse *VF sentences.
• But performance on *VF is not significantly below chance.
• “Verb-final rule” more difficult to acquire?
• Possible match with developmental data.
0
10
20
30
40
50
60
70
80
90
100
*V1 V2 *V3 *V4 *VF V2-VF *V1-VF VF-V1 *VF-V2
Intentional (Exp 2) Incidental (Exp 1)
*
*
*
*
* * *
*
*
*
67
Results: Learning effect • Significant learning effect in both groups.
• Intentional group outperformed incidental group:
Instructing subjects to look for ‘rules’ promotes learning.
• Interestingly, this holds even subjects are not actually able to figure out the rules (only 4 subjects actually did).
Results: Confidence ratings
1. ZCC: Subjects were equally confident in correct and in incorrect decisions. They were partially unaware of the knowledge they acquired.
2. Guessing Criterion: When intentional learners claimed to be guessing, their performance was significantly above chance. Unconscious (judgment) knowledge.
Intentional group: Accuracy and Proportions (in Percent) across Confidence Ratings
Guess Somewhat confident Very confident
Accuracy 80* 74* 81*
Proportion 6 56 38
Significance from chance: * p < .05.
69
Results: Source attributions
• When attributing grammaticality decisions to guessing and intuition, subjects perform significantly above chance.
• Evidence for both conscious and unconscious structural knowledge (under intentional learning conditions).
Intentional group:
Accuracy and Proportions (in Percent) across Source Attributions
Guess Intuition Memory Rule
Accuracy 65.9+ 68.9** 61.9* 83.6**
Proportion 4.6 21.8 13.1 60.6
Significance from chance: + p < .06 * p < .05. ** p < .001.
70
Results: Verbal reports • Only four subjects were able to give accurate verbalizations
(in this case, provide examples of licensed sentences).
• These subjects performed categorically on the grammaticality judgment task (90% endorsement of grammatical sentences, 90% rejection of ungrammatical items)
• Evidence for rule application/knowledge in these subjects.
• The verbal reports of the remaining subjects were similar to the ones of the incidental learners. No relevant knowledge was reported.
71
Summary: Experiment 3 • Significant learning effect in incidental and intentional
group.
• Intentional group outperformed incidental group: Instructing subjects to look for ‘rules’ results in more learning.
Summary: Experiment 3 • Measures of awareness suggest that both incidental and
intentional learners develop implicit and explicit knowledge.
• But how much of each they develop depends on learning context!
• “Explicit learning experiments” do not necessarily result
in explicit knowledge. “Explicit instruction” can also result in implicit knowledge.
• It is important to measure implicit and explicit knowledge.
Experiment 4
Tagarelli, Borges Costa, & Rebuschat (2011)
Individual differences
in implicit and explicit learning
Individual differences:
• Do variables like intelligence, motivation, learning style, personality, working memory capacity, etc. affect the way we learn language?
• How do these variables affect L2 acquisition?
Tagarelli, Borges Costa, & Rebuschat (2011) • Investigated how WM capacity affect implicit and
explicit learning of languages.
• Does working memory capacity (WMC) affect an individual’s ability to learn the syntax of a novel language?
• Does WMC have a differential effect on implicit and explicit learning?
Subjects: NS of English, no German language background • Incidental group (n = 31) • Intentional group (n = 31)
Working Memory Assessment 1. Operation-word span task 2. Letter-number ordering task
Letter-number ordering task
Participant hears… A-7 9-C-3 W-1-K-5 8-A-6-G-1 W-8-J-5-F-3 R-3-A-4-Z-1-G 7-K-2-S-6-F-1-Z
Participant says… 7-A 3-9-C 1-5-K-W 1-6-8-A-G 3-5-8-F-J-W 1-3-4-A-G-R-Z 1-2-6-7-F-K-S-Z
Exposure: • Same as experiment 3. • Incidental group: Plausibility judgments • Intentional group: Rule-search task Testing phase: • Same as previous experiments. • Grammaticality judgments • 60 new sentences
Results
• Significant learning effect in both groups.
• Intentional group outperformed incidental group: More learning under rule-search conditions.
Replicates Rebuschat & Williams (2009, 2011) and Rebuschat & Williams (in prep)
Results: Working Memory
1. No correlation between WM capacity and learning in the incidental group.
2. Significant correlation between WM capacity and learning in the intentional group.
• Results support distinction between implicit and explicit learning.
• Results supports language teaching approaches that emphasize implicit learning, e.g. TBLT.
1. Adult learners are able to acquire L2 syntax without intending to and without becoming aware of the knowledge they have acquired.
2. The acquired knowledge is abstract: subjects are able to rate completely novel sequences, and changing modality from exposure to test does not affect this ability.
3. Subjects do not acquire ‘rules’ in the traditional sense, but rather syntactic patterns.
88
4. The relationship between learning condition and the acquired knowledge is complex. Subjects often acquire both implicit and explicit knowledge.
5. Working memory capacity plays a role when subjects are exposed under explicit learning conditions, but not under implicit learning conditions.
6. Adding an explicit component (e.g. rule presentation) to an otherwise implicit learning task can result in greater learning effect.
89