Perceptions on L2 fluency-perspectives of untrained raters by The Foreign Language Assessment Group CFP meeting 06 March 2009
Human Raters’ Perceptions Classroom recording (video + audio) Annotation of Speaker Turn Transcription Random Selection Extracted Audio files Automated Phone Segmentation Manual Checking Data Analysis Human Raters’ Perceptions Data Analysis Data Analysis
Status of the two rating studies Summer & Fall 2008 1st Rating study 38 untrained raters from different Mandarin-speaking regions Data collection completed Data analysis currently underway Spring 2009 2nd Rating study Target 10-20 participants (i.e. trained raters) from linguistic-related disciplines Determine factors that affect raters’ perception on L2 fluency, visual impacts on rating performance and influence of variety of Mandarin
Research Purpose To explore untrained raters’ rating patterns and their perceptions on L2 fluency To compare raters’ performance between untrained and trained raters
Research Questions What kinds of rating patterns do untrained raters show? Which assessment criteria predict the L2 fluency? Are there any interaction effects between visual and audio inputs based on the rating results by untrained raters? What are implications for the automated speech recognition tool from these results?
Research Procedures (1) Target raters 38 Native Speakers of Chinese “Untrained” people in rating Mini-Training session Familiarization: 1.5 Hours Workshop Brief directions on rating scales descriptors /the rating procedures
Research Procedures (2) Session: First Rating Second Rating → Two Weeks → Type: Audio Video Audio Video R1- R19 R20-R38 R20-R38 R1- R19 120 120 120 120 120 120 120 120
Rater Procedures 38 Native Speakers of Chinese 6 or 7 assessment criteria used depending on method types Web-rating tool used Note1: e.g. Disfluency, Pronunciation, Nativeness, Communication, Syntax, Lexicon, Gesture Note 2: e.g. Audio/ video
Web-Rating Frame used
Methodology Target Raters Data Analyses Rating results of 33 untrained raters were analyzed Data Analyses Descriptive Statistics, Correlation analysis Analysis of Repeated Measures to look at the audio/visual interaction effects Logistic Regression
Q1.What kinds of rating patterns do untrained raters show? Session Type Dis Pron Natn Comn Synta Lexic Guest First Rating Audio 2.89 3.01 2.51 2.98 3.10 2.83 Visual 2.80 2.97 2.55 2.99 3.04 2.81 2.60 Mean 2.85 2.53 3.07 2.82 Second 2.76 2.48 2.92 2.91 2.77 2.87 2.68
Severity level (Mean of ratings)
Correlation Analysis (1) First Round Fluent2 Disqual Pron Natness Comm Syntax Lexicon Gesture 1 .689(**) .660(**) .698(**) .646(**) .709(**) .760(**) .691(**) .752(**) .806(**) .739(**) .633(**) .697(**) .743(**) .700(**) .775(**) .748(**) .750(**) .822(**) .783(**) .425(**) .495(**) .526(**) .503(**) .577(**) .567(**) .580(**)
Correlation Analysis (2) Second Round Fluent Disqual Pron Natness Comm Syntax Lexicon Gesture 1 .705(**) .659(**) .729(**) .660(**) .715(**) .769(**) .694(**) .762(**) .799(**) .746(**) .672(**) .725(**) .776(**) .726(**) .835(**) .679(**) .740(**) .738(**) .808(**) .805(**) .373(**) .446(**) .471(**) .439(**) .501(**) .482(**) .484(**)
Q2.Which assessment criteria can predict fluency(Logistic_1)? Rater Group 1(second) Group 2 (first) Video Exp(B) Sig. Disqual 4.552 0.000 * 4.293 Pron 1.351 0.079 2.077 Natness 2.361 1.902 Comm 1.728 0.004 * 2.092 Syntax 0.001 * 1.819 Lexicon 1.567 0.005 * 2.351 Gesture 1.108 0.391 0.932 0.578
Q2.Which assessment criteria can predict the fluency(1_2)? Group 1(S2): Y=-11.358+ 1.515X1+0.301X2+0.859X3+0.547X4 +0.643X5+0.449X6 + 0.102X7 Group 2(S1): Y=-12.620+1.457X1+0.731X2+0.643X3+0.738X4 +0.598X5+0.855X6+ -0.070X7
Q2.Which assessment criteria can predict the fluency(Logistic_1)? Rater Group 1(first) Group 2(second) Audio Exp(B) Sig. Disqual 3.632 0.000 * 6.085 Pron 1.476 0.004 * 5.002 Natness 2.419 1.051 0.715 Comm 1.500 0.008 * 1.914 Syntax 1.817 1.361 0.069 Lexicon 1.584 0.002 * 1.918
Q2.Which assessment criteria can predict the fluency(2_2)? Group 1(S1): Y=-10.188+1.290X1+0.389X2+0.883X3+0.405X4 +0.597X5+0.460X6 Group 2(S2): Y=-12.606+1.806X1+1.610X2+0.050X3+0.649X4 +0.308X5+0.651X6
Q3. Are there any interaction effects between visual and audio inputs based on the rating results by untrained raters? Within-subject Effects Source F Sig. Main Effects Session *7.190 0.007 Input Type *6.127 0.013 Interaction Effects Session*Type *12.252 0.000
Findings (1) Gestures show relatively low correlation with Fluency in both rating sessions. Gestures and Pronunciation are variables that do not predict the fluency level in G1 (video samples). Nativeness and syntax do not predict the fluency level in G2 (audio samples).
Findings (2) Interaction effects are significant. Implies that raters show different rating patterns depending on the rating session and input types (audio/video).
2nd Rating Study of Trained Raters Comparisons with rating results of 1st untrained raters Find differences in ratings between two different groups
Methodology of 2nd rating study Rating scale 6-7 Assessment criteria used in the 1st study Same speech samples used in the 1st study Visual/audio effects on rating Same input types used in the 1st study Same rating procedures used in the 1ST study Individual raters’ rating patterns
Raters Target raters Propose Rater training model for trained raters 10 - 20 Native Speakers of Chinese Teaching Experience at UIUC and other area Propose Rater training model for trained raters
Rater Training Model Actual Ratings Training Workshop STEP 1 STEP 2 Practice & Discussion Session II Practice & Discussion Session I Actual Ratings STEP 4
Rater Training Training materials Using the same rating scale descriptors and rating procedures as used in the 1st study One day Workshop (3-4 hours) More practice during on-site workshops Familiarization and norming sessions by providing lecture, practices, and discussions
Methodology in the 2nd Study For trained raters Repeated measures to look at the audio/visual interaction effect Logistic Regression analysis Correlation analysis For comparisons with two groups T-test for two group mean differences in terms of assessment criteria FACETS analysis
Validity is the underlying objective – to validate the measures developed in this project. Messick, 1989: Validity is a unitary concept which includes test use and consequences “Validity is an integrated evaluative judgment of the degree to which empirical evidence and theoretical rationales support the adequacy and appropriateness of inferences and actions based on test scores or other modes of assessment”. Kane, 2006 : Validation is a concept which understands the procedures to connect test scores/score-based inferences to test use/ the consequences of test use. “To validate an interpretation or use of measurements is to evaluate the rationale, or argument, for the claims being made, and this in turn requires a clear statement of the proposed interpretations and uses and a critical evaluation of these interpretations and uses”
Given the above definitions of validity: Further refinement of the rating scale may be suggested based on: Empirical evidence to support changes made to the rating scale New rating tool for untrained raters Wording of descriptors e.g. Phonological Control 7 Assessment criteria used in the study Propose a training model for trained raters group
Thank you for your time!!
Grammatical & Lexical Accuracy CHINESE FLUENCY PROJECT–RATING RUBRICS Version 0.43 Fluency Level Speech Flow Phonological Control Grammatical & Lexical Accuracy Delivery Skills 4 Speech is clear with fluid and ease expression; no inappropriate use of filler words, repetitions, self-correction or noticeable errors occurred that affect overall intelligibility. Phonological features (tones, rhythm, stress) used properly and easily understood by speakers unaccustomed to dealing with non-native speakers of Chinese. Excellent use of grammatical structures. Wide range of vocabulary usage. Intended meanings can be understood easily. Style of delivery (eye contact, gestures, facial expressions) effectively enhances the communication of the message. 3 Speech is clear though occasional listener effort is needed; meaning may be obscured from time to time due to a few errors and some inappropriate use of filler words, repetitions, or self-correction. Phonological features sometimes improperly used but can be understood, according to the context, by speakers accustomed to dealing with non-native speakers of Chinese. Although some errors in grammatical accuracy and selection of vocabulary, these generally do not interfere with communication. Style of delivery (eye contact, gestures, facial expressions) adequately enhances the communication of the message. 2 Speech is somewhat comprehensible despite difficulties in formulating ideas; pausing for grammatical and lexical planning and repair is very evident. Phonological features often used improperly that even speakers accustomed to dealing with non-native speakers of Chinese receive limited clues from the context and understand fragmented utterances only. Use of vocabulary and grammatical structure somewhat cause confusion and intended meanings not fully delivered. Style of delivery (eye contact, gestures, facial expressions ) is fragmented and inconsistent and this frequently impedes the communication of the message. 1 Speech contains frequent or long pausing to search for expressions; only isolated and short phrases may be understood through considerable listener effort. Phonological features mostly misused that even Chinese instructors accustomed to dealing with non-native speakers of Chinese understand very few phrases. Grammatical and vocabulary control limited to simple, short, and familiar sentences and words only. Style of delivery (eye contact, gestures, facial expressions) impedes communication of the message.
Correlation Analysis (Audio 1) First Round Fluent2 Disqual Pron Natness Comm Syntax Lexicon 1 .680(**) .639(**) .682(**) .642(**) .661(**) .731(**) .660(**) .734(**) .751(**) .666(**) .712(**) .732(**) .813(**) .677(**) .754(**) .722(**) .743(**) .792(**) .775(**)
Correlation Analysis (Audio 2) Second Round Fluent2 Disqual Pron Natness Comm Syntax Lexicon 1 .697(**) .684(**) .743(**) .600(**) .736(**) .803(**) .681(**) .783(**) .804(**) .747(**) .631(**) .750(**) .763(**) .833(**) .669(**) .795(**) .764(**) .746(**) .832(**) .822(**)