Difference between revisions of "Fluency Summer Intern Project"

From LearnLab
Jump to: navigation, search
(Added results)
m
 
(3 intermediate revisions by one other user not shown)
Line 37: Line 37:
 
%pos: det|the [*] n|soccer v|be&3S [*] adv:int|very adj|famous .  
 
%pos: det|the [*] n|soccer v|be&3S [*] adv:int|very adj|famous .  
  
%err: $DET:ART:SUB:ZERO
+
%err: $DET:ART:SUB:ZERO; $DET:ART:LOS:A
  
 
'''163:''' game all [/] all over the world .  
 
'''163:''' game all [/] all over the world .  
Line 46: Line 46:
  
  
The %mor tier codes morphemic segments by type and part of speech.  If a word could  
+
The %mor tier codes morphemic segments by type and part of speech.  If a word could  
possibly function as more than one part of speech, options are separated by a ^ symbol.
+
possibly function as more than one part of speech, options are separated by a ^ symbol.
 +
 
 +
The %pos tier codes morphemic segments by type and part of speech as well.  However,
 +
all options have been eliminated besides the correct type and part of speech for each word.
  
The %pos tier codes morphemic segments by type and part of speech as well.  However,
 
all options have been eliminated besides the correct type and part of speech for each word.
 
  
 
; CLAN: CLAN performs automatic analyses of the transcriptions.  CLAN processed the CHAT files and produced the probable part of speech for each word in the transcript.  Where there was still uncertainty concerning the part of speech, options were given and a choice was made by the coder.  An error coding file developed by the principal investigator, Nel de Jong, was referenced for lines of the transcript containing error.  From this file, CLAN produced a hierarchy of the part of speech codes in order to determine which form of the part of speech should have been used versus the form used in error by the particpant.  CLAN can also perform other processes aside from morphosyntactic analysis.  Such processes include the generation of a word frequency list and a lexicon of all words used in the transcript.
 
; CLAN: CLAN performs automatic analyses of the transcriptions.  CLAN processed the CHAT files and produced the probable part of speech for each word in the transcript.  Where there was still uncertainty concerning the part of speech, options were given and a choice was made by the coder.  An error coding file developed by the principal investigator, Nel de Jong, was referenced for lines of the transcript containing error.  From this file, CLAN produced a hierarchy of the part of speech codes in order to determine which form of the part of speech should have been used versus the form used in error by the particpant.  CLAN can also perform other processes aside from morphosyntactic analysis.  Such processes include the generation of a word frequency list and a lexicon of all words used in the transcript.
Line 73: Line 74:
  
 
== Results ==
 
== Results ==
'''Lexical overlap'''
+
''Lexical overlap''. To verify whether students tried to express the same meaning in the retellings, we calculated the lexical overlap between the first and second recording, and between the second and third recording in each fluency training session. Lexical overlap is the number of words that were used in both recordings, divided by the total number of words that were used in the two recordings. It was indeed higher for students who repeated their speech than for students who talked about three different topics (.15-.31 for no repetition vs. .35-.59 for repetition).
To check whether students were trying to express the same meaning in the retellings, we calculated the lexical overlap between the three recordings in each session (the number of words that were used in both recordings, divided by the total number of words that were used in the two recordings). Lexical overlap was indeed higher for students who repeated their speech than for students who talked about three different topics (.15-.31 for no repetition vs. .35-.59 for repetition).
 
  
'''Fluency'''
+
''Fluency''. Fluency developed across the three recordings in that fluent runs became longer, while pauses length and phonation/time ratio were stable or improved. In the first 4/3/2 training session, all students whose fluency improved from the the pretest to the posttest also showed fluency development in training. On the other hand, two out of three students whose fluency did not develop on the pre- and posttests also did not develop fluency during training. In the second training session, no fluency development took place for any students, while in the third session, fluency developed for all six students.
Fluency developed across the three recordings in that fluent runs became longer, and pauses length and phonation/time ratio were stable or improved. In the first 4/3/2 training session, all students whose fluency improved in the pre- and posttest also showed fluency development in training. On the other hand, two out of three students whose fluency did not develop on the pre- and posttests also did not develop fluency during training. In the second training session, no fluency development took place for any students. In the third session, fluency developed for all six students.
+
Fluency development was also tracked by the number of repetitions per minute (e.g., ''it's ... it's imp-... important'' has two repetitions). This number decreased across retellings in most cases, which indicates that students had less difficulty retrieving words and structures while they were speaking.  
Fluency development was also tracked by the number of repetitions per minute (e.g., it's ... it's imp-... important has two repetitions). This number decreased across retellings in most cases, which indicates that students had less difficulty retrieving words and structures while they were speaking.  
 
  
'''Accuracy and complexity'''
+
''Accuracy and complexity''. Accuracy was measured by the target-like use of subject-verb agreement (third person -s). This measure remained mostly stable, but was fairly high, over .75 for most recordings, which indicates there may have been a ceiling effect. In addition, the number of self-corrections was low, on average between 0.4 and 1.1. per minute, and remained stable across retellings, which indicates that students did not monitor their speech extensively.  
Accuracy was measured by the target-like use of subject-verb agreement (third person -s). This measure remained mostly stable, but was fairly high, over .75 for most recordings, which indicates there may have been a ceiling effect. In addition, the number of self-corrections was low, on average between 0.4 and 1.1. per minute, and remained stable across retellings, which indicates that students did not monitor their speech extensively.  
 
  
'''Lexical variety'''
+
''Lexical variety''. As a measure for lexical variety we used the Mean Segmental Type-Token Ratio, the mean type-token ratio for all segments of 40 words in a recording. Lexical variety did not increase in the retellings, but remained mostly stable. It did, however, increase from pretest to immediate posttest for those students whose fluency improved as well.
As a measure for lexical variety we used the Mean Segmental Type-Token Ratio, the mean type-token ratio for all segments of 40 words in a recording. Lexical variety did not increase in the retellings, but remained mostly stable. It did, however, increase from pretest to immediate posttest for those students whose fluency improved as well.
 
  
In summary, it seems that fluency increased during the retellings in the 4/3/2 procedure, mostly in terms of temporal measures. Accuracy for subject-verb agreement was high, and lexical variety was mostly stable. Further analyses of other, more problematic morphosyntactic structures may still reveal an increase in accuracy. It seems that most students were able to improve the fluency of their speeches during training, but only some of them were able to transfer this gain to the posttest. Future studies will need to address why this would be the case. Also, analyses of data from more students may reveal whether this is a general pattern or occurred only for few students.
+
In summary, it seems that fluency increased during the retellings in the 4/3/2 procedure, mostly in terms of temporal measures. Accuracy for subject-verb agreement was high, and lexical variety was mostly stable. Further analyses of other, more problematic morphosyntactic structures may still reveal an increase in accuracy. Most students seemed to improve the fluency of their speeches during training, but only some of them were able to transfer this gain to the posttest. Future studies will need to address why this would be the case (e.g., motivation, focus on accuracy or complexity instead of fluency). In addition, analyses of data from more students may reveal whether this is a general pattern or occurred only for few students.
  
 
== More information ==
 
== More information ==
 
For more information about the fluency project, see [[Fostering fluency in second language learning]].
 
For more information about the fluency project, see [[Fostering fluency in second language learning]].

Latest revision as of 04:32, 4 August 2007

Purpose of the Project

The goal of this project is to gain further insight into the development of fluency, accuracy and complexity during fluency training, and the retention and transfer of that development to a delayed posttest. The fluency training is the 4/3/2 task, developed by Nation (1989), in which students prepare a four minute talk and repeat it twice: first in three minutes, then in two minutes.

Research Questions

1. Does the repetition of the second and third telling of the subject's speech, in which new semantic information is not required, result in changes in fluency, morphosyntactic accuracy, and complexity?

2. What types of changes, if any, occur? What causes these changes?

3. Are the changes retained for atleast one week?

Data Collection

Data was collected under two different conditions. The first condition was that students would repeat their speech, developed from a prompt concerning a singular topic, on three occasions. The second condition was that students would be asked to address three different topics. Data consists of recordings from the 4/3/2 task, as well as from pre- and posttests. Articulation rate, morphology, syntax, mean length of fluent runs, and mean length of pauses will all be examined.

Data preparation

Recordings from the pre- and posttests were transcribed and analyzed for fluency before. In the present project, recordings from the training itself were transcribed and coded for morphosyntactic accuracy, parts of speech, and repetitions and reformulations.


Tools Used to Analyze Data

PRAAT
PRAAT is a computer program that allows multiple tier transcriptions to be displayed along with sound wave files. For the purposes of this project, PRAAT was used to transcribe the 4/3/2 recordings. A spectrogram was referenced while listening to the sound files in order to determine appropriate time intervals of pauses and speech utterances.


CHAT
CHAT is a transcription format in which the transcribed data can be coded. The transcribed data was marked for part of speech, a range of morphological and syntactic errors, and repetitions using CHAT.


Example of CHAT format

163: the [*] soccer is [*] very famous: .

%mor: det|the [*] n|soccer v|be&3S^v:aux|be&3S [*] adv:int|very adj|famous .

%pos: det|the [*] n|soccer v|be&3S [*] adv:int|very adj|famous .

%err: $DET:ART:SUB:ZERO; $DET:ART:LOS:A

163: game all [/] all over the world .

%mor: v|game^n|game qn|all prep|over^adv:loc|over det|the n|world .

%pos: n|game qn|all adv:loc|over det|the n|world .


The %mor tier codes morphemic segments by type and part of speech. If a word could possibly function as more than one part of speech, options are separated by a ^ symbol.

The %pos tier codes morphemic segments by type and part of speech as well. However, all options have been eliminated besides the correct type and part of speech for each word.


CLAN
CLAN performs automatic analyses of the transcriptions. CLAN processed the CHAT files and produced the probable part of speech for each word in the transcript. Where there was still uncertainty concerning the part of speech, options were given and a choice was made by the coder. An error coding file developed by the principal investigator, Nel de Jong, was referenced for lines of the transcript containing error. From this file, CLAN produced a hierarchy of the part of speech codes in order to determine which form of the part of speech should have been used versus the form used in error by the particpant. CLAN can also perform other processes aside from morphosyntactic analysis. Such processes include the generation of a word frequency list and a lexicon of all words used in the transcript.


Error coding

Errors which were coded include:

  • Morphological errors such as:
    • inflection
    • noun (number)
    • pronoun (number, gender, loss, addition, substitution)
    • determiner (article substitution, absence, addition)
    • verb (tense, addition, loss, substitution, subject-verb agreement)
    • relative pronoun (loss, addition, substitution)
    • conjunction (loss, addition, substitution)
  • Repetitions were marked by: [/] which represents an exact repetition, [//] which represents retracing with correction, and [1] which represents retracing with reformulation.
  • Syntactical errors will eventually be addressed, but not at this moment.

Results

Lexical overlap. To verify whether students tried to express the same meaning in the retellings, we calculated the lexical overlap between the first and second recording, and between the second and third recording in each fluency training session. Lexical overlap is the number of words that were used in both recordings, divided by the total number of words that were used in the two recordings. It was indeed higher for students who repeated their speech than for students who talked about three different topics (.15-.31 for no repetition vs. .35-.59 for repetition).

Fluency. Fluency developed across the three recordings in that fluent runs became longer, while pauses length and phonation/time ratio were stable or improved. In the first 4/3/2 training session, all students whose fluency improved from the the pretest to the posttest also showed fluency development in training. On the other hand, two out of three students whose fluency did not develop on the pre- and posttests also did not develop fluency during training. In the second training session, no fluency development took place for any students, while in the third session, fluency developed for all six students. Fluency development was also tracked by the number of repetitions per minute (e.g., it's ... it's imp-... important has two repetitions). This number decreased across retellings in most cases, which indicates that students had less difficulty retrieving words and structures while they were speaking.

Accuracy and complexity. Accuracy was measured by the target-like use of subject-verb agreement (third person -s). This measure remained mostly stable, but was fairly high, over .75 for most recordings, which indicates there may have been a ceiling effect. In addition, the number of self-corrections was low, on average between 0.4 and 1.1. per minute, and remained stable across retellings, which indicates that students did not monitor their speech extensively.

Lexical variety. As a measure for lexical variety we used the Mean Segmental Type-Token Ratio, the mean type-token ratio for all segments of 40 words in a recording. Lexical variety did not increase in the retellings, but remained mostly stable. It did, however, increase from pretest to immediate posttest for those students whose fluency improved as well.

In summary, it seems that fluency increased during the retellings in the 4/3/2 procedure, mostly in terms of temporal measures. Accuracy for subject-verb agreement was high, and lexical variety was mostly stable. Further analyses of other, more problematic morphosyntactic structures may still reveal an increase in accuracy. Most students seemed to improve the fluency of their speeches during training, but only some of them were able to transfer this gain to the posttest. Future studies will need to address why this would be the case (e.g., motivation, focus on accuracy or complexity instead of fluency). In addition, analyses of data from more students may reveal whether this is a general pattern or occurred only for few students.

More information

For more information about the fluency project, see Fostering fluency in second language learning.