Scribd
Upload
4 views

Labelling Rules

Labelling rules

Uploaded by

hafeezfatima27
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
4 views

Labelling Rules

Labelling rules

Uploaded by

hafeezfatima27
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 4

Rule document

1. We are looking for a quality inspection supplier this time, and you need to conduct quality
inspection on the audio that has been marked;
2. The quality inspection standards are completely consistent with the labeling specifications
written in this document. Quality control colleagues need to listen to the entire audio in its
entirety. If there are individual transcription spelling/sentence segmentation errors, they need
help to correct them. If it is obvious that the annotator has misunderstood the rules, return the
entire article.
3. The annotation personnel who have been returned with data need to make modifications. If
there are unclear rules and inconsistent standards, contact the responsible personnel to organize
records and send them to Aopeng for evaluation and confirmation.
4. As the quality inspection standards are consistent with the annotation specifications written in
this document, the quality inspection supplier also needs to annotate the original audio during
the trial phase, and determine whether the quality inspection supplier has the ability to execute
the project based on the trial results.

Firstly, read the content of the 'Annotation Standards', and then click on the' Key Questions and
Answers' online form to carefully review it
Online forms are very important!!!!

[Labeling Standards]
Before annotating, the annotator needs to assess the audio quality:
Confirm whether it is the data of a two person conversation in the corresponding language. If
not, directly determine that the entire audio is invalid
No need to label (if a whole audio segment is not a two person conversation, the data is invalid!)
Confirm the background sound of the audio, and any data with BGM laughter or applause will
directly determine that the entire audio is invalid
No need to label
Accent, accent data does not need to be annotated
1. For English data - if the speaker has a non European or American accent (which can be
understood as invalid data as long as the human ear judges that it is not a regular English
pronunciation, such as Indian English), the entire audio is directly judged as invalid
2. For Chinese data - if the speaker has a distinct regional accent, the entire audio will be deemed
invalid

be careful!! Regardless of whether the template is selected as valid or invalid, the "Valid Data"
button on the left side at the bottom of the page should never be moved! Click directly to
confirm completion and load the next one“
Thi
s
con
ten
t is
for
int
ern
al
use
onl
y
be careful!! If the entire audio is invalid, select "Invalid" and click Ctrl+Delete to clear all audio
segments. Select "none" for the line attribute to submit! (If delete cannot be restored, use with
caution)
Text pre transcription is configured in Appen task, and the model is used to automatically
recognize and distinguish the speaker's role.
However, the tool recognition is not entirely accurate. The task you need to do is to manually
move and modify the boxes and text marked incorrectly by the tool, and correct the machine's
cutting errors
be careful!! Do not let the pre recognition results interfere with manual judgment. Whether it's
the recognition of characters or the completeness and accuracy of each transcribed text
sentence. Based on human hearing and judgment, manual annotation shall prevail. (Note that it
is not based on pre identification)

1. Handling of front and rear silent segments: If the front and rear silent segments exceed 500ms,
they will be marked; No labeling is required within 500ms# none
2. Use tag # none for the middle silent section
In the original audio, if the pronunciation of the first and last sentences is cut off by the machine,
the incomplete sentences need to be framed as a whole, and the invalid data at the beginning
and end should be cut off when selecting #. Until the complete sentence appears, start selecting
the corresponding character and transcribing it

4. Regarding the segmentation criteria in the middle of paragraphs:


The first principle is to divide by role and select the corresponding role label (role 1/2)
From the beginning of character 1's speech to the beginning of character 2's speech, select #
speechr1 for this interval, without marking a silent section in between; From the beginning of
character 2's speech to the beginning of the next character 1's speech, select # speechr2 without
marking a silent section in between.
At the same time, it should be noted that the longest time for splitting is 10 seconds (no labels
need to be selected for the cut, just transcribe the next sentence directly), and within 10
seconds, use punctuation marks to naturally break sentences according to semantics (use a
period if the topic ends, and a comma if the topic continues)
Two people speaking at the same time, such as interrupting: choose # Two people speaking at
the same time without transcription
(If two people interrupt each other slightly, the sentence that needs to be transcribed should be
kept as complete as possible; but if two people keep interrupting each other, then choose the
whole sentence. # Two people speak at the same time without transcription.)
Short and ambiguous content: Choose # Two people speaking at the same time without
transcription
Thi If there are background sound effects (such as applause and laughter sound effects) but no one
s speaks, select the tag: # There are background sound effects but no one speaks
con
ten
t is In the trial data, it was found that some annotators rely entirely on pre recognition for the
for standard of audio framing, which is a wrong idea! The basis for segmentation is: sentence
int meaning+maximum length of 10 seconds, followed by the annotation rule for segmentation. If
ern
al
use
onl
y
pre recognition breaks a complete sentence into scattered segments, it requires manual deletion
of unnecessary and redundant cuts in the middle. One sentence and one box are enough

5. Spaces on both sides of numbers or letters in Chinese do not need to be marked


6. If there are other conflicts with the company or conversations introducing the audio source in
the audio, you need to select # Invalid data needs to be cut off
If the original audio contains inappropriate comments such as yellow violence, you need to select
# Invalid data needs to be cut off

【 Key Question Answering 】 Online Form Address:


https://docs.qq.com/sheet/DZHVSYXZKdVB6RldX?tab=BB08J2

Other instructions
1. Clicking on the "little robot" in the text box will link to pre transcribed text, saving time and
improving efficiency
2. Platform function keys, click on the "?" in the upper right corner to view the function guide

2. When selecting speaker 1 and speaker 2, the transcription content cannot be empty
3. The effective fragment cannot exceed 10 seconds
4. Required attributes

5. The first 'none' label in the left column is the platform default and cannot be deleted or
changed. At the same time, due to the platform's configuration of the "tag must be selected"
feature limitation, submissions cannot be made without selecting tags, which means that each
line must have two tags to submit for acceptance.

Thi
s
con
ten
t is
for
int
ern
al
use
onl
y
Thi
s
con
ten
t is
for
int
ern
al
use
onl
y

You might also like