Saturday, January 18, 2020

ELAN Transcription and Translation Workflow

Below is a description of the workflow I follow for transcribing and translating oral texts into ELAN. 

 This post fills in the sketch for ELAN workflow that I gave in an earlier post on rate of transcription:

The workflow assumes that the media files (audio and video) have been uploaded to ELAN and synchronized (using Media Synchronization Mode). It also assumes that the oral text has been segmented in ELAN. I generally work in Annotation Mode, so I can easily change segment boundaries. But I also skip back and forth to Transcription Mode, which has some nice features. For example, the view in Transcription Mode gives a very clear overview of what has been transcribed and translated already, and what remains to be done.

My ELAN project is set up with three tiers (a Sasi tier, with no parent tier), and two translation tiers, one for Setswana and one for English. Each of the translation tiers has the Sasi tier as parent. I do not set up a glossing tier in ELAN, since it is much easier to gloss in FLEx.

See the following post for converting an ELAN file to FLEx, and back again:

My working group consists of the linguist (me), a Setswana translator and two Sasi consultants. The linguist speaks English, and some Setswana and Sasi. The Setswana translator speaks English and standard Setswana (but no Sasi at all). The Sasi consultants speak Sasi and the Sengwato dialect of Setswana (but no English at all). I always try to include the original speaker in the group if possible. The Sasi consultants are illiterate, so they do not know how to read or write in any language.

The translator is necessary (a) to communicate with the consultants, especially about complicated issues, and (b) to create accurate and natural Setswana translations for the videos (since the resulting video clips will be made available to Setswana speaking people in Botswana).

This work mostly takes place in the village. There is only one computer, a Macbook Pro, powered by a solar panel. The computer is running ELAN, but not FLEx. The translator does not have a computer, and there is no internet.

Linguists with a different set-up (e.g., only one translation language) will not use the same workflow. However, I advise people to write down their workflow. This will help them make the process more efficient. A single 10 minute text may take 10 hours or more to process, so working out the exact steps can save a lot of time. I know this from experience!

Steps in Workflow

The linguist plays a segment from ELAN to the group one or more times.
[Press “Play selection”, immediately above the waveform in ELAN.]

The linguist transcribes the segment in modified IPA, writing by hand into a notebook.
[In general, I try to stick closely to what the consultant actually says. The rough transcription in step 2 is sometimes modified after the translation steps below. It can be more efficient to do a bunch of segment transcriptions ahead of time (without the translator and consultants).]

The linguist refers to the Sasi dictionary to check transcriptions and glosses of words where needed.
[The Sasi dictionary is a .pdf version, since FLEx only runs on a PC.]

The linguist asks the consultant to repeat the segment, if it is unclear.
[This task is difficult for many consultants to carry out.]

The linguist asks the consultant any questions needed to clarify segment.
[For example, Who said this? Who are they talking to? Who does ‘he’ refer to? Where does ‘there’ refer to? How is this word pronounced?]

The linguist asks the consultant to define any new words found in the current segment.
[Just a quick gloss is needed for now, more systematic lexical work can be done in a separate session. The larger the existing dictionary of the language, the less time that needs to be spent on defining new words.]

Words and phrases that are difficult to transcribe can be marked with (???), optionally followed by the best guess transcription.
[See Appendix below.]

The linguist adjusts the segment boundaries were necessary (splitting segments, merging segments, etc.) using ELAN commands in Annotation Mode.
[In transcribing, it becomes clearer what the segment boundaries should be. Usually the segments correspond to complete sentences, or long clauses of complete sentences. ELAN uses the confusing word “annotation” to refer to the individual segments of an oral text. Some useful commands that I use regularly are: New Annotation Here, Merge with Next Annotation, Delete Annotation. All these are found by right clicking on a particular segment on the top tier.]

The linguist hands the notebook to the translator.
[In simple cases, the linguist just translates the Sasi into Setswana and English, bypassing steps 9-15. In other simple cases, the linguistics translates directly from Sasi to English and hands the notebook to the translator for Setswana. As should be clear, the notebook is needed in the process to give the translator a place to write the Setswana translations to share and discuss with the linguist.]

The translator asks the consultant to translate the segment into Setswana.
[The translator says: Bua ka Setswana “Speak in Setswana” or Ka Setswana “In Setswana”]

The translator writes down Setswana translation by hand in notebook.
[The translator mostly writes in standard Setswana.]

The translator asks consultant questions about Setswana translation.
[For example, questions about unknown words, ambiguous words, dialect issues. Also, the consultant may have to repeat the Setswana translation.]

The translator translates Setswana into English, writing by hand in the notebook.
[At this point, the notebook has three hand-written lines: Sasi, Setswana and English.]

The translator uses the Setswana-English dictionary for translations and spellings where needed.
[In our work, the use of the dictionary has turned out to be a time sink so I discourage it, unless it is necessary.]

The translator hands the notebook back to the linguist.

The linguist checks consistency of the Sasi transcription, and the Setswana and English translations, and addresses any issues that come up.
[This step can sometimes involve lengthy discussions with the consultants and the translator about the correct translations. It can also lead to discussions of cultural and historical matters.]

The linguist types Sasi into MS Word.
[The Sasi keyboard (for the Macbook Pro) does not work directly in ELAN. This step takes place at the same time as steps 10-14, so no time is lost.

The linguist copies Sasi from MS Word and pastes into ELAN.
[This step also takes place at the time as steps 10-14.]

The linguist copies the Setswana and English translations from the notebook, by typing them into the ELAN translation tiers.

The linguist uses the ELAN commands to move to the next segment, and the process starts over.
[Press “Go to next annotation” which is a right arrow, just above the wave form in ELAN.]

After transcribing and translating an oral text, the whole text should be reviewed (in Transcription Mode), to check for errors and to check the consistency of the running text in each of translation tiers (e.g., Are all the tenses and pronouns consistent? Does the oral text make sense? Are there small notes that can be inserted in the translations to help the reader? etc.). Ideally, the translator should be present for this step. For a 15 minute long oral text, this step itself can take a few hours (e.g., 2-4 hours), and has its own internal workflow.

Transferring the oral text from ELAN to FLEx to ELAN might result in further minor corrections in the three tiers.

Appendix [Step 7]: Resolving (???)
As noted in step 7, words and phrases that are difficult to transcribe can be marked with (???), optionally followed by the best guess transcription. Unresolved issues arise when the linguist is not able to transcribe a word or a phrase or a whole segment, and the consultants are also unable to make out the word or phrase or segment. The reasons that this happens are numerous:

The speaker is speaking very softly.
The speaker is speaking very fast.
The speaker is speaking very loudly (clipping the sound file).
The speaker does not complete some phrases or sentences.
There is background noise (e.g., wind, music, people speaking, clothing on lavs).
The quality of the recording is poor (e.g., using the built-in camera mic, echo in room).
The mic was positioned too far from the speakers.
The speaker turned away from the mic.

In my case, the speakers often do not remember exactly what they said. In one case, the speaker has a hearing disability, so they cannot help with transcription of difficult areas of their own texts.

Areas marked with (???) can sometimes be resolved in the translation steps. They can sometimes be resolved in the review step, once the entire text has been transcribed and translated. If you put the text on the back burner for a few days and revisit it, sometimes the transcription will become clear. Lastly, some consultants are better than others at transcription, so if you find some difficult spots in a text, you can try to work them out with a different consultant.

Even for unresolved words, phrases and segments, try to give rough transcriptions with as much detail as possible, since the more you chip away at a difficult area, the greater the chance that you will have of resolving it later on.

As a last resort, you might consider developing a policy where unresolved words and phrases are replaced with reasonable alternatives that make sense in the context of the oral text.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.