Create dataset loader for M3LS

[SamuelCahyawijaya]

Dataloader name: m3ls/m3ls.py
DataCatalogue: http://seacrowd.github.io/seacrowd-catalogue/card.html?m3ls

Dataset	m3ls
Description	The multilingual multimodal summarization dataset (M3LS) consists of over a million instances of document-image pairs along with a professionally annotated multimodal summary for each pair. It is derived from news articles published by the British Broadcasting Corporation (BBC) over a decade and spans 20 total languages.
Subsets	-
Languages	ind
Tasks	Summarization
License	MIT (mit)
Homepage	https://github.com/anubhav-jangra/M3LS
HF URL	-
Paper URL	https://aclanthology.org/2023.eacl-main.263/

[sedrickkeh]

#self-assign

[github-actions]

Hi, may I know if you are still working on this issue? Please let @holylovenia @SamuelCahyawijaya @sabilmakbar know if you need any help.

[sedrickkeh]

Working on it. Will try to finish this week

[holylovenia]

Working on it. Will try to finish this week

No problem! Feel free to let us know anytime you would like to discuss.

[github-actions]

Hi, may I know if you are still working on this issue? Please let @holylovenia @SamuelCahyawijaya @sabilmakbar know if you need any help.

[sabilmakbar]

#self-assign

[sabilmakbar]

and for this one, I'll try to do this immediately after this one is PR-ed:
#449

[sabilmakbar]

Hi, seems the dataset itself is exceedingly large (the zipped version is around 14GB, unsure abt the actual size after unzipping -- now I'm doing it).

Also, there is a forseeable blocker on bypassing Google Drive downloading process by passing the GDrive URL to either datasets.DownloadManager or gdown.download. I'm trying to fix the issue and see other workarounds as well (last time I saw a possible workaround in #206 discussion).

If all of them aren't possible, maybe the last resort is to change the format into local-based dataset

[sabilmakbar]

updates:

1. There are multiple files that can be used, but there's no clear documentation on the contents/folder structuring itself on the description. Prob needs to skim the paper or the scrapper codes to get some hints.

2. From what I inspect on the , prob this data contains much more info than we thought initially (only text summarization from article). Idk if the images there can be stitched together to get a multimodal dataset, tho (would be valuable if we can somehow pull it).
   

any thoughts or ideas? @holylovenia @SamuelCahyawijaya

[holylovenia]

Thanks for inspecting this dataset, @sabilmakbar! I think this dataset is a multimodal dataset, precisely a multilingual
multimodal summarization dataset. Or did you mean stitching up another multimodal dataset?