Amr Keleg
@amr-keleg.bsky.social
PhD student at the University of Edinburgh.
Co-creator of AlWird (Arabic Wordle).
Research interests: Diversity of Arabic Dialects, Arabic NLP, Multilinguality.
https://amr-keleg.github.io/
Co-creator of AlWird (Arabic Wordle).
Research interests: Diversity of Arabic Dialects, Arabic NLP, Multilinguality.
https://amr-keleg.github.io/
I realized the instructions were circulated as part of the acceptance email. This caught me by surprise as well, especially that it's different from how things were previously managed.
All the best with your preparations, though!
All the best with your preparations, though!
June 24, 2025 at 11:24 PM
I realized the instructions were circulated as part of the acceptance email. This caught me by surprise as well, especially that it's different from how things were previously managed.
All the best with your preparations, though!
All the best with your preparations, though!
Glad to know you liked it 😀
March 24, 2025 at 4:59 PM
Glad to know you liked it 😀
I share some preliminary thoughts for four steps that could help in building culturally representative models.
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
March 21, 2025 at 6:56 PM
I share some preliminary thoughts for four steps that could help in building culturally representative models.
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
* The NLP community acknowledges the rich diversity of the Arabic dialects, which are a manifestation of cultural differences across the region.
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
March 21, 2025 at 6:56 PM
* The NLP community acknowledges the rich diversity of the Arabic dialects, which are a manifestation of cultural differences across the region.
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* Arabic speakers have substantial cultural similarities (see map below). This does not imply they have one single homogenous culture!
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
March 21, 2025 at 6:55 PM
* Arabic speakers have substantial cultural similarities (see map below). This does not imply they have one single homogenous culture!
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
I share some preliminary thoughts for four steps that could help in building culturally representative models.
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
March 21, 2025 at 6:53 PM
I share some preliminary thoughts for four steps that could help in building culturally representative models.
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
Lastly, I hope this will spark discussions within the Arabic NLP community, and the broader NLP community interested in serving marginalized speech communities!
(4/4)
* The NLP community acknowledges the rich diversity of the Arabic dialects, which are a manifestation of cultural differences across the region.
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
March 21, 2025 at 6:53 PM
* The NLP community acknowledges the rich diversity of the Arabic dialects, which are a manifestation of cultural differences across the region.
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* While Arabic-specific LLMs are still marketed as serving all Arabs, our alignment data/benchmarks are scarce and not inclusive enough!
(3/4)
* Arabic speakers have substantial cultural similarities (see map below). This does not imply they have one single homogenous culture!
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
March 21, 2025 at 6:53 PM
* Arabic speakers have substantial cultural similarities (see map below). This does not imply they have one single homogenous culture!
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)
* Their views tend to be ignored, even for largely diverse alignment datasets (e.g., PRISM, Kirk et al., 2024).
(2/4)