We are pleased to announce that Wissam Antoun ranked 1st in the Arabic Sentiment Analysis 2021 @ KAUST competition. Arabic Sentiment Analysis is one of the most popular tasks in Arabic Natural Language Processing (ANLP), with this competition being one of the largest in terms of the prize and dataset size.
![](https://sites.aub.edu.lb/mindlab/files/2020/02/arabert_logo.png)
AraBERT
The competition required building machine learning models that can determine the sentiment (positive, negative, neutral) behind Arabic text (tweets), with a prize of 10000 USD, 5000 USD, and 2000 USD to the first, second and third place winners, respectively. The organizers provided participants with 55,000 annotated tweets named ASAD (Arabic Sentiment Analysis Dataset) for training and developing their machine learning models. The provided dataset targeted multiple Arabic dialects and topics. The participants then had to predict and submit the sentiment of 20,000 tweets, which are used to calculate the public leaderboard score. Note that the baseline scores provided by the organizers were achieved using AraBERT a language understanding model created by Wissam, Fady Baly, and Hazem Hajj.
![](https://sites.aub.edu.lb/mindlab/files/2021/08/wti4.jpg)
Competition Timeline
![](https://sites.aub.edu.lb/mindlab/files/2021/08/wti3.jpg)
Stage 1: Public leaderboard scores (not final)
The competition ran for 3 months, with 74 teams participating and submitting their predictions. Top scores were close, with only a 0.006 difference between the first and the fifth submission. Wissam ranked 3rd on the public leaderboard when the submission window closed. Top-ranked participants were then invited to submit and share their codes to the organizers for the final evaluation on another private 20,000 tweets dataset. Wissam’s submission scored the highest on the private dataset.
![](https://sites.aub.edu.lb/mindlab/files/2021/08/wti1-1.jpg)
Stage 2: Final results
Wissam’s winning submission consisted of aggregating the predicting of multiple models (also known as ensembling) each trained to recognize a specific sentiment characteristic from a tweet (i.e. Sarcasm, Emojis…) while employing thorough data cleaning and text pre-processing techniques. A detailed description of the approach, code, and models will be released soon.
Congratulations Wissam!!
definately enjoy every little bit of it and I have you bookmarked to check out new stuff of your blog a must read blog!