 هلو أجل، today we will be presenting our banking 77 an intent detection dataset for Arabic. This is a joint work between Birzate University in Palestine and Cistek in Turkey. The goal is to classify a given query into an intent. For instance ماذا أفعل إذا لم أستلم بطاقة الجديدة which translates to what do I do if I still have not received my new card to the intent card arrival. مرور مرور جداً لدى بلاد زوجات، وضع مرور جداً والجانب لأعطاء كبير الإنسان. لقد ساعدت نضمن مرور جداً، وحدث أن يتعلق لكلام مرور مرور جداً. مرور جداً is the R-Banking 77 dataset, بصورة 31,000 قويريز، مرور جداً لكي يامل 77 مرور ويقوم بخبر مرور المرور جداً للأربية والبلستينية والدائل، ومع مخطوطاً لتعديد فتاة عربية V2 وهي مخطوطة ترانسية المدينة لتعديد فتاة عربية ومدينة تحصل على افون سكور 92% على مدينة مدينة ومع 90% على مدينة بلسطينية الانتبارة 77 مدينة تباً على الانتبارة 77 مدينة ويجب أن يتضغط تسريق 13 مجموعة ونقلت سليس and 77 مجموعة أو تسريق ويجب أن يتضغطها في مجموعة وعنها أن تضغطها في مجموعة ويجب أن تضغطها في مجموعة ويجب أن تحاول الماللما يوجد أعطان رقمية يقرأ افضل مجتموعة سليس 77 مجموعة تضغط مماتة لعبارة مجموعة لأننا لدينا one query in MSA and one query in Palestinian dialect and sometimes we have more as we are going to discuss in a bit. The mutation process went on for a few months. We used 26 annotators and there are two main phases for the annotation. The first one is Arabization and localization and the second one is a final review of the data. For Arabization localization we performed translation from Banking 77 English into MSA using Google Translate and then we performed manual annotations and revision. So we revised any translation errors in the MSA and the annotator then was given the option to write an additional query in MSA and then we have one Palestinian query for each MSA query and optionally a second Palestinian query for each corresponding MSA query if there is a new variation of it. The review basically looked at how acceptable the queries are and how semantically correct. We also performed some correctness on the queries and errors that comes from the translation misspellings. We have not corrected everything just to make sure this reflects real life scenarios which is in live chat queries. We also removed duplicate queries by adding variation to the query that is duplicated. There is a big لixical difference between MSA and Palestinian. If somebody asks why you want to do Palestinian dialect as well MSA, they are lixically different and we could see that in the Jacquard index which is 0.16 very low but that indicates that Arabic is highly the classic language and training one model on one dialect doesn't mean it's going to translate well and have good performance on different dialect or different variation of the language. We trained the transformer based model on the dataset we proposed like we mentioned Arabic V2 but before we evaluated Arabic V2 on our data we performed zero shout learning. We want to see how the model performs without even looking at Palestinian dialect data or even Arabic data. So first we trained an Arabic 77 data MSA and we evaluated on Palestinian. This is zero shout learning on Palestinian and we see the performance about 0.6 F1 score very low. If we take the English data and train giga bird or multilingual bird we see a significant drop on MSA and Palestinian. So we need to have a model trained on our data. So when we train our model on our data we see big F1 score big boost 92 F1 on MSA and about 90 F1 on Palestinian. Arabic V2 was the best performing model. To conclude we proposed Arabic 77 intent detection data. We benchmarked against multiple models Arabic V2 was the best model we benchmarked against. There is a demo here is a link be sure to go and give it a try and thank you for listening.