This paper describes the Rosetta system for Subtask 1 (Context-Aware English-to-Dialectal Arabic Dialogue Translation) of the AlexandriaX shared task, participating in both constrained and unconstrained tracks. The approach fine-tunes a LoRA adapter on NileChat-3B using structured system/user prompts that condition generation on dialect and dialogue context. For the unconstrained track, the adapter is additionally pretrained on MADAR and PADIC. Rosetta ranked 4th in the constrained track (spBLEU 26.10) and 5th in the unconstrained track (spBLEU 25.09). The experimental results demonstrate that external pretraining helps only two of thirteen dialects while slightly hurting overall performance, suggesting negative transfer.

In the contemporary digital landscape, social media has emerged as a prominent means of communication and information dissemination, offering a rapid outreach to a broad audience compared to traditional communication methods. Unfortunately, the escalating prevalence of abusive language and hate speech on these platforms has become a pressing issue. Detecting and addressing such content on the Internet has garnered considerable attention due to the significant impact it has on individuals. The advent of deep learning has facilitated the use of pre-trained deep neural network models for text classification tasks. While these models demonstrate high performance, some exhibit a substantial number of parameters. In the DravidianLangTech@EACL 2024 task, we opted for the Distilbert-base-multilingual-cased model, an enhancement of the BERT model that effectively reduces the number of parameters without compromising performance. This model was selected based on its exceptional results in the task. Our system achieved a commendable Macro F1 score of 0.6369%.