Ahmed Abdelali’s research while affiliated with Qatar Computing Research Institute and other places

We present ALLaM: Arabic Large Language Model, a series of large language models to support the ecosystem of Arabic Language Technologies (ALT). ALLaM is carefully trained considering the values of language alignment and knowledge transfer at scale. Our autoregressive decoder-only architecture models demonstrate how second-language acquisition via vocabulary expansion and pretraining on a mixture of Arabic and English text can steer a model towards a new language (Arabic) without any catastrophic forgetting in the original language (English). Furthermore, we highlight the effectiveness of using parallel/translated data to aid the process of knowledge alignment between languages. Finally, we show that extensive alignment with human preferences can significantly enhance the performance of a language model compared to models of a larger scale with lower quality alignment. ALLaM achieves state-of-the-art performance in various Arabic benchmarks, including MMLU Arabic, ACVA, and Arabic Exams. Our aligned models improve both in Arabic and English from their base aligned models.

Recent advancements in Large Language Models (LLMs) have significantly influenced the landscape of language and speech research. Despite this progress, these models lack specific benchmarking against state-of-the-art (SOTA) models tailored to particular languages and tasks. LAraBench addresses this gap for Ara-bic Natural Language Processing (NLP) and Speech Processing tasks, including sequence tagging and content classification across different domains. We utilized models such as GPT-3.5-turbo, GPT-4, BLOOMZ, Jais-13b-chat, Whisper, and USM, employing zero and few-shot learning techniques to tackle 33 distinct tasks across 61 publicly available datasets. This involved 98 experimental setups, encompassing ∼296K data points, ∼46 hours of speech, and 30 sentences for Text-to-Speech (TTS). This effort resulted in 330+ sets of experiments. Our analysis focused on measuring the performance gap between SOTA models and LLMs. The overarching trend observed was that SOTA models generally outperformed LLMs in zero-shot learning, with a few exceptions. Notably, larger computational models with few-shot learning techniques managed to reduce these performance gaps. Our findings provide valuable insights into the applicability of LLMs for Arabic NLP and speech processing tasks.

Despite their remarkable ability to capture linguistic nuances across diverse languages, questions persist regarding the degree of alignment between languages in multilingual embeddings. Drawing inspiration from research on high-dimensional representations in neural language models, we employ clustering to uncover latent concepts within multilingual models. Our analysis focuses on quantifying the \textit{alignment} and \textit{overlap} of these concepts across various languages within the latent space. To this end, we introduce two metrics \CA{} and \CO{} aimed at quantifying these aspects, enabling a deeper exploration of multilingual embeddings. Our study encompasses three multilingual models (\texttt{mT5}, \texttt{mBERT}, and \texttt{XLM-R}) and three downstream tasks (Machine Translation, Named Entity Recognition, and Sentiment Analysis). Key findings from our analysis include: i) deeper layers in the network demonstrate increased cross-lingual \textit{alignment} due to the presence of language-agnostic concepts, ii) fine-tuning of the models enhances \textit{alignment} within the latent space, and iii) such task-specific calibration helps in explaining the emergence of zero-shot capabilities in the models.\footnote{The code is available at \url{https://github.com/baselmousi/multilingual-latent-concepts}}

Recent advancements in the field of natural language processing have markedly enhanced the capability of machines to comprehend human language. However, as language models progress, they require continuous architectural enhancements and different approaches to text processing. One significant challenge stems from the rich diversity of languages, each characterized by its distinctive grammar resulting in a decreased accuracy of language models for specific languages, especially for low-resource languages. This limitation is exacerbated by the reliance of existing NLP models on rigid tokenization methods, rendering them susceptible to issues with previously unseen or infrequent words. Additionally, models based on word and subword tokenization are vulnerable to minor typographical errors, whether they occur naturally or result from adversarial misspellings. To address these challenges, this paper presents the utilization of a recently proposed free-tokenization method, such as Cannine, to enhance the comprehension of natural language. Specifically, we employ this method to develop an Arabic-free tokenization language model. In this research, we will precisely evaluate our model’s performance across a range of eight tasks using Arabic Language Understanding Evaluation (ALUE) benchmark. Furthermore, we will conduct a comparative analysis, pitting our free-tokenization model against existing Arabic language models that rely on sub-word tokenization. By making our pre-training and fine-tuning models accessible to the Arabic NLP community, we aim to facilitate the replication of our experiments and contribute to the advancement of Arabic language processing capabilities. To further support reproducibility and open-source collaboration, the complete source code and model checkpoints will be made publicly available on our Huggingface1 . In conclusion, the results of our study will demonstrate that the free-tokenization approach exhibits comparable performance to established Arabic language models that utilize sub-word tokenization techniques. Notably, in certain tasks, our model surpasses the performance of some of these existing models. This evidence underscores the efficacy of free-tokenization in processing the Arabic language, particularly in specific linguistic contexts.

The recent development and success of Large Language Models (LLMs) necessitate an evaluation of their performance across diverse NLP tasks in different languages. Although several frameworks have been developed and made publicly available, their customization capabilities for specific tasks and datasets are often complex for different users. In this study, we introduce the LLMeBench framework. Initially developed to evaluate Arabic NLP tasks using OpenAI's GPT and BLOOM models; it can be seamlessly customized for any NLP task and model, regardless of language. The framework also features zero- and few-shot learning settings. A new custom dataset can be added in less than 10 minutes, and users can use their own model API keys to evaluate the task at hand. The developed framework has been already tested on 31 unique NLP tasks using 53 publicly available datasets within 90 experimental setups, involving approximately 296K data points. We plan to open-source the framework for the community (https://github.com/qcri/LLMeBench/). A video demonstrating the framework is available online (https://youtu.be/FkQn4UjYA0s).

The abundance of social media data in the Arab world, specifically on Twitter, enabled companies and entities to exploit such rich and beneficial data that could be mined and used to extract important information, including sentiments and opinions of people towards a topic or a merchandise. However, with this plenitude comes the issue of producing models that are able to deliver consistent outcomes when tested within various contexts. Although model generalization has been thoroughly investigated in many fields, it has not been heavily investigated in the Arabic context. To address this gap, we investigate the generalization of models and data in Arabic with application to sentiment analysis, by performing a battery of experiments and building different models that are tested on five independent test sets to understand their performance when presented with unseen data. In doing so, we detail different techniques that improve the generalization of machine learning models in Arabic sentiment analysis, and share a large versatile dataset consisting of approximately 1.64M Arabic tweets and their corresponding sentiment to be used for future research. Our experiments concluded that the most consistent model is trained using a dataset labelled by a cascaded approach of two models, one that labels neutral tweets and another that identifies positive/negative tweets based on the Arabic emoji lexicon after class balancing. Both the BERT and the SVM models trained using the refined data achieve an average F-1 score of 0.62 and 0.60, and standard deviation of 0.06 and 0.04 respectively, when evaluated on five diverse test sets, outperforming other models by at least 17% relative gain in F-1. Based on our experiments, we share recommendations to improve model generalization for classification tasks.

With large Foundation Models (FMs), language technologies (AI in general) are entering a new paradigm: eliminating the need for developing large-scale task-specific datasets and supporting a variety of tasks through set-ups ranging from zero-shot to few-shot learning. However, understanding FMs capabilities requires a systematic benchmarking effort by comparing FMs performance with the state-of-the-art (SOTA) task-specific models. With that goal, past work focused on the English language and included a few efforts with multiple languages. Our study contributes to ongoing research by evaluating FMs performance for standard Arabic NLP and Speech processing, including a range of tasks from sequence tagging to content classification across diverse domains. We start with zero-shot learning using GPT-3.5-turbo, Whisper, and USM, addressing 33 unique tasks using 59 publicly available datasets resulting in 96 test setups. For a few tasks, FMs performs on par or exceeds the performance of the SOTA models but for the majority it under-performs. Given the importance of prompt for the FMs performance, we discuss our prompt strategies in detail and elaborate on our findings. Our future work on Arabic AI will explore few-shot prompting, expand the range of tasks, and investigate additional open-source models.