Mixture-of-Supernets: Improving Weight-Sharing Supernet Training with Architecture-Routed Mixture-of-Experts

Weight-sharing supernet has become a vital component for performance estimation in the state-of-the-art (SOTA) neural architecture search (NAS) frameworks. Although supernet can directly generate different subnetworks without retraining, there is no guarantee for the quality of these subnetworks because of weight sharing. In NLP tasks such as machine translation and pre-trained language modeling, we observe that given the same model architecture, there is a large performance gap between supernet and training from scratch. Hence, supernet cannot be directly used and retraining is necessary after finding the optimal architectures. In this work, we propose mixture-of-supernets, a generalized supernet formulation where mixture-of-experts (MoE) is adopted to enhance the expressive power of the supernet model, with negligible training overhead. In this way, different subnetworks do not share the model weights directly, but through an architecture-based routing mechanism. As a result, model weights of different subnetworks are customized towards their specific architectures and the weight generation is learned by gradient descent. Compared to existing weight-sharing supernet for NLP, our method can minimize the retraining time, greatly improving training efficiency. In addition, the proposed method achieves the SOTA performance in NAS for building fast machine translation models, yielding better latency-BLEU tradeoff compared to HAT, state-of-the-art NAS for MT. We also achieve the SOTA performance in NAS for building memory-efficient task-agnostic BERT models, outperforming NAS-BERT and AutoDistil in various model sizes.

N-Shot Benchmarking of Whisper on Diverse Arabic Speech Recognition

Whisper, the recently developed multilingual weakly supervised model, is reported to perform well on multiple speech recognition benchmarks in both monolingual and multilingual settings. However, it is not clear how Whisper would fare under diverse conditions even on languages it was evaluated on such as Arabic. In this work, we address this gap by comprehensively evaluating Whisper on several varieties of Arabic speech for the ASR task. Our evaluation covers most publicly available Arabic speech data and is performed under n-shot (zero-, few-, and full) finetuning. We also investigate the robustness of Whisper under completely novel conditions, such as in dialect-accented standard Arabic and in unseen dialects for which we develop evaluation data. Our experiments show that although Whisper zero-shot outperforms fully finetuned XLS-R models on all datasets, its performance deteriorates significantly in the zero-shot setting for five unseen dialects (i.e., Algeria, Jordan, Palestine, UAE, and Yemen).

On the Robustness of Arabic Speech Dialect Identification

Arabic dialect identification (ADI) tools are an important part of the large-scale data collection pipelines necessary for training speech recognition models. As these pipelines require application of ADI tools to potentially out-of-domain data, we aim to investigate how vulnerable the tools may be to this domain shift. With self-supervised learning (SSL) models as a starting point, we evaluate transfer learning and direct classification from SSL features. We undertake our evaluation under rich conditions, with a goal to develop ADI systems from pretrained models and ultimately evaluate performance on newly collected data. In order to understand what factors contribute to model decisions, we carry out a careful human study of a subset of our data. Our analysis confirms that domain shift is a major challenge for ADI models. We also find that while self-training does alleviate this challenges, it may be insufficient for realistic conditions.

GPTAraEval: A Comprehensive Evaluation of ChatGPT on Arabic NLP

The recent emergence of ChatGPT has brought a revolutionary change in the landscape of NLP. Although ChatGPT has consistently shown impressive performance on English benchmarks, its exact capabilities on most other languages remain largely unknown. To better understand ChatGPT's capabilities on Arabic, we present a large-scale evaluation of the model on a broad range of Arabic NLP tasks. Namely, we evaluate ChatGPT on 32 diverse natural language understanding and generation tasks on over 60 different datasets. To the best of our knowledge, our work offers the first performance analysis of ChatGPT on Arabic NLP at such a massive scale. Our results show that, despite its success on English benchmarks, ChatGPT trained in-context (few-shot) is consistently outperformed by much smaller dedicated models finetuned on Arabic. These results suggest that there is significant place for improvement for instruction-tuned LLMs such as ChatGPT.

Dolphin: A Challenging and Diverse Benchmark for Arabic NLG

We present Dolphin, a novel benchmark that addresses the need for an evaluation framework for the wide collection of Arabic languages and varieties. The proposed benchmark encompasses a broad range of 13 different NLG tasks, including text summarization, machine translation, question answering, and dialogue generation, among others. Dolphin comprises a substantial corpus of 40 diverse and representative public datasets across 50 test splits, carefully curated to reflect real-world scenarios and the linguistic richness of Arabic. It sets a new standard for evaluating the performance and generalization capabilities of Arabic and multilingual models, promising to enable researchers to push the boundaries of current methodologies. We provide an extensive analysis of Dolphin, highlighting its diversity and identifying gaps in current Arabic NLG research. We also evaluate several Arabic and multilingual models on our benchmark, allowing us to set strong baselines against which researchers can compare.

LaMini-LM: A Diverse Herd of Distilled Models from Large-Scale Instructions

Large language models (LLMs) with instruction finetuning demonstrate superior generative capabilities. However, these models are resource intensive. To alleviate this issue, we explore distilling knowledge from instruction-tuned LLMs to much smaller ones. To this end, we carefully develop a large set of 2.58M instructions based on both existing and newly-generated instructions. In addition to being sizeable, we design our instructions to cover a broad set of topics to ensure. A thorough investigation of our instruction data demonstrate their diversity, and we generate responses for these instructions using gpt-3.5-turbo. We then exploit the instructions to tune a host of models, dubbed LaMini-LM, of varying sizes, both from the encoder-decoder as well as the decoder-only families. We evaluate our models both automatically (on 15 different NLP benchmarks) and manually. Results show that our proposed LaMini-LM are on par with competitive baselines while being nearly 10 times smaller in size.

Zero-Shot Slot and Intent Detection in Low-Resource Languages

Intent detection and slot filling are critical tasks in spoken and natural language understanding for task-oriented dialog systems. In this work we describe our participation in the slot and intent detection for low-resource language varieties (SID4LR; Aepli et al. (2023)). We investigate the slot and intent detection (SID) tasks using a wide range of models and settings. Given the recent success of multitask-prompted finetuning of large language models, we also test the generalization capability of the recent encoder-decoder model mT0 (Muennighoff et al., 2022) on new tasks (i.e., SID) in languages they have never intentionally seen. We show that our best model outperforms the baseline by a large margin (up to +30 F1 points) in both SID tasks

UBC-DLNLP at SemEval-2023 Task 12: Impact of Transfer Learning on African Sentiment Analysis

We describe our contribution to the SemEVAl 2023 AfriSenti-SemEval shared task, where we tackle the task of sentiment analysis in 14 different African languages. We develop both monolingual and multilingual models under a full supervised setting (subtasks A and B). We also develop models for the zero-shot setting (subtask C). Our approach involves experimenting with transfer learning using six language models, including further pertaining of some of these models as well as a final finetuning stage. Our best performing models achieve an F1-score of 70.36 on development data and an F1-score of 66.13 on test data. Unsurprisingly, our results demonstrate the effectiveness of transfer learning and fine-tuning techniques for sentiment analysis across multiple languages. Our approach can be applied to other sentiment analysis tasks in different languages and domains.

ORCA: A Challenging Benchmark for Arabic Language Understanding

Due to their crucial role in all NLP, several benchmarks have been proposed to evaluate pretrained language models. In spite of these efforts, no public benchmark of diverse nature currently exists for evaluation of Arabic. This makes it challenging to measure progress for both Arabic and multilingual language models. This challenge is compounded by the fact that any benchmark targeting Arabic needs to take into account the fact that Arabic is not a single language but rather a collection of languages and varieties. In this work, we introduce ORCA, a publicly available benchmark for Arabic language understanding evaluation. ORCA is carefully constructed to cover diverse Arabic varieties and a wide range of challenging Arabic understanding tasks exploiting 60 different datasets across seven NLU task clusters. To measure current progress in Arabic NLU, we use ORCA to offer a comprehensive comparison between 18 multilingual and Arabic language models. We also provide a public leaderboard with a unified single-number evaluation metric (ORCA score) to facilitate future research.

SERENGETI: Massively Multilingual Language Models for Africa

Multilingual language models (MLMs) acquire valuable, generalizable linguistic information during pretraining and have advanced the state of the art on task-specific finetuning. So far, only ~ 28 out of ~2,000 African languages are covered in existing language models. We ameliorate this limitation by developing SERENGETI, a set of massively multilingual language model that covers 517 African languages and language varieties. We evaluate our novel models on eight natural language understanding tasks across 20 datasets, comparing to four MLMs that each cover any number of African languages. SERENGETI outperforms other models on 11 datasets across the eights tasks and achieves 82.27 average F-1. We also perform error analysis on our models' performance and show the influence of mutual intelligibility when the models are applied under zero-shot settings. We will publicly release our models for research.