Computing servers for NLP applications

NLP应用的计算服务器

• 批准号: RTI-2022-00466
• 负责人: Nie, JianYun
• 金额: $ 10.77万
• 依托单位: Université de Montréal
• 依托单位国家: 加拿大
• 项目类别: Research Tools and Instruments
• 财政年份: 2021
• 资助国家: 加拿大
• 起止时间: 2021-01-01 至 2022-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=742172
• 关键词: Computing; servers; NLP; applications

This RTI application aims to equip the RALI lab on Computational Linguistics and Applications (Recherche appliquée en linguistique informatique) with adequate computation resources necessary for our research. The RALI lab is specialized in natural language processing (NLP) and applications. Our research addresses both fundamental and application questions related to NLP: how can we understand languages and how can we use natural languages in different applications. Our research work covers a large number of topics such as information retrieval (search engine), information extraction, question answering, dialogue systems, recommendation systems, social media mining, medical information retrieval and question answering, and so on. All these applications are highly in demand in industry, and more generally in our society. The RALI lab has been performing high-quality research work, widely recognized by the peers. We publish at the very top conferences and journals in this domain and our research work has made significant impact in the community (e.g. Test of Time honorable mention of ACM SIGIR in information retrieval). However, our current equipment is no longer adequate to support our research in NLP and applications. The domain has made a drastic move in recent years toward methods based on deep learning and deep neural networks, that exploit huge amounts of textual data and large pre-trained models (from 0.34 billions parameter for BERT to 175 billion for GPT-3). Intensive computation on GPU servers is required to benefit from these models. The NLP domain is becoming more competitive than ever, in the sense that not only innovative ideas, but also powerful computation resources, are required. While most top NLP labs have updated their computer servers to meet the need, the RALI lab is well behind: we only have a small number of workstations with GPU cores shared by 3 faculty members, 3 research associates and over 40 graduate students, most of whom use neural models and require GPUs. The poor equipment is causing serious delays in our research and publications. The requested equipment - GPU server and storage server, will allow us to remain competitive in NLP research and offer adequate environment for HQP training. The equipment will also help us attract more good students, especially female students and students from visible minorities, for whom the computation environment is an important criterion when choosing a lab for their studies.

该 RTI 应用程序旨在为 RALI 计算语言学和应用（Recherche appliquée en linguistique informatique）实验室配备我们研究所需的充足计算资源。 RALI 实验室专门从事自然语言处理 (NLP) 和应用研究。以及与NLP相关的应用问题：我们如何理解语言以及我们如何在不同的应用中使用自然语言。我们的研究工作涵盖了信息检索（搜索）等大量主题。引擎）、信息提取、问答、对话系统、推荐系统、社交媒体挖掘、医疗信息检索和问答等，所有这些应用在工业界以及更广泛的社会中都有很高的需求。我们一直在开展高质量的研究工作，受到同行的广泛认可，我们在该领域的顶级会议和期刊上发表文章，并且我们的研究工作在社区中产生了重大影响（例如，ACM SIGIR 在信息领域获得了 Test of Time 荣誉奖）。然而，我们现有的设备已不足以支持我们在 NLP 和应用方面的研究，近年来，该领域已经朝着基于深度学习和深度神经网络的方法迈出了一大步，这些方法利用了大量的文本数据和数据。大型预训练模型（从 BERT 的 3.4 亿个参数到 GPT-3 的 1750 亿个参数）需要在 GPU 服务器上进行密集计算才能从这些模型中受益，从某种意义上说，NLP 领域的竞争变得比以往任何时候都更加激烈。不仅需要创新的想法，还需要强大的计算资源。虽然大多数顶级 NLP 实验室都更新了计算机服务器来满足需求，但 RALI 实验室却远远落后：我们只有少量的工作站，GPU 核心由 3 个共享。教职员工、3 名研究员和 40 多名研究生，其中大多数使用神经模型并需要 GPU，设备较差导致我们的研究和出版物严重延迟，而所需的设备 - GPU 服务器和存储服务器将使我们能够继续存在。 NLP 领域的竞争力这些设备还将帮助我们吸引更多优秀的学生，特别是女学生和有色人种学生，对于他们来说，计算环境是选择实验室学习的重要标准。