Enabling Deep Learning for Multilingual Sociopragmatics

为多语言社交语用学提供深度学习

• 批准号: RGPIN-2018-04267
• 负责人: AbdulMageed, Muhammad
• 金额: $ 2.04万
• 依托单位: University of British Columbia
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2020
• 资助国家: 加拿大
• 起止时间: 2020-01-01 至 2021-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=712829
• 关键词: Enabling; Deep; Learning; Multilingual; Sociopragmatics

Natural language processing (NLP) is the exciting field focused at teaching computers to understand and generate human language. Recently, deep learning, a class of machine learning methods inspired by information processing in the human brain, has broken records on many NLP tasks for which large amounts of labeled data are available (e.g., machine translation, speech recognition). Due to these advances and the pervasive technologies it enables, deep learning of natural language is currently a strategic area of high socioeconomic impact. This makes it ripe time for creating models that understand human language at the level of sociopragmatics (i.e., the meaning of a proposition depends on the social context in which it is uttered). Many challenges, however, remain. Two prominent, inter-related, examples are (a) the high costs associated to labeling data, and (b) the bias in existing labeled data. Absence of labeled data hinders progress on building powerful deep learning models since these models scale exclusively given large amounts of labeled data. Biased labeled data result in creating technologies that serve particular dominant groups better than others, which can have serious social and economic repercussions. My research program aims at developing methods to accelerate deep learning of natural language at the level of sociopragmatics by targeting these two core problems, with a focus on bringing NLP technologies to wider demographics across several languages and language varieties. The proposal has three key objectives: 1. Cross-Lingual Surrogate Labeling: This involves developing methods for automatically labeling data for sociopragmatic tasks (e.g., user intention modeling, empathy detection), with a focus on English and all Arabic varieties (i.e., varieties representing all the 22 Arab countries). 2. Deep Generative Semi-Supervised Learning: I will develop methods that exploit deep generative models, a class of deep learning methods that can generate sensible language that can be leveraged as labeled data. This will help solve the two data-focused problems above (i.e., a and b). 3. Toward Social Machines With Controlled Sociopragmatics: My goal is to develop sociopragmatically intelligent conversational models capable of dynamic customization in response to conversant attributes (e.g., emotionally intelligent language generation, gender- and personality-specific conversational agents). The research will have a wide range of applications in various fields, including decision making, health and well-being, education, recreation, and entertainment. Since it is a specialized subfield at the junction of a number of already supply-constrained fields, deep learning of natural language currently suffers from acute shortage of talent. HQP training provided by my program will contribute to fulfilling these ever-rising needs.

自然语言处理（NLP）是一个令人兴奋的领域，专注于教计算机理解和生成人类语言。最近，深度学习，一类受人脑信息处理启发的机器学习方法，在许多需要大量标记数据的NLP任务（例如，机器翻译，语音识别）上打破了记录。由于这些进步及其所带来的普及技术，自然语言的深度学习目前是一个具有高度社会经济影响的战略领域。这使得创建在社会语用学层面上理解人类语言的模型的时机成熟（即，命题的意义取决于它所说的社会背景）。然而，许多挑战依然存在。两个突出的、相互关联的例子是(a)与标记数据相关的高成本，以及(b)现有标记数据中的偏差。缺乏标记数据阻碍了构建强大的深度学习模型的进展，因为这些模型只能在给定大量标记数据的情况下进行扩展。有偏见的标签数据导致创造出比其他群体更好地服务于特定主导群体的技术，这可能会产生严重的社会和经济影响。