Augmenting text data for sentence classification using weakly supervised multi-reward reinforcement learning

• Main Author: ARAKI JUN
• Format: Patent
• Language: Chinese; English
• Published: 09.09.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

Text data for sentence classification is augmented using weak supervised multi-reward reinforcement learning. A system and method are disclosed that enable fast and cost effective human-in-loop synthesis of domain-specific text training data for deep learning models. The data expansion process comprises a sentence generator, a sentence classifier and weak supervision of field experts in a loop. Typically, both a sentence generator and a sentence classifier are implemented as machine learning models. A sentence generator generates new sentences based on the artificially tagged sentences, and a sentence classifier generates tags for the newly generated sentences. The new sentence is corrected or validated by a domain expert and then used to retrain one or both of the sentence generator and the sentence classifier. 使用弱监督多奖励强化学习扩充用于句子分类的文本数据。公开了一种系统和方法，其使得能够实现针对深度学习模型的领域特定文本训练数据的快速且成本有效的人在回路合成。数据扩充过程包括句子生成器、句子分类器和"回路中"的领域专家的弱监督。通常，句子生成器和句子分类器两者都被实现为机器学习模型。句子生成器基于人工标记的句子生成新的句子，并且句子分类器针对新生成的句子生成标签。新句子由领域专家校正或验证，并且然