Blackbird's Language Matrices (BLMs)
收藏arXiv2022-05-23 更新2024-08-06 收录
下载链接：
http://arxiv.org/abs/2205.10866v1
下载链接
链接失效反馈官方服务：
资源简介：
Blackbird's Language Matrices (BLMs)是由日内瓦大学开发的一个新颖的语法数据集，旨在测试语言版本的Raven's Progressive Matrices智力测试。该数据集包含44800个句子，通过生成模型构建，用于研究当前模型对语法一致性规则的掌握及其泛化能力。数据集的创建过程涉及定义语言现象、规则的自动生成过程以及大规模数据样本的自动创建。BLMs数据集主要应用于理解神经网络的泛化和抽象能力，为语言任务提供了一个新的挑战性测试平台。
Blackbird's Language Matrices (BLMs) is a novel grammatical dataset developed by the University of Geneva, designed to serve as the linguistic counterpart of the classic Raven's Progressive Matrices intelligence test. This dataset contains 44,800 sentences constructed using generative models, and is used to investigate the mastery of grammatical agreement rules and the generalization capabilities of current models. The creation process of the BLMs dataset involves defining linguistic phenomena, automated rule generation procedures, and the automatic generation of large-scale data samples. Primarily, the BLMs dataset is applied to understanding the generalization and abstraction capabilities of neural networks, providing a new challenging testbed for linguistic tasks.
提供机构：
日内瓦大学
创建时间：
2022-05-23
