Blackbird Language Matrices (BLM)
收藏资源简介:
Blackbird语言矩阵(BLM)是由Idiap研究所和日内瓦大学联合开发的创新型语言任务数据集,旨在通过结构化多选问题探究语言模型的系统性泛化能力。该数据集包含法语、意大利语和英语等多种语言的语法现象(如动词一致性、论元结构),通过人工构建和规则生成相结合的方式,形成包含上下文序列和对比答案的复杂语言谜题。其数据特点在于多层级结构设计(句子内、跨句子、候选答案间),支持对模型语言对象识别、模式检测等核心能力的多维度评估。该数据集主要应用于自然语言处理领域,用于诊断语言模型的语法归纳、结构依赖和组合系统性等类人语言能力,并为可解释性研究提供结构化数据支持。
The Blackbird Language Matrix (BLM) is an innovative language task dataset co-developed by Idiap Research Institute and the University of Geneva, designed to investigate the systematic generalization capabilities of language models via structured multiple-choice questions. This dataset covers linguistic phenomena across multiple languages including French, Italian and English, such as verb agreement and argument structure. It combines manual construction and rule-based generation to create complex linguistic puzzles that contain contextual sequences and contrasting answer options. The dataset features a multi-level structural design (intra-sentence, cross-sentence, and between candidate answers), enabling multi-dimensional evaluation of core capabilities of language models such as linguistic object recognition and pattern detection. This dataset is primarily applied in the field of natural language processing, used to diagnose human-like linguistic abilities of language models including grammar induction, structural dependency and compositional systematicity, and provides structured data support for interpretability research.



