nb-asr-numerics-categorized
收藏资源简介:
该数据集是挪威语(Bokmål)数字表达式分类数据集,代表了挪威语数字数据处理流程的第二阶段。其核心目标是对第一阶段从大量文本中收集到的、包含数字或数字表达式的挪威语(Bokmål)句子进行语义验证和精细分类。数据集的主要用途是为后续生成新的、有效的、包含不同数字值的句子提供高质量的模板基础。在处理过程中,每个句子都会根据一个封闭的、包含超过60个预定义类别(如基数、序数、数量、日期、年份、时间、价格、百分比、测量值、标识号等)的本体进行分类。核心标注决策是二元的:保留(keep)或丢弃(drop)。保留的句子中的所有数字实体都必须被精确分类,并且可以标注实体之间的语义关系约束(如小于、大于、时间顺序等)。丢弃则适用于非挪威语、语法错误、不完整、数字表达式无效、功能不明确或无法分类等情况。数据集通过指令微调的大语言模型(google/gemma-4-12B-it)进行自动化处理,并遵循严格的提示模板(categorize_v1)以确保一致性和质量。最终数据集包含约176.9万条被保留的句子(保留率约46.4%)和相应的结构化标注信息,适用于文本生成(尤其是数字替换)、信息抽取、自然语言理解以及挪威语计算语言学相关的研究与开发。
This dataset is a Norwegian (Bokmål) numerical expression classification dataset, representing the second stage of the Norwegian numerical data processing pipeline. Its core objective is to perform semantic validation and fine-grained classification of Norwegian (Bokmål) sentences containing numbers or numerical expressions, which were collected from large-scale texts in the first stage. The primary purpose of the dataset is to provide a high-quality template foundation for subsequently generating new, valid sentences with different numerical values. During processing, each sentence is classified based on a closed ontology with over 60 predefined categories (such as cardinal numbers, ordinal numbers, quantities, dates, years, time, prices, percentages, measurements, identification numbers, etc.). The core annotation decision is binary: keep or drop. All numerical entities in keep sentences must be precisely classified, and semantic relationship constraints between entities (e.g., less than, greater than, temporal order, etc.) can be annotated. Drop applies to cases such as non-Norwegian language, grammatical errors, incompleteness, invalid numerical expressions, ambiguous functionality, or unclassifiable content. The dataset is processed automatically using an instruction-tuned large language model (google/gemma-4-12B-it) and follows a strict prompt template (categorize_v1) to ensure consistency and quality. The final dataset contains approximately 1.769 million retained sentences (retention rate about 46.4%) with corresponding structured annotation information, suitable for text generation (especially numerical substitution), information extraction, natural language understanding, and research and development related to Norwegian computational linguistics.
数据集概述
数据集名称: Norwegian Bokmål Numeric Expression Categorized Dataset
描述: 该数据集是挪威语(书面挪威语)数字数据处理流程的第二阶段(Stage 2),包含对第一阶段收集的数字表达式句子进行的语义验证和分类注释。
任务与语言
- 任务类别: 文本生成
- 语言: 挪威语 (书面挪威语, nb)
来源与处理
- 源数据集: pere/nb-asr-numerics-harvested (约380万行,分为6个分片)
- 推断模型:
google/gemma-4-12B-it(指令微调模型,BF16精度) - 并行化: 6个独立的单GPU任务(在H200系统上的0-5号GPU上运行)
- 提示模板版本:
categorize_v1 - 解码设置:
do_sample: false,temperature: 0.0,top_p: 1.0
核心注释策略
- 所有数字表达式必须分配到封闭本体论中的类别。
- 二元决策:
keep(保留)或drop(丢弃)。不确定性、畸形或不受支持的类别强制drop决策。 - 重叠范围和未匹配的数字会被严格验证,强制
drop决策。 - 不要求事实正确性,仅要求语法和语义上下文有效性。
数据模式
- categories: 句子中唯一、排序的类别名称列表。
- category_signature: 以"+"分隔的、排序的类别名称(无重复)。
- entity_count: 句子中数字实体的总数。
- has_constraints: 句子是否有数字约束。
- decision:
keep(保留) 或drop(丢弃)。 - entities: 已注释的语义实体(偏移量、文本、类别、预跨度映射)。
- constraints: 结构和语义关系约束。
聚合运行统计
- 总源行数: 3,814,821
- 保留行数: 1,769,510
- 丢弃行数: 2,045,311
- 保留率: 0.46385
- 首次通过JSON失败数: 841,335
- 重试次数: 841,335
- 重试成功率: 0.75369
- 模型输出失败数: 207,228
- 多类别句子数: 131,586
- 复合实体数: 63,675
类别计数(前20位)
- quantity: 100,001
- date: 100,000
- percentage: 100,000
- year: 100,000
- duration: 94,904
- measurement: 77,615
- age: 74,329
- price: 73,724
- cardinal: 59,851
- range: 57,257
- product_model: 53,419
- legal_section: 46,814
- distance: 43,884
- time: 42,677
- ordinal: 38,075
- chapter_number: 34,746
- code: 30,458
- money: 22,446
- weight: 18,800
- score: 18,800
许可证
- 许可证: cc-by-4.0





