TibNER:Tibetan Named Entity Recognition Dataset

Structured linguistic resources are an important foundation for natural language processing. Currently, due to the lack of open-source datasets, the research on Tibetan named entity recognition progresses slowly and the results accumulate less. Based on this, this paper semi-automatically constructs a Tibetan named entity recognition dataset (TibNER) using an entity dictionary. In order to ensure the quality of the dataset, the automatic annotation results are manually proofread.TibNER contains 20,096 sentences, with an average sentence length of 44.2069 syllables, and the annotated entities include names of people, places, and organizations, with a total number of 43,678 in the three types of entities.In order to validate the validity of the dataset, this paper conducts a comparative test on three types of mainstream sequence annotation models, with an F1 value of up to 80.60%. After the study, this data provides data construction experience for low-resource languages, and provides certain data basis for studies such as Tibetan named entity recognition.结构化语言资源是自然语言处理（Natural Language Processing, NLP）的重要基础。当前，由于开源数据集的匮乏，藏语命名实体识别领域的研究进展缓慢，相关成果积累有限。为此，本文依托实体词典，采用半自动方式构建了藏语命名实体识别数据集（Tibetan Named Entity Recognition, TibNER）。为保障数据集质量，本文对自动标注结果进行了人工校对。该数据集共计包含20096条语句，平均句长为44.2069个音节；标注实体涵盖人名、地名与组织机构名三类，三类实体总数量达43678个。为验证数据集的有效性，本文针对三类主流序列标注模型开展了对比实验，最高F1值可达80.60%。本研究不仅为低资源语言的数据集构建提供了实践经验，同时也为藏语命名实体识别等相关研究提供了可靠的数据支撑。

搜集汇总

背景与挑战

背景概述

TibNER是一个针对低资源藏语构建的命名实体识别数据集，通过基于实体词典的半自动方法创建并经过人工校对，包含20,096个句子和43,678个人名、地名、组织名实体标注。该数据集在主流序列标注模型上验证有效，最高F1值达80.60%，为藏语自然语言处理研究提供了重要数据基础。