2026-10-14T12:44:43Z https://www.scidb.cn/oai

10.57760/sciencedb.35283 2026-09-23T16:19:14Z

2026-09-23 A Chinese-Tibetan Anchor-Enhanced Parallel Corpus – CT-AEPC doi:10.57760/sciencedb.35283 en Tibetan is a typical low resource language, and publicly available large-scale, high-quality datasets with parallel text, retrieval structures, and factual anchor annotations are still relatively limited. To this end, this article constructs a Chinese Tibetan anchor enhanced parallel corpus CT-AEPC. This corpus is based on publicly accessible Chinese webpage text and is formed through steps such as webpage body extraction, noise cleaning, Chinese sentence segment screening, Tibetan candidate translation generation, manual proofreading, anchor point extraction, quality scoring, duplicate sample removal, sensitive information filtering, and structured encapsulation. CT-AEPC contains a total of 99907 parallel text samples from China and Tibet, with a data format of JSONL and a data size of approximately 203.77MB. Each sample includes fields such as qid, query, positive, query_date, anchors, features, and quality_store. Among them, 'query' is a Chinese sentence segment, 'positive' is the corresponding Tibetan text, 'anchors' records verifiable factual anchor points such as date, time, and quantity,' features' records text length, anchor point quantity, language character ratio, length matching, and numerical consistency, and 'quality_store' represents the automatic structural quality assessment results of the sample. The statistical results show that the average length of Chinese text is 102.19 characters, and the average length of Tibetan text is 412.80 characters. Each sample contains an average of 4.01 factual anchor points, and the average automatic quality score is 0.9478. This language database can be used for tasks such as Chinese Tibetan machine translation, cross language retrieval, retrieval enhanced generation of positive samples, fine-tuning of low resource language models, evaluation of reordering models, and analysis of factual consistency.  Chinese–Tibetan parallel corpus; anchor enhancement; low-resource language; machine translation; cross-lingual retrieval; factual consistency Ran Bowen Wan Fucheng Shu Wenjun Yang Denghui PUBLIC https://creativecommons.org/licenses/by-nc-sa/4.0/ dataset Science Data Bank