Scholarlink

Parallel corpora that align Knowledge Graph (KG) triples with natural language text serve as essential resources for training relation extraction and automatic KG construction systems. Although several existing datasets utilize Wikidata/DBpedia triples aligned to Wikipedia or synthetic text, Wikidata references, which point to diverse, real-world Web pages, remain underutilized. In this work, we propose using LLMs to extract and validate the text entailing a given triple in a web page, using few-shot in-context learning and Chain-Of-Thought prompting. With this, we build a dataset with 80.5K triple-to-text alignments with diverse text covering 90K entities and 910 relations. Dataset: https://huggingface.co/datasets/sven-h/wikidata_reference .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards Extracting Triple-to-Text Alignments from Wikidata References Using LLMs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nandana Mihindukulasooriya,
  • Sven Hertling

摘要

Parallel corpora that align Knowledge Graph (KG) triples with natural language text serve as essential resources for training relation extraction and automatic KG construction systems. Although several existing datasets utilize Wikidata/DBpedia triples aligned to Wikipedia or synthetic text, Wikidata references, which point to diverse, real-world Web pages, remain underutilized. In this work, we propose using LLMs to extract and validate the text entailing a given triple in a web page, using few-shot in-context learning and Chain-Of-Thought prompting. With this, we build a dataset with 80.5K triple-to-text alignments with diverse text covering 90K entities and 910 relations. Dataset: https://huggingface.co/datasets/sven-h/wikidata_reference .