MehdiAstaraki/multi-lingual-qac-alias-graph

Multi-lingual chemical QAC — Alias-Graph Retrieval benchmark

Given a chemistry concept (named in several languages), can a retriever find the documents that genuinely talk about it, across languages, without being fooled by documents about chemically similar look-alike concepts?

Configs: corpus (gold + hard-negative documents), queries (technical questions about each concept, plus the concept's multilingual name_set and the source_publication each query was generated from), qrels (every document about the concept: gold docs = score 1, designated look-alike docs = score 0), source_qrels (the exact publication each query was generated from and its translations — the gold document in every language), hard_negatives (each look-alike's neighbor concept and relation), qac (full triplets), and concepts (the per-concept alias-graph record). Each config has a train split.

Data source and license

• Source dataset: Patent text (titles, abstracts) in this dataset is derived from Google Patents Public Data on BigQuery (patents-public-data.patents.publications), provided by IFI CLAIMS Patent Services and Google. See Marketplace and announcement.
• License: That source data is made available under CC BY 4.0 (Creative Commons Attribution 4.0).
• This dataset: The corpus, questions, and answers (including all Q&A pairs and translations) form a derived/adapted dataset based on that source.
• No endorsement: This dataset is not affiliated with, endorsed by, or officially connected with Google or IFI CLAIMS. Only the underlying patent publication text is from that source; the Q&A generation and benchmark design are independent.
• Scope: Attribution and license refer only to the patent dataset content (bibliographic and abstract text from the public BigQuery tables). They do not cover other Google services, products, or UI content.