bowang0911/DAPRNQHardChunkRetrieval

DAPR NQ-Hard — Chunk-level Retrieval Eval

Contextualized chunk-level (query2chunk) retrieval eval for NQ-Hard, the hard subset of Natural Questions from the DAPR benchmark (Document-Aware Passage Retrieval, Wang, Reimers & Gurevych, ACL 2024, arXiv:2305.13915).

NQ-Hard queries are hand-selected because understanding the document context is required to retrieve the relevant passage — the gold passage often refers to the query's entity only by coreference, main-topic ellipsis, an acronym, or via multi-hop reasoning. This makes it a focused probe of context-aware retrieval.

Corpus scope: gold-document-scoped

Each hard query is about one Wikipedia document (query_id == doc_id). The corpus here is the union of all passages of the 479 query documents — i.e. each query must find its gold passage(s) among the passages of the relevant document set.

This is the gold-document-scoped setting: a lightweight, self-contained probe of context-dependent passage selection (~16k passages, encodes in seconds). It is easier than DAPR's canonical full-corpus NQ setting (retrieval over 2.68M passages) and scores are not directly comparable to published DAPR numbers.

	count
queries	479
gold (query, passage) pairs	516
corpus passages	16,063
documents	479 (mean 33.5 passages/doc, max 228)

Hardness categories (per gold pair)

category	pairs
coreference	223
main_topic	205
multi-hop	88
acronym	13

A pair may have multiple categories. Use the queries.categories field to slice metrics by reasoning type — e.g. measure whether the documents_coref variant specifically lifts the 223 coreference queries.

Configs

documents / documents_coref

field	type	notes
chunk_id	string	DAPR passage id ({doc}-{paragraph})
chunk	string	passage text (coreference-resolved in documents_coref)
source_url	string	document id
title	string	document title
chunk_idx	int64	paragraph number within the document
chunk_start_char / chunk_end_char	int64	char offset of the passage in the reconstructed document

documents_coref reuses the same char offsets as documents (plain-text coordinate space) so the single queries config drives overlap-based gold matching for both configs; only the chunk text differs.

queries

field	type	notes
original_query / query	string	the query text
answer	list[string]	gold passage text(s)
score	list[int64]	relevance per gold passage
source_url	list[string]	document id per gold passage
frag_start_char / frag_end_char	list[int64]	char span of each gold passage
n_gold	int64	number of gold passages
categories	list[string]	hardness reason(s): coreference / main_topic / multi-hop / acronym
url	string	source Wikipedia URL

Gold matching

A chunk is gold if it shares source_url with a gold fragment and overlaps its [frag_start_char, frag_end_char). Each gold fragment is a whole passage and each chunk is a whole passage, so every gold maps to exactly one chunk (verified: 516 / 516 in both documents and documents_coref).

Provenance

Built from UKPLab/dapr nq-hard (queries + gold, with categories/url) and the DAPR NaturalQuestions-corpus / -corpus_coref (test), filtered to the 479 gold documents. Documents are reconstructed by concatenating passages in paragraph_no order (joined with \n) to assign char offsets.

Citation

@article{wang2023dapr,
    title  = "DAPR: A Benchmark on Document-Aware Passage Retrieval",
    author = "Kexin Wang and Nils Reimers and Iryna Gurevych",
    journal= "arXiv preprint arXiv:2305.13915",
    year   = "2023",
    url    = "https://arxiv.org/abs/2305.13915"
}