Emulated-Inc/reading-comprehension-qa-training-pool

Reading comprehension question answering training pool

Public question answering data from five datasets, every row a question, the passages it is answered from and every acceptable answer, read at the pinned revisions named below and laid out twice. Train on either layer or on both.

pool.jsonl

Every source rewritten into one shape, 311462 rows, one JSON object per line, with these fields.

Field	What it holds
id	a row identifier unique within this file
question	the question text, as its source publishes it
passages	the passages the question is answered from, each a title and a text
answers	every acceptable form of the answer, as a list of strings
n_passages	how many passages the row carries
source	the name of the source directory the row came from
source_repo, source_revision	the dataset and the revision it was read at
source_subset, source_split, source_id	where the row sits in that set
provenance_class	how the row came to exist
licence	the licence of the source it came from

A row carries at most five passages and the ones that carry an answer come first. How many there are depends on what the source offers, and the section for each source below says. A row is only here when one of its answers appears in the passages it carries, compared after lowercasing, deleting punctuation, removing the articles and collapsing whitespace, so every row can be answered from what it holds. Rows are deduplicated across sources on the normalised question, keeping the first source that carries it in the order of the sections below.

sources/

The same data untouched, 374104 rows, one directory per source, holding the files at the paths, in the format and with the fields its own repository publishes. Nothing here was renamed, reshaped, reordered or deduplicated. Use this layer if you want a field the rewritten one drops, such as the character offset of a SQuAD answer, the sentence-level supporting facts of HotpotQA or the answer types of DROP, or if you would rather choose the passages yourself. There is one exception to untouched, and it is that the two retrieval sources keep at most 10 retrieved passages per row rather than the hundred they publish. Their published files are 1.3 GB each and nine tenths of those bytes are passages ranked below the tenth.

The sources

sources/natural_questions

Real queries issued to a search engine, with the answers an annotator marked on the Wikipedia page the query returned, paired here with hundred-word passages of a December 2018 Wikipedia dump. From Tevatron/wikipedia-nq at revision 37a7d5bb286c76f1b406d76a7434e8eeb4385203, files nq- train.jsonl.gz. 58519 rows here, of which 57373 also appear in pool.jsonl, each with one passage that contains an answer and four of the highest-ranked retrieved passages that do not. Provenance class human, licence cc-by-sa-3.0. Its own fields are query_id, query, answers, positive_passages, negative_passages.

Worth knowing. The repository carries no licence tag. The questions and answers are Natural Questions under CC BY-SA 3.0 and the passages are Wikipedia under the same term, which is what this pool records.

sources/triviaqa

Trivia questions written by enthusiasts, with the long alias lists their answers carry, paired here with passages of the same Wikipedia dump. From Tevatron/wikipedia-trivia at revision e73506da2be3e431b2900ad58f7cb983818e27a4, files trivia-train.jsonl.gz. 60149 rows here, of which 59713 also appear in pool.jsonl, each with one passage that contains an answer and four of the highest-ranked retrieved passages that do not. Provenance class human, licence apache-2.0. Its own fields are query_id, query, answers, positive_passages, negative_passages.

Worth knowing. The passages were matched to the questions by distant supervision, so a passage can carry the answer string without being the evidence for it. The alias lists are long and generous, which is the published way to score this set.

sources/squad

Crowdworker questions written against a paragraph of a Wikipedia article, each answered by a span of that paragraph. From rajpurkar/squad at revision 7b6d24c440a36b6815f21b70d25016731768db1f, files plain_text/train-00000-of-00001.parquet. 87594 rows here, of which 85658 also appear in pool.jsonl, each with the one paragraph the question was written against. Provenance class human, licence cc-by-sa-4.0. Its own fields are id, title, context, question, answers with parallel text and answer_start lists.

Worth knowing. The questions were written by people looking at the paragraph, so they lean on its wording more than a search query does.

sources/hotpotqa

Questions that need two Wikipedia paragraphs to answer, shipped with those two paragraphs and eight retrieved distractors. From hotpotqa/hotpot_qa at revision 1908d6afbbead072334abe2965f91bd2709910ab, files distractor/train-00000-of-00002.parquet, distractor/train-00001-of-00002.parquet. 90442 rows here, of which 85027 also appear in pool.jsonl, each with the two supporting paragraphs named by supporting_facts, then distractors up to five. Provenance class human, licence cc-by-sa-4.0. Its own fields are id, question, answer, type, level, supporting_facts, context.

Worth knowing. Its comparison questions are answered yes or no, and those rows do not survive the rule that an answer must appear in the passages kept, so the rewritten layer holds the span- answered part of this set and the untouched layer holds all of it.

sources/drop

Questions over a paragraph of narrative text that need counting, sorting or arithmetic over what the paragraph says. From ucinlp/drop at revision 95cda593fae71b60b5b19f82de3fcf3298c1239c, files data/train-00000-of-00001.parquet. 77400 rows here, of which 23691 also appear in pool.jsonl, each with the one paragraph the question is about. Provenance class human, licence cc-by-sa-4.0. Its own fields are section_id, query_id, passage, question, answers_spans with spans and types.

Worth knowing. Many of its answers are computed rather than quoted, and those rows do not survive the rule that an answer must appear in the passage kept, so the rewritten layer holds its quotable part. It also publishes no passage title, so the section identifier stands in for one.

Provenance and licences

Every row was written by a person. All five sources are questions people wrote, either as real search queries, as trivia, or as reading questions about a paragraph in front of them, which is the provenance class human. Nothing in the pool was generated by a model.

The pool as a whole is offered under cc-by-sa-4.0, which is the most restrictive term its sources compose to. SQuAD, HotpotQA and DROP are cc-by-sa-4.0, the Natural Questions rows and every Wikipedia passage are cc-by-sa-3.0, and TriviaQA is apache-2.0. Each rewritten row carries its own in the licence field and each directory under sources/ is one source, so a subset under a single licence can be selected. Attribution for the share-alike sources goes to their authors and to the Wikipedia contributors whose text the passages are.

Filtering

Rows whose question duplicated a question in a held-out evaluation set were removed before publication, from both layers alike, by an exact comparison of the normalised question (2 rows) and then by a word 8-gram overlap check (258 rows). That evaluation set is not distributed here. Only the question was compared. The passages were not, because they are public Wikipedia that every source here draws on. Nothing else was filtered out of the untouched layer: no subject, no difficulty and no answer type was selected for or against. The rewritten layer holds fewer rows than the untouched one, 53587 of them because no answer of the row appears in the passages the row carries, which is what happens to a yes or no comparison question and to a computed number, and 9055 because another source already carried the same question.