EntailmentBank

EntailmentBank is a dataset of multistep entailment trees for open-domain science question answering. Each example links a question and answer to a structured proof: a tree of multi-premise entailment steps from known facts, through intermediate conclusions, to a hypothesis.

This repository contains the EMNLP 2021 v2 release in JSONL format, with four configs:

Config	Description
task1	Generate an entailment-tree proof from gold supporting facts
task2	Generate a proof from gold facts plus distractor facts
task3	Retrieve supporting facts from a candidate pool (no gold proof)
worldtree_corpus	Extended WorldTree fact bank used to build contexts

Dataset Summary

EntailmentBank contains 1,840 science QA examples (1,313 train / 187 validation / 340 test), each annotated with:

• a natural-language question and answer
• a hypothesis combining the question and answer
• a context of candidate facts (sent1, sent2, ...)
• a gold proof in structured entailment-tree format (tasks 1 and 2)

The three task configs represent increasing difficulty:

1. Task 1 — all supporting facts are given; generate the proof.
2. Task 2 — supporting facts plus irrelevant distractors are given; generate the proof.
3. Task 3 — a larger candidate pool is given; retrieve the relevant facts (proof field is empty).

The worldtree_corpus config provides the underlying knowledge base of 11,941 science facts (9,011 from WorldTree + 2,930 EntailmentBank additions).

Supported Tasks

• Entailment-tree generation (task1, task2): generate a valid multistep proof from a context of facts.
• Supporting-fact retrieval (task3): select relevant facts from a candidate pool.
• Explainable QA: use structured proofs for interpretable reasoning chains.

Usage

from datasets import load_dataset

# Task 1: proof generation with gold facts
task1 = load_dataset("sxiong/entailmentbank", name="task1")
print(task1["train"][0]["question"])
print(task1["train"][0]["proof"])

# Task 2: proof generation with distractors
task2 = load_dataset("sxiong/entailmentbank", name="task2")

# Task 3: fact retrieval
task3 = load_dataset("sxiong/entailmentbank", name="task3")

# WorldTree extended corpus
corpus = load_dataset("sxiong/entailmentbank", name="worldtree_corpus")

Or load JSONL files directly:

from datasets import load_dataset

task1 = load_dataset(
    "json",
    data_files={
        "train": "task1/train.jsonl",
        "validation": "task1/validation.jsonl",
        "test": "task1/test.jsonl",
    },
)

Data Fields

task1 / task2

Field	Type	Description
id	string	Unique example ID
context	string	Candidate facts as sent1: ... sent2: ...
question	string	Science question
answer	string	Gold answer
hypothesis	string	Combined question + answer as an entailment target
proof	string	Gold proof in structured format, e.g. sent1 & sent3 -> int1: ...; int1 & sent2 -> hypothesis;
full_text_proof	string	Proof rendered as natural-language entailment steps
depth_of_proof	int	Maximum depth of the entailment tree
length_of_proof	int	Number of entailment steps
meta	object	Metadata (see below)

task3

Same as above, except proof is empty and full_text_proof, depth_of_proof, and length_of_proof are absent.

meta (tasks 1–3)

Key	Description
triples	Map from sentX IDs to fact text
intermediate_conclusions	Map from intX IDs to intermediate entailments (task 1/2)
distractors	List of distractor sentX IDs (task 2)
distractors_relevance	Retrieval scores for distractors (task 2)
worldtree_provenance	Mapping from sentences to WorldTree UUIDs and relevance scores
core_concepts	Key concepts for the example
add_list / delete_list	Facts added to or removed from the WorldTree pool
lisp_proof / polish_proof / step_proof	Alternative proof representations

worldtree_corpus

Field	Type	Description
uuid	string	Fact ID (WorldTree UUID or ET_addon:*)
text	string	Canonical fact text
source	string	worldtree or et_addon

Data Splits

Config	Train	Validation	Test
task1	1,313	187	340
task2	1,313	187	340
task3	1,313	187	340

Validation splits correspond to the original dev.jsonl files.

Source Data

• Original release: allenai.org/data/entailmentbank
• Evaluation code: github.com/allenai/entailment_bank
• Annotation tool: github.com/cognitiveailab/entailmentbank-tree-annotation-tool
• Readable overview: EntailmentBank book (May 2022)

Example

Task 1 (task1/train.jsonl):

{
  "id": "MDSA_2009_4_30",
  "question": "Stars are organized into patterns called constellations. One constellation is named Leo. Which statement best explains why Leo appears in different areas of the sky throughout the year?",
  "answer": "Earth revolves around the sun.",
  "hypothesis": "the earth revolving around the sun causes leo to appear in different areas in the sky at different times of year",
  "context": "sent1: leo is a kind of constellation sent2: the earth revolving around the sun causes stars to appear in different areas in the sky at different times of year sent3: a constellation contains stars",
  "proof": "sent1 & sent3 -> int1: leo is a constellation containing stars; int1 & sent2 -> hypothesis; ",
  "depth_of_proof": 2,
  "length_of_proof": 2,
  "meta": { "...": "..." }
}

WorldTree corpus (worldtree_corpus.jsonl):

{"uuid": "db62-4c2e-6a9d-8b12", "text": "a constellation contains stars", "source": "worldtree"}

Licensing

This dataset follows the license of the original EntailmentBank release. See allenai.org/data/entailmentbank for details.