A Puzzle-Based Dataset for Natural Language Inference
Roxana Szomiu, Adrian Groza
TL;DR
The paper introduces PuzzTE, a puzzle-based dataset for textual entailment and question answering derived from natural-language logic puzzles in three domains: comparison, knights-and-knaves, and zebra puzzles. PuzzTE automatically translates puzzles into First-Order Logic, generates the full set of atomic questions, and uses Mace4/Prover9 to assign labels of entailment, contradiction, or unknown, enabling automated, reasoning-driven evaluation distinct from crowdsourced labeling. The dataset comprises 16,745 puzzle-question pairs across unambiguous and ambiguous puzzles, with detailed counts of entailments, contradictions, and unknowns, and is supported by a four-module pipeline for NL-to-FOL translation, question generation, theorem proving, and data assembly. The work highlights two puzzle properties—each piece of information being necessary and no unnecessary information—making PuzzTE a challenging resource for machine comprehension and reasoning-based QA, and outlines plans for a Kaggle competition and future automatic generation of QA data from FOL without human verification.
Abstract
We provide here a dataset for tasks related to natural language understanding and natural language inference. The dataset contains logical puzzles in natural language from three domains: comparing puzzles, knighs and knaves, and zebra puzzles. Each puzzle is associated with the entire set of atomic questions that can be generated based on the relations and individuals occurring in the text. For each question we provide the correct answer: entailment, contradiction or ambiguity. The answer's correctness is verified against theorem provers. Good puzzles have two properties: (i) each piece of information is necessary and (ii) no unnecessary information is provided. These properties make puzzles interesting candidates for machine comprehension tasks.