Link Search Menu Expand Document

The OpenCSR Data

Table of contents

The OpenCSR Datasets

We present the three datasets used for studying OpenCSR, which we got by reformatting the multiple-choice QA datasets – ARC, OBQA, and QASC.

Download the OpenCSR datasets

Please put the datasets under the folder drfact_data/datasets/[ARC|OBQA|QASC]/.

This zip file contains train/dev/test data in jsonl format (i.e., each line is a dictionary in json) for ARC, OBQA, and QASC. Note that the gold annotations (i.e.,all_answer_concepts items) for the test sets are hidden for hosting a shared task to fairly compare the submissions. Below are two examples for training/validation instances:

// Example 1 from OBQA
  "_id": "257",  
    // unique example id
  "question": "Global warming is lowering the world's amount of ( ___ ) ",  
    // input: question text
  "entities": [{"kb_id": "global warming", "name": "global warming"}, ...],  
    // supplementary info: simple concept matching over the question text. "kb_id" and "name" are the same.
  "all_answer_concepts": [{"kb_id": "ice", "name": "ice"}],
    // output: answer concepts that a model needs to retrieve. there is a duplicate key "answer_concepts"
  "all_answer_concepts_decomp": [{"kb_id": "ice", "name": "ice"}],
    // supplementary info: the decomposition of the "all_answer_concepts"
  "answer": "ice"
    // supplementary info: original text of the correct choice.

// Example 2 from ARC
  "_id": "Mercury_7267873", 
  "question": "What provide the best evidence that life could develop on Mars?", 
  "entities": [{"kb_id": "good", "name": "good"}, {"kb_id": "evidence", "name": "evidence"}, {"kb_id": "life", "name": "life"}, {"kb_id": "develop", "name": "develop"}, {"kb_id": "mars", "name": "mars"}],
  "all_answer_concepts": [
    {"kb_id": "organic molecule", "name": "organic molecule"}, 
    {"kb_id": "ice", "name": "ice"}
  "all_answer_concepts_decomp": [{"kb_id": "organic molecule", "name": "organic molecule"}, {"kb_id": "ice", "name": "ice"}, {"kb_id": "organic", "name": "organic"}, {"kb_id": "molecule", "name": "molecule"}],
  "answer": "its ice and organic molecules" // original text of the correct choice.

Here is an example of the test data (the answers are hidden for the shared task):

  "_id": "Mercury_SC_400328", 
    // unique example id
  "question": "What item is used for protection from chemical splashing?", 
  "entities": ["chemical splashing", "item", "use", "protection", "chemical"], 
    // supplementary info: simple concept matching over the question text.
  "all_answer_concepts": [ // hidden info; used for evaluaiton. 
    {"kb_id": "safety goggle", "name": "safety goggle"},
    {"kb_id": "lab coat", "name": "lab coat"},
    {"kb_id": "rubber boot", "name": "rubber boot"},
    {"kb_id": "protective glove", "name": "protective glove"},
    {"kb_id": "face shield", "name": "face shield"},
    {"kb_id": "helmet", "name": "helmet"},
    {"kb_id": "shoe", "name": "shoe"},
    {"kb_id": "glass", "name": "glass"},
    {"kb_id": "goggle", "name": "goggle"},
    {"kb_id": "glove", "name": "glove"},
    {"kb_id": "coverall", "name": "coverall"}
  ],  // Note that the first one is from the original choice, and others are from our crowdsourcing.
  "answer": "safety goggle",  // original text of the correct choice.
    // hidden info.

Please also cite the papers of these original datasets if you use the data here.

  title={Think you have solved question answering? try arc, the ai2 reasoning challenge},
  author={Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind},
  journal={arXiv preprint arXiv:1803.05457},

    title = "Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering",
    author = "Mihaylov, Todor  and Clark, Peter  and Khot, Tushar  and Sabharwal, Ashish",
    booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
    year = "2018"

  author    = {Tushar Khot and  Peter Clark and Michal Guerquin and Peter Jansen and Ashish Sabharwal},
  title     = {QASC: A Dataset for Question Answering via Sentence Composition},
  booktitle = {The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI
               2020, New York, NY, USA, February 7-12, 2020},
  year      = {2020}

The Commonsense Knowledge Corpus

Download the Preprocessed Corpus

We use the GenericsKB as our knowledge corpus. We preprocess the corpus, extract the frequent concepts, and finally link the facts to their mentioned concepts. The link contains two files:

  • gkb_best.drfact_format.jsonl consists of generics commonsense facts, each of which is a statement of common knowledge. (See an example below.)
  • gkb_best.vocab.txt is the vocabulary of the concepts (i.e., noun chunks) sorted by their frequency.

Here is an example json line for the commonsense fact: “Trees remove dust and pollution from the air.

    "id": "gkb-best#934338", "url": "gkb-best#934338",
    "context": "Trees remove dust and pollution from the air .",
    "mentions": [
        { "kb_id": "tree", "start": 0, "text": "Trees", "sent_id": 0 },
        { "kb_id": "dust", "start": 13, "text": "dust", "sent_id": 0 },
        { "kb_id": "pollution", "start": 22, "text": "pollution", "sent_id": 0 },
        { "kb_id": "air", "start": 41, "text": "air", "sent_id": 0 }
    "title": "Trees", "kb_id": "tree"

Please also cite the GenericsKB paper if you use the data here.

  title={GenericsKB: A Knowledge Base of Generic Statements},
  author={Sumithra Bhakthavatsalam and Chloe Anastasiades and P. Clark},