CommaQA: Communicating with Agents for QA

CommaQA Dataset is a QA benchmark for learning to communicate with agents. It consists of three datasets capturing three forms of multi-hop reasoning -- explicit(CommaQA-E), implicit(CommaQA-I), and numeric(CommaQA-N).

Paper Link: Semantic Scholar

Citation:

@article{Khot2021LearningTS,
  title={Learning to Solve Complex Tasks by Talking to Agents},
  author={Tushar Khot and Kyle Richardson and Daniel Khashabi and Ashish Sabharwal},
  journal={ArXiv},
  year={2021},
  volume={abs/2110.08542}
}

Dataset
- Download
- Formats
Models
Code

Dataset

Download

Download the datasets:

Formats

Each dataset contains three formats:

commaqa/: This is default CommaQA format that contains the raw facts, verbalized sentences, agent language specification, QA pairs and associated theories. Each JSON file consists of list of items with each item containing a KB and a group of questions. The JSON file uses the following key structure:
- kb: a map between predicate names and the facts associated with each predicate
- context: verbalized context used by black-box models
- per_fact_context: map between the facts in the kb and the corresponding verbalized sentence in context
- pred_lang_config: map between the agent name and the questions answerable by this agent (see ModelQuestionConfig for more details)
- qa_pairs: list of QA pairs associated with this context using the keys:
  - id: question id
  - question: question
  - answer: numeric or list answer
  - config: specific theory config used to construct this example (see TheoryConfig for more details)
  - assignment: assignment to the variables in the theory
  - decomposition: agent-based decomposition to answer this question
  - facts_used: facts needed to answer this question
drop/: This contains the dataset in the default DROP dataset format by converting the context and qa_pairs from the commaqa/ format.
seq2seq/: This contains the dataset in a simple text format (one example per line) by converting the context and qa_pairs from the commaqa/ format into <context> Q: <question> A: <answer>.

Additional Datasets

Decompositions: The training data for the decomposer can be downloaded from here. The data uses the JSONL format with train_seqs field containing the decompositions for each question. Each entry in the train_seqs array corresponds to one step in the decomposition chain. E.g.
```
QC: What awards have the actors of the Hallowcock winning movies received?
QI: (select) [table] The award Hallowcock has been awarded to which movies? A: [\"Clenestration\"]
QS: (project_values_flat_unique) [text] Who all acted in the movie #1?
```

can be used to train a model to generate the question (project_values_flat_unique) [text] Who all acted in the movie #1? (string following QS:) given the previous questions and answers.

Language: The language specification for agents in each dataset can be downloaded from here. It contains to files:
- operations.txt: File containing the list of operations for CommaQA
- model_questions.tsv: A TSV file where the first field corresponds to the model name and all the subsequent fields contain valid questions that can be asked to this model.

Models

We also provide the T5-Large models trained to produce the next sub-question based on the oracle decompositions. These models can be used to perform inference as described [here]((commaqa/inference/README.md) to reproduce the TMN results.

Code

Refer to the individual READMEs in each package for instructions on:

FranxYao/CommaQA