Grounded_Keys2Text

This is a dataset and codebase for EMNLP'22 paper titled: "Grounded Keys-to-Text Generation: Towards Factual Open-Ended Generation"

You can download the dataset from here.

The dataset contains the following files:

• [train/test/dev].jsonl
• [train/test/dev]_docs.pt

The *.jsonl files contain the main data including the entity name, section titles, and a set of factual and topical keys. Below are information about main fields in the data:

• doc_title: entity name
• sec_title: section and subsection title(s)
• topical_keys: topical keys
• infobox_keys: factual keys (infobbox portion)
• wikidata_keys: factual keys (Wikidata portion)
• text: description to be generated (section text)

Other meta fields includes:

• infobox_keyval: infobox key-value pairs a long with their bertscore wrt to section text. The values will be used for evaluation.
• wikidata_keyval: Wikidata key-value pairs a long with their bertscore wrt to section text. The values will be used for evaluation.

The *_docs.pt files contain the corresponding grounding documents for each entity in the *.jsonl files. Additionally, we include the oracle ranking of the grounding documents based on Rouge2 scores (please refer to paper Sec. 5.1).

To load the grounding *_docs.pt files, run:

import torch
documents = torch.load('test_docs.pt')

# list fields for each entry
print(documents[0].keys())