WikiAsp: A Dataset for Multi-domain Aspect-based Summarization

This repository contains the dataset from the paper "WikiAsp: A Dataset for Multi-domain Aspect-based Summarization".

WikiAsp is a multi-domain, aspect-based summarization dataset in the encyclopedic domain. In this task, models are asked to summarize cited reference documents of a Wikipedia article into aspect-based summaries. Each of the 20 domains include 10 domain-specific pre-defined aspects.

Dataset

Download

WikiAsp is a available via 20 zipped archives, each of which corresponds to a domain. More than 28GB of storage space is necessary to download and store all the domains (unzipped). The following command will download all of them and extract archives:

./scripts/download_and_extract_all.sh /path/to/save_directory

Alternatively, one can individually download an archive for each domain from the table below. (Note: left-clicking will not prompt downloading dialogue. Open the link in a new tab, or save from the context menu on your OS, or use wget.)

Format

Each domain includes three files {train,valid,test}.jsonl, and each line represents one instance in JSON format. Each instance forms the following structure:

{
    "exid": "train-1-1",
    "input": [  
        "tokenized and uncased sentence_1 from document_1",
        "tokenized and uncased sentence_2 from document_1",
        "...",
        "tokenized and uncased sentence_i from document_j",
        "..."
    ],
    "targets": [ 
        ["a_1", "tokenized and uncased aspect-based summary for a_1"],
        ["a_2", "tokenized and uncased aspect-based summary for a_2"],
        "..."
    ]
}

where,

• exid: str
• input: List[str]
• targets: List[Tuple[str,str]]

Here, input is the cited references and consists of tokenized sentences (with NLTK). The targets key points to a list of aspect-based summaries, where each element is a pair of a) the target aspect and b) the aspect-based summary.

Inheriting from the base corpus, this dataset exhibits the following characteristics:

• Cited references are composed of multiple documents, but the document boundaries are lost, thus expressed simply in terms of list of sentences.
• Sentences in the cited references (input) are tokenized using NLTK.
• The number of target summaries for each instance varies.

Citation

If you use the dataset, please consider citing with

@article{hayashi20tacl,
    title = {WikiAsp: A Dataset for Multi-domain Aspect-based Summarization},
    author = {Hiroaki Hayashi and Prashant Budania and Peng Wang and Chris Ackerson and Raj Neervannan and Graham Neubig},
    journal = {Transactions of the Association for Computational Linguistics (TACL)},
    month = {},
    url = {https://arxiv.org/abs/2011.07832},
    year = {2020}
}

LICENSE

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.