NewsQuote Dataset: Quote-Speaker Pairs from New Articles

To seek reliable information sources for news events, we introduce a novel task of expert recommendation, which aims to identify trustworthy sources based on their previously quoted statements. To achieve this, we built a novel dataset, called NewsQuote, consisting of 23,571 quote-speaker pairs sourced from a collection of news articles.

We built our NewsQuote dataset from the AYLIEN coronavirus dataset, which contains news articales published between November 2019 and August 2020. Apart from text, each article is also accompanied with the meta data such as authors, keywords, summary, source, publishing time, topical categories coded by both the Interactive Advertising Bureau (IAB) taxonomy and the IPTC NewsCodes, as well as the recognized entities and entity links from the DBpedia.

Our data covers three categories of quotes: direct quote, indirect quote and mixed quote. We roughly estimated the weight of each quotation type on the dataset by the number and position of quotation marks: 81% for indirect quotes, 11% for direct quotes, and 7% for mixed quotes. In the test set, there are 1,582 (79%) indirect quotes, 178 (9%) mixed quotes, and 240 direct quotes (12%). The following table shows the statistics of our final NewsQuote dataset. In summary, we have a total of 23,571 English speaker-quote pairs with 2,843 speakers from 263 global domains.

For each sample in the dataset:

'ID': sample id. In the id, number before '-' is the corresponding article id in the AYLIEN coronavirus dataset

'Sentence': main sentence.

'Tags': semantic role labels.

'Left_sentence': context before the sentence.

'Right_sentence': context after the sentence.

'Speaker': speaker.

'Words': words in the sentence.

'Verb': trigger verb that indicates the quotation.

'Quotation1': possible quotation.

'Time': publish time.

'Entity_link': speaker's Dbpedia link.