Source-linked AI summary
News Category Dataset
Rishabh Misra
TL;DR
News coverage trends can provide insight into what society deems important at a given time. The paper collects and processes HuffPost news headlines with category and publication metadata into a News Category Dataset. The dataset analysis reports variation in short-description lengths across categories, different category emphases in 2013 and 2017, and visibly distinct headline vocabulary across categories.
Problem
News coverage trends can provide insight into what society deems important at a given time.
Source: Introduction · p. 1Method
The paper collects and processes HuffPost news headlines with category and publication metadata into a News Category Dataset.
Source: Introduction · p. 1, Data Source Discovery · p. 1Results
The dataset analysis reports variation in short-description lengths across categories, different category emphases in 2013 and 2017, and visibly distinct headline vocabulary across categories.
Source: Exploratory Data Analysis · pp. 2–3Takeaways & Limitations
The dataset is presented as a benchmark for NLP tasks involving the syntax and semantics of general-purpose news or news from specific categories.
Source: Concluding Remarks · p. 4HuffPost's archive is much thinner after May 2018, with about 200,000 headlines from 2012 to May 2018 and about 10,000 from May 2018 to September 2022; the authors recommend 2012–2017 for temporal analysis.
Source: Exploratory Data Analysis · p. 2
Abstract
from arXiv · showhide
People rely on news to know what is happening around the world and inform their daily lives. In today's world, when the proliferation of fake news is rampant, having a large-scale and high-quality source of authentic news articles with the published category information is valuable to learning authentic news' Natural Language syntax and semantics. As part of this work, we present a News Category Dataset that contains around 210k news headlines from the year 2012 to 2022 obtained from HuffPost, along with useful metadata to enable various NLP tasks. In this paper, we also produce some novel insights from the dataset and describe various existing and potential applications of our dataset.
Introduction
News coverage both informs public understanding and reflects what society considers important, while also contributing to policy debate. The paper introduces a dataset with category and date metadata to support analysis of journalistic trends and related NLP work.
Introduction
- News coverage can inform public decisions and reveal what society considers important, while helping raise awareness, initiate dialogue, and set policy agendas.The paper illustrates this role with Indian media coverage of food-price increases after the 2008 global food crisis.Source: Introduction · p. 1
- The News Category Dataset is introduced to help researchers discover journalistic trends using headline category and publication-date metadata.The metadata is presented as enabling analyses that extract varied insights and support NLP tasks such as semantic tagging, named entity recognition, and word sense disambiguation.Source: Introduction · p. 1
- The paper reports collecting and processing the dataset, presenting insights from it, describing its community adoption, and outlining potential applications.Source: Introduction · p. 1
Dataset
The dataset contains 210,294 HuffPost headlines published from 2012 to 2022, with fine-grained category information and article text. Its construction uses source-selection criteria, archive collection, and a category-count threshold to limit data skew.
Data Source Discovery
- Source selection required an archive, sufficient article volume, category and date metadata, and cleanly collectible text; HuffPost met these criteria with an archive dating to 2012.Its archive also supplied article authors and short descriptions.Source: Data Source Discovery · pp. 1–2
Dataset
- 210,294 news headlines published from 2012 to 2022 form the dataset, whose distinctive contribution is fine-grained category information alongside high-quality text.Source: Data Source Discovery · p. 2
Data Source Discovery
- Each record includes its category, headline, authors, article link, short description, and publication date.Source: Data Source Discovery · p. 2
- Tables 1 and 2 identify the top and bottom ten categories by article count.Source: Data Source Discovery · p. 2, Table 1 · p. 2, Table 2 · p. 2
Data Curation Method
- The collection process extracts article links, categories, headlines, abstracts, authors, and publication dates from HuffPost archive pages, then uses Selenium to advance through pages.BeautifulSoup is used to extract the article information.Source: Data Curation Method · p. 2
- Categories with fewer than 1,000 articles were removed to reduce data skew, and the headline text received no additional preprocessing.The authors describe the professional news headlines as reasonably high quality, without misspellings or shorthand.Source: Data Curation Method · p. 2
Exploratory Data Analysis
The dataset is temporally uneven, with far fewer articles after 2018. Its exploratory analyses compare category-level headline and short-description lengths, yearly category distributions, and vocabulary across categories.
Exploratory Data Analysis
- About 200k headlines cover 2012 to May 2018, versus about 10k from May 2018 to September 2022; the authors recommend 2012–2017 for temporal analysis.HuffPost stopped maintaining an extensive archive sometime after the dataset’s initial 2018 release.Source: Exploratory Data Analysis · p. 2, Figure 1 · p. 2
- Short-description lengths vary more across categories than headline lengths do.Figures 2 and 3 show category-level average headline and short-description word counts, respectively.Source: Exploratory Data Analysis · p. 2, Figure 2 · p. 3, Figure 3 · p. 3
- In 2013, reporting focused more on Wellness, parenting, and beauty; by 2017, it focused more on Politics, World News, and Entertainment.Source: Exploratory Data Analysis · p. 3
- A word cloud of headlines shows clear distinctions in words used across news categories.The authors suggest these distinctions may be useful for studying linguistic phenomena.Source: Exploratory Data Analysis · p. 3
Applications
The paper reviews research uses of the dataset and outlines further applications. Examples span sarcasm, epidemiological text classification, gender representation, media language, and potential NLP tasks.
Applications
- Since its 2018 release, the dataset has seen wide research-community adoption, with the section surveying applications that may inspire further use.Source: Applications · p. 3
Sarcasm Detection
- Misra and colleagues combined randomly sampled dataset headlines with The Onion’s headlines to build a sarcasm dataset, which they report has less label and language noise than Twitter-based datasets and more labels than hand-labeled ones.They used a hybrid neural network with attention to investigate what constitutes sarcasm.Source: Sarcasm Detection · p. 3
Epidemiological Text Classification
- Mutuvi and colleagues filtered out articles mentioning disease outbreaks and combined the remainder with epidemiology news to create a dataset for epidemiological text classification.They explored deep-learning architectures for the task.Source: Epidemiological Text Classification · p. 3
Gender Representation in News
- Dacon and Liu used the dataset to examine gender bias in English news abstracts, finding socially constructed biases through measurements of distribution, content, labeling, and word choice.Source: Gender Representation in News · p. 3
The Media Language of Police Killings
- A BERT-based study used the dataset’s Crime category to identify police-killing stories and found more responsibility-obscuring language structures in coverage of police killings than civilian homicides.The study highlights semantic structure alongside coverage amount and slant when analyzing media influence on perceptions.Source: The Media Language of Police Killings · p. 4
Other Applications
- The authors propose using the corpus as a benchmark for NLP tasks involving syntax and semantics in general-purpose or category-specific news.Potential applications include tagging untracked articles, tracking shifts in journalism’s focus, comparing category writing styles, and tagging tone.Source: Other Applications · p. 4
Concluding Remarks
The paper presents a large-scale, high-quality News Category Dataset as a benchmark for NLP tasks involving news syntax and semantics, and reports insights and applications drawn from it.
Concluding Remarks
- The News Category Dataset is presented as a large-scale, high-quality benchmark for NLP tasks involving general-purpose or category-specific news syntax and semantics.The paper also reports novel dataset insights and describes existing and potential applications.Source: Concluding Remarks · p. 4