metadata

license: apache-2.0
task_categories:
  - text-generation
  - summarization
language:
  - en
size_categories:
  - 100K<n<1M

Dataset Card for PENS: PErsonalized News headlineS

PENS is an English dataset for Personalized News Headline Generation Research. It contains two parts for training and test individually. The training set was collected from anonymized user impressions logs of Microsoft News website, and the test set is manually-created by hundreds of native speakers to enable a fair testbed for evaluating models in an offline mode.

PENS contains about 113k English news articles whose topics are distributed into 15 categories and 500k impression logs generated by over 445k users for training. In detail, every news article contains rich textual content including title, body, category and corresponding entities. Each impression log contains the click events, non-clicked events and historical news click behaviors of this user before this impression. To provide an offline testbed, we invited 103 English native speakers to manually create a test set by two stages. In detail, there are over 100k personalized news headlines generated.

Notice that each user was de-linked from the production system when securely hashed into an anonymized ID to protect user privacy. For more detailed information about the PENS dataset, you can refer to the following paper: PENS: A Dataset and Generic Framework for Personalized News Headline Generation.

Dataset Details

Dataset Description

PENS (PErsonalized News headlineS) is an English dataset tailored for Personalized News Headline Generation research. The dataset is divided into training and test sets to support both model development and evaluation.

Training Set: Collected from anonymized user impression logs of the Microsoft News website, containing approximately 113k English news articles across 15 categories and 500k impression logs from over 445k users. Each news article includes a title, body, category, and associated entities. Impression logs detail click events, non-click events, and historical news click behaviors of users prior to each impression.
Test Set: Manually created by 103 native English speakers to provide a fair and reliable benchmark for offline model evaluation. This set includes over 100k personalized news headlines generated through a two-stage annotation process.

PENS ensures user privacy by de-linking user data from the production system and anonymizing user IDs through secure hashing.

Curated by: Microsoft Research
Shared by [optional]: Microsoft Research
Language(s) (NLP): English
License: Microsoft Research License Terms

Dataset Sources [optional]

Repository: PENS Dataset Repository
Paper: Link to the PENS Paper
Related Works: Available Related Works on PENS

Dataset Structure

PENS consists of two main components: the training set and the test set.

Training Set:
- News Articles: ~113,762 entries
  - Fields:
    - News ID: Unique identifier for each news article (e.g., N10000).
    - Category: One of 15 predefined categories (e.g., sports).
    - Topic: Specific topic within the category (e.g., soccer).
    - Headline: The title of the news article.
    - News body: The full content of the news article.
    - Title entity: JSON mapping between phrases in the title and their corresponding entities in WikiData.
    - Entity content: Detailed information about each entity from WikiData.
- Impression Logs: 500,000 entries
  - Fields:
    - UserID: Anonymous identifier for the user (e.g., U335175).
    - ClicknewsID: Space-separated list of historically clicked news IDs.
    - Dwelltime: Space-separated list of browsing durations for each clicked news.
    - Exposure_time: Space-separated list of exposure times for historical clicked news (separated by #TAB#).
    - Pos: Space-separated list of clicked news IDs in the current impression.
    - Neg: Space-separated list of unclicked news IDs in the current impression.
    - Start: Start time of the impression (e.g., 7/3/2019 6:43:49 AM).
    - End: End time of the impression (e.g., 7/3/2019 7:06:06 AM).
    - Dwelltime_pos: Space-separated list of browsing durations for clicked news in the current impression.
Validation Set:
- Structured similarly to the training set, used for validating model performance during training.
Test Set:
- Personalized Headlines: >100,000 entries
  - Fields:
    - userid: Unique identifier for each of the 103 test users (e.g., NT1).
    - clicknewsID: Comma-separated list of historically clicked news IDs from the first stage.
    - posnewID: Comma-separated list of exhibited news IDs in the second stage.
    - rewrite_titles: Space-separated list of manually written personalized headlines (separated by #TAB#).

The dataset is structured to facilitate the mapping between user behaviors and personalized headline generation.

Uses

Direct Use

PENS is intended for research in personalized news headline generation. Suitable use cases include:

Developing models that generate personalized news headlines based on user interaction data.
Evaluating the effectiveness of personalization algorithms in news dissemination.
Analyzing user behavior and preferences in the context of news consumption.

Out-of-Scope Use

PENS is not suitable for:

Commercial use without explicit permission from Microsoft Research.
Applications requiring real-time data updates, as the dataset is static.
Tasks unrelated to text generation or summarization, such as image processing.

Dataset Structure

PENS consists of two main components: the training set and the test set.

Training Set:
- News Articles: ~113k entries
  - Fields: title, body, category, entities
- Impression Logs: ~500k entries
  - Fields: click_events, non_click_events, historical_click_behaviors
Test Set:
- Personalized Headlines: >100k entries
  - Fields: original_article, personalized_headline, user_id

The dataset is structured to facilitate the mapping between user behaviors and personalized headline generation.

Dataset Creation

Curation Rationale

The PENS dataset was created to address the need for comprehensive resources in the domain of personalized news headline generation. By leveraging real user interaction data and manually curated headlines, PENS provides a robust foundation for developing models that can tailor news content to individual preferences.

Source Data

Data Collection and Processing

Training Data: Extracted from anonymized user impression logs of the Microsoft News website. The data includes detailed user interactions with news articles, ensuring a rich source of information for personalization.
Test Data: Generated through a two-stage manual annotation process involving 103 native English speakers. This ensures high-quality, personalized headlines for reliable evaluation.

Who are the source data producers?

Training Data Producers: Users of the Microsoft News website whose interactions were anonymized and aggregated.
Test Data Producers: 103 native English speakers recruited to generate personalized headlines.

Annotations [optional]

Annotation process

The test set annotations were created in two stages:

Initial Generation: Annotators generated personalized headlines based on provided news articles and user interaction histories.
Validation: A secondary review ensured consistency and quality of the generated headlines.

Who are the annotators?

The annotators are 103 native English speakers with expertise in journalism and content creation, ensuring that the personalized headlines are both accurate and engaging.

Personal and Sensitive Information

PENS contains anonymized user interaction data. User IDs are securely hashed to prevent re-identification, and no personal, sensitive, or private information is included in the dataset.

Bias, Risks, and Limitations

PENS may contain inherent biases present in user interaction data, such as demographic or preference-based biases. Additionally, the dataset is limited to English-language news and may not generalize to other languages or cultural contexts.

Recommendations

Users should critically assess the dataset for potential biases and ensure that models trained on PENS are evaluated for fairness and representativeness. It's recommended to complement PENS with additional data sources to mitigate identified biases.

Citation

BibTeX:

@inproceedings{pens2024,
  title={PENS: PErsonalized News headlineS for Personalized News Headline Generation},
  author={Author Names},
  booktitle={Proceedings of the Conference on Natural Language Processing},
  year={2024},
  organization={Microsoft Research}
}

APA: Ao, X., Wang, X., Luo, L., Qiao, Y., He, Q., & Xie, X. (2021, August). PENS: A dataset and generic framework for personalized news headline generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) (pp. 82-92).

Glossary

Impression Log: A record of a user's interaction with news articles, including clicks and non-clicks.
Personalized Headline: A news headline tailored to an individual user's preferences and interaction history.

More Information

Put Your Voice on Stage: Personalized Headline Generation for News Articles

Fact-Preserved Personalized News Headline Generation

Dataset Card Contact

For any inquiries or feedback regarding the PENS dataset, please contact [email protected].

Download

The PENS dataset is available for free download for research purposes under the Microsoft Research License Terms. Please ensure you have read and agree to the license terms before downloading.

Download Link

License

The PENS dataset is distributed under the Microsoft Research License Terms. Please review the license agreement before using the dataset.