Abstract
In this work, we define the task of teaser generation and provide an evaluation benchmark and baseline systems for the process of generating teasers. A teaser is a short reading suggestion for an article that is illustrative and includes curiosity-arousing elements to entice potential readers to read particular news items. Teasers are one of the main vehicles for transmitting news to social media users. We compile a novel dataset of teasers by systematically accumulating tweets and selecting those that conform to the teaser definition. We have compared a number of neural abstractive architectures on the task of teaser generation and the overall best performing system is See et al.(2017)'s seq2seq with pointer network.
Introduction: The Rise of the Teaser
In the digital news era, publishers face a constant struggle to capture reader attention without resorting to deceptive "clickbait." While traditional headlines prioritize relevance and factual summary, clickbait often exploits information gaps in a non-bona-fide manner. This paper identifies a middle ground: the teaser.
A teaser is designed to be both bona-fide (accurately reflecting the story) and "teasing" (arousing curiosity). Crucially, teasers are highly abstractive; they are rarely direct substrings of the article. Generating them requires a model to understand the core narrative and rephrase it to create a psychological "hook" that motivates a click.
The Teaser Recognition Pipeline
To build a robust dataset for this task, the authors developed a multi-stage unsupervised algorithm to filter 1.4 million tweets. The pipeline employs three primary filters:
- Extractivity Check: Filters out tweets that are simple substrings of the article.
- Abstractivity Check: Uses unigram overlap scores to ensure the tweet maintains a specific balance of similarity (between 20% and 80%) with the article's most prominent sections.
- Teasingness Check: Employs a Domain Relevance ($dr$) metric to identify "unusual" words—terms that are rare in the specific news domain but common elsewhere—which often signal a stylistic choice intended to pique interest.
Dataset Characteristics and Abstractivity
The resulting Teaser Corpus consists of 330,000 cleaned pairs from major English-language outlets. Quantitative analysis confirms that teasers are significantly more abstractive than headlines. While headlines typically derive ~80% of their content directly from the source, teasers share only ~32%. This disparity highlights the creative challenge of teaser generation: models cannot simply rely on extraction but must synthesize new, engaging phrasing.
Neural Generation and Performance
The authors evaluated two primary abstractive architectures: a standard Attentive Seq2Seq model and a Pointer-Generator Network. The Pointer-Generator proved to be the superior baseline, as its ability to "point" to (copy) specific words from the source text allows it to effectively handle the "unusual" words that are key to a teaser's effectiveness.
However, the task remains difficult. The Pointer-Generator achieved a ROUGE-L of 18.49 on teasers, compared to 21.68 for headlines and 31.53 for highlights. This lower score reflects the inherent creativity of teasers; because they are so detached from the source text, standard overlap-based metrics like ROUGE struggle to fully capture their quality. Human evaluations supported this, noting that while models can produce "stimulating" content, human editors still excel at crafting the most engaging social media hooks.
Conclusion and Future Directions
This research formalizes the teaser as a distinct NLP target and provides a large-scale benchmark for future study. The authors suggest that future work should move beyond simple word-overlap metrics to better model the "information gap" and explore multi-task learning to capture the stylistic nuances between headlines, highlights, and teasers.
Summary updated