[Paper Review] The Style-Content Duality of Attractiveness: Learning to Write Eye-Catching Headlines via Disentanglement

Mingzhe Li, Xiuying Chen+4 more|arXiv (Cornell University)|Dec 14, 2020

Video Analysis and Summarization45 references4 citations

TL;DR

This paper proposes DAHG, a disentanglement-based model that separates attractive content and style from prototype headlines to generate eye-catching, content-accurate headlines. By disentangling style and content in latent space and using them to guide headline generation, DAHG achieves state-of-the-art performance, triggering 22% more clicks than baselines on the Kuaibao dataset.

ABSTRACT

Eye-catching headlines function as the first device to trigger more clicks, bringing reciprocal effect between producers and viewers. Producers can obtain more traffic and profits, and readers can have access to outstanding articles. When generating attractive headlines, it is important to not only capture the attractive content but also follow an eye-catching written style. In this paper, we propose a Disentanglement-based Attractive Headline Generator (DAHG) that generates headline which captures the attractive content following the attractive style. Concretely, we first devise a disentanglement module to divide the style and content of an attractive prototype headline into latent spaces, with two auxiliary constraints to ensure the two spaces are indeed disentangled. The latent content information is then used to further polish the document representation and help capture the salient part. Finally, the generator takes the polished document as input to generate headline under the guidance of the attractive style. Extensive experiments on the public Kuaibao dataset show that DAHG achieves state-of-the-art performance. Human evaluation also demonstrates that DAHG triggers 22% more clicks than existing models.

Motivation & Objective

• To address the gap in headline generation models that fail to jointly optimize for both content relevance and stylistic attractiveness.
• To model attractiveness as a duality of content (what is being said) and style (how it is said), rather than treating it as a monolithic attribute.
• To improve headline generation by leveraging a prototype headline’s disentangled content and style representations to guide the generation process.
• To ensure generated headlines are both engaging and faithful to the source document, reducing clickbait risks.

Proposed method

• A disentanglement module separates the attractive content and style of a prototype headline into distinct latent spaces using two auxiliary constraints to enforce disentanglement.
• The learned content space is used to iteratively highlight salient, attractive parts in the input document representation.
• The headline generator produces a headline conditioned on the polished document representation and the disentangled attractive style embedding.
• The model is trained end-to-end on the Kuaibao dataset using ROUGE and BLEU metrics, with human evaluation for click-through rate.
• Prototype selection is based on semantic similarity between input document and prototype headlines to ensure relevant style and content transfer.
• The disentanglement is enforced via auxiliary constraints that encourage disentanglement in the latent space, ensuring style and content are independent.

Experimental results

Research questions

• RQ1 Can disentangling content and style in headlines improve both factual faithfulness and attractiveness in generated headlines?
• RQ2 How does the choice of prototype headline affect the quality and relevance of generated headlines?
• RQ3 Can a disentangled style representation guide the generation of more engaging, click-worthy headlines without sacrificing content accuracy?
• RQ4 To what extent does the disentanglement mechanism improve headline quality compared to end-to-end abstractive models?

Key findings

• DAHG achieves state-of-the-art performance on the Kuaibao dataset, outperforming all abstractive and extractive baselines in ROUGE and BLEU metrics.
• Human evaluation shows that headlines generated by DAHG trigger 22% more clicks than the strongest baseline model.
• When the prototype headline is semantically similar to the input document, DAHG generates headlines that accurately reflect key events and attract attention.
• Incorrect or mismatched prototype headlines lead to poor generation, especially when the prototype emphasizes irrelevant content (e.g., gender over conflict).
• The disentanglement mechanism successfully isolates style and content, enabling the model to preserve factual content while enhancing stylistic appeal.
• Baselines like PESG and Sensation generate fluent but often unfaithful or overly sensationalized headlines, whereas DAHG maintains content consistency while improving attractiveness.