Personalized Title Generation and Automatic Cover Image Synthesis for Content Feeds

The presentation introduces the importance of attractive titles and cover images in information‑flow scenarios such as QQ Browser, where they heavily influence user click‑through and consumption behavior.

Personalized Title Generation

Three main application scenarios are discussed: recommendation systems, search engines, and creator platforms. The key challenges include representing scene information (user interests, queries, author style) and designing interaction mechanisms between scene representation and article content.

Three technical approaches are described:

Keyword‑based title generation: Keywords (tags, interests, queries) are integrated into a Transformer model by adding a dedicated keyword representation layer or by using the keywords as queries in a multi‑head attention module, improving relevance in recommendation and search.

Click‑sequence‑based title generation: A Transformer encoder extracts article semantics while an LSTM decoder incorporates user click history via a user embedding that can (a) initialize the LSTM hidden state, (b) participate in attention distribution, or (c) be used in the gating network, resulting in titles aligned with individual user preferences.

Author‑style‑based title generation: Historical titles of the same author are used to construct <article, historical title, target title> triples. A contrastive learning scheme treats titles from the same author as positive pairs and different authors as negatives, enabling the Transformer to capture author style beyond pure semantics.

Experimental results show significant improvements in Rouge and BLEU scores, as well as higher user engagement metrics, when incorporating keywords, click sequences, or author style.

Automatic Cover Image Synthesis

The article emphasizes the balance between simplicity and richness in cover images and recommends embedding key information (title, tags) without obstructing important visual elements.

The proposed synthesis pipeline consists of:

Image restoration: Faster R‑CNN based detection and inpainting remove watermarks, subtitles, and other distractions.

Seq2Seq key‑phrase extraction: A pointer‑augmented T5 model extracts salient phrases from titles and tags.

Object detection: CNN‑based detectors identify faces, objects, etc., to avoid occlusion during text overlay.

Layout generation: A layout‑generation method determines optimal text placement and style, merging extracted key information with the cleaned image.

Both parts conclude that leveraging advanced AI techniques for title generation and cover image synthesis can substantially enhance content attractiveness and user experience.