Personalized Title Generation and Automatic Cover Image Synthesis for Information‑Flow Scenarios

In information‑flow scenarios such as QQ Browser, titles and cover images are critical cues that influence whether users click to view detailed content, making them essential for creators, platforms, and advertisers.

Personalized Title Generation aims to produce titles tailored to individual users or specific contexts. Three main use cases are described: recommendation systems, search engines, and creator platforms. Challenges include representing user interests, queries, or author style, and designing interaction mechanisms between scene representations and the article.

Three solution families are introduced:

Keyword‑based title generation: integrates keywords (e.g., tags, interests, queries) into a Transformer model by adding a dedicated keyword representation layer or by using the keywords as queries in a multi‑head attention module, resulting in titles that better match user interests.

Click‑sequence‑based title generation: combines a Transformer encoder for article content with an LSTM decoder guided by a user embedding derived from historical click sequences. The user embedding can initialize the LSTM hidden state, participate in attention distribution, or be incorporated into the gating network, producing titles aligned with the user’s past behavior.

Author‑style‑based title generation: incorporates an author’s historical titles into a Transformer model using contrastive learning to capture stylistic features, leading to titles that preserve the author’s brand while improving Rouge and BLEU scores.

Automatic Cover Image Synthesis addresses the need for visually appealing yet informative cover images. The pipeline includes:

Image restoration to remove watermarks, subtitles, or logos using techniques such as Faster R‑CNN and inpainting.

Seq2Seq (Pointer‑augmented T5) extraction of key information from titles and tags.

Object detection (e.g., Faster R‑CNN) to locate faces, objects, etc., ensuring they are not occluded.

Layout Generation to automatically place extracted textual information onto the cleaned image, producing a balanced and readable cover.

Experimental results show that both personalized title models and the cover‑image synthesis pipeline significantly improve click‑through rates and user satisfaction.

The presentation concludes with acknowledgments and a thank‑you to the audience.