Chinese Title Generation for Short Videos: Dataset, Metric and Algorithm

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Existing video captioning methods lack human interest and attractiveness.
• Video title generation (titling) lacks standardized benchmarks for evaluating attractive titles.
• There is a need for comprehensive resources to advance Chinese short video analysis.

Purpose of the Study:

• To introduce CREATE, the first large-scale Chinese dataset for video retrieval and title generation.
• To propose ACTEr, a novel metric for evaluating the attractiveness and relevance of generated video titles.
• To present ALWIG, a multi-modal model for simultaneous video titling, captioning, and retrieval.

Main Methods:

• Developed CREATE dataset with 210K labeled, 3M/10M pre-training data, covering 51 categories and extensive annotations.
• Introduced ACTEr (Attractiveness-Consensus-based Title Evaluation) metric using semantic correlation and consensus weights.
• Proposed ALWIG (multi-modal ALignment WIth Generation) model with tag-driven alignment and GPT-based generation.

Main Results:

• The CREATE dataset provides a robust foundation for Chinese short video research.
• ACTEr objectively evaluates video title quality by considering attractiveness and relevance.
• The ALWIG model demonstrates strong performance as a baseline for multi-modal video tasks.

Conclusions:

• The release of CREATE, ACTEr, and ALWIG is expected to stimulate further research in Chinese short video analysis and creation.
• This work bridges the gap between objective video description and engaging title generation.
• The developed resources facilitate advancements in video titling, captioning, and retrieval applications.