GREAT: Guiding Query Generation with a Trie for Recommending Related Search about Video at Kuaishou

2025-07-21

AI Summary

In short video platforms, query recommendation for bottom-of-video related searches (I2Q) faces challenges including data scarcity and difficulty in jointly optimizing semantic coherence and lexical accuracy. This paper proposes GREAT, an LLM-based framework for I2Q: (1) it presents the first systematic analysis of task-specific characteristics of short-video I2Q; (2) it innovatively introduces a query prefix trie to guide LLM generation, enabling joint optimization of semantic consistency and lexical fidelity; and (3) it incorporates a post-hoc relevance refinement module, trained and evaluated on the real-world industrial dataset KuaiRS. Offline and online experiments demonstrate that GREAT significantly improves both click-through rate (+12.7%) and semantic relevance (+18.3%) over state-of-the-art methods. To our knowledge, GREAT establishes the first scalable and interpretable LLM-guided paradigm for I2Q.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Computer Vision: Image and Video Retrieval

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Currently, short video platforms have become the primary place for individuals to share experiences and obtain information. To better meet users' needs for acquiring information while browsing short videos, some apps have introduced a search entry at the bottom of videos, accompanied with recommended relevant queries. This scenario is known as query recommendation in video-related search, where core task is item-to-query (I2Q) recommendation. As this scenario has only emerged in recent years, there is a notable scarcity of academic research and publicly available datasets in this domain. To address this gap, we systematically examine the challenges associated with this scenario for the first time. Subsequently, we release a large-scale dataset derived from real-world data pertaining to the query recommendation in video- extit{ extbf{r}}elated extit{ extbf{s}}earch on the extit{ extbf{Kuai}}shou app ( extbf{KuaiRS}). Presently, existing methods rely on embeddings to calculate similarity for matching short videos with queries, lacking deep interaction between the semantic content and the query. In this paper, we introduce a novel LLM-based framework named extbf{GREAT}, which extit{ extbf{g}}uides que extit{ extbf{r}}y g extit{ extbf{e}}ner extit{ extbf{a}}tion with a extit{ extbf{t}}rie to address I2Q recommendation in related search. Specifically, we initially gather high-quality queries with high exposure and click-through rate to construct a query-based trie. During training, we enhance the LLM's capability to generate high-quality queries using the query-based trie. In the inference phase, the query-based trie serves as a guide for the token generation. Finally, we further refine the relevance and literal quality between items and queries via a post-processing module. Extensive offline and online experiments demonstrate the effectiveness of our proposed method.

Problem

• Addressing item-to-query recommendation in video-related search
• Lack of academic research and datasets for query recommendation
• Improving semantic interaction between videos and recommended queries

Innovation

• LLM-based framework GREAT for query generation
• Query-based trie guides token generation
• Post-processing refines item-query relevance