CroPS: Improving Dense Retrieval with Cross-Perspective Positive Samples in Short-Video Search

TL;DR

CroPS tackles filter bubbles in industrial dense retrieval by enriching positive training signals with three cross-perspective sources: query-level reformulation positives, system-level recommendations, and world-knowledge generated by LLMs. It introduces a Hierarchical Label Assignment to provide graded supervision and an H-InfoNCE loss to leverage multi-level signals effectively, achieving strong offline improvements and real-world gains in online deployments. Evaluations on CPSQA show substantial recall and ranking benefits, and live A/B tests in Kuaishou demonstrate improved recall, engagement, and reduced query reformulation, with CroPS now deployed for hundreds of millions of users. The approach is architecture-agnostic and scalable, offering a practical path to mitigate information cocoon effects in large-scale short-video search systems.

Abstract

Dense retrieval has become a foundational paradigm in modern search systems, especially on short-video platforms. However, most industrial systems adopt a self-reinforcing training pipeline that relies on historically exposed user interactions for supervision. This paradigm inevitably leads to a filter bubble effect, where potentially relevant but previously unseen content is excluded from the training signal, biasing the model toward narrow and conservative retrieval. In this paper, we present CroPS (Cross-Perspective Positive Samples), a novel retrieval data engine designed to alleviate this problem by introducing diverse and semantically meaningful positive examples from multiple perspectives. CroPS enhances training with positive signals derived from user query reformulation behavior (query-level), engagement data in recommendation streams (system-level), and world knowledge synthesized by large language models (knowledge-level). To effectively utilize these heterogeneous signals, we introduce a Hierarchical Label Assignment (HLA) strategy and a corresponding H-InfoNCE loss that together enable fine-grained, relevance-aware optimization. Extensive experiments conducted on Kuaishou Search, a large-scale commercial short-video search platform, demonstrate that CroPS significantly outperforms strong baselines both offline and in live A/B tests, achieving superior retrieval performance and reducing query reformulation rates. CroPS is now fully deployed in Kuaishou Search, serving hundreds of millions of users daily.

Figures (8)

• Figure 1: Overview of CroPS, where "+" ("-") represents positives (negatives). It introduces positives from three complementary sources: Query-level Augmentation, System-level Expansion, and World Knowledge Enrichment.
• Figure 2: The overall framework of the CroPS, where "+" ("-") represents positives (negatives). It consists of three components: CroPS Data Engine, Hierarchical Label Assignment (HLA), and H-InfoNCE Optimization. CroPS Data Engine expands the positive samples from: (1) Query-level Augmentation, (2) System-level Expansion, and (3) World Knowledge Enrichment.
• Figure 3: Case Study. Blue block indicates test results, while green block shows evidence traces from our method.
• Figure 4: The algorithmic process of H-InfoNCE. As an example, we use training data consisting of 2 queries ($q_1$ and $q_2$ with their corresponding documents: $d_1,d_2,d_3$ for $q_1$, and $d_4,d_5$ for $q_2$, respectively) for a single forward pass. It can primarily be divided into three steps: Compute Similarity Scores, Hierarchy-Aware Masking and Contrastive Calculation.
• Figure 5: LLM Prompt in World Knowledge Enrichment.