PaperlayerView exact revision on arXiv

Source-linked AI summary

Preference Shapes Relevance: Cross-component Hierarchical Semantic Alignment for Personalized Generative Retrieval

Gaoming Zhang, Angqing Jiang, Jianchun Song, Kena Qi, Dayao Chen, Wei Lin, Defu Lian

arXiv:2608.30553v1cs.IRcs.AI

TL;DR

Existing generative retrieval uses static item-derived Semantic IDs that may not align with dynamic user intents, models user behavior incompletely, and incurs high autoregressive decoding latency. CHAP introduces hierarchical semantic alignment, personalized modeling with discrete SIDs and continuous representations, and single-pass Residual Cascading Generation. Experiments across four datasets and online A/B tests report superiority, while the authors identify limitations in dynamic personalization, identifier generality, and scaling-efficiency characterization.

  • Problem

    Existing item-content-only Semantic IDs do not explicitly align dynamic user intents with static item representations, while current generative approaches rarely model behavior sequences and face costly multi-step decoding.

  • Method

    CHAP aligns query latent representations with item quantization paths and combines discrete SIDs with dense representations in a personalized framework using single-pass Residual Cascading Generation.

  • Results

    CHAP establishes a new state-of-the-art across four diverse real-world datasets and an online A/B test, demonstrating superiority and industrial efficiency.

  • Takeaways & Limitations

    CHAP provides a personalized generative-retrieval framework that jointly addresses semantic alignment and inference-latency bottlenecks within the reported evaluation settings.

  • Takeaways & Limitations

    Freezing the item-side codebook may limit personalization in highly dynamic environments, and generalization beyond RQ-VAE-style or scaling behavior remains under investigation.

Abstract

from arXiv · show

Generative Retrieval (GR) has emerged as a promising paradigm by mapping queries directly to Semantic IDs (SIDs) with powerful representation capabilities for candidate items. However, existing SIDs derived solely from item content create a semantic gap, failing to align dynamic query intents with static item representations. Furthermore, current generative paradigms rarely model user behavior sequences and are always bottlenecked by the high inference latency of beam-search autoregressive decoding. To address these challenges, we propose $\textbf{C}$ross-component $\textbf{H}$ierarchical semantic $\textbf{A}$lignment for $\textbf{P}$ersonalized generative retrieval ($\textbf{CHAP}$), a novel personalized GR framework from a hierarchical perspective. First, we design a Hierarchical Semantic Alignment module to align query's latent space with item's quantization path and synchronize multi-granular semantics. Second, we construct a personalized GR framework that models user behavior by synergizing discrete SIDs for structural guidance and continuous representations for fine-grained semantic refinement. Notably, we introduce a Residual Cascading Generation mechanism to restrict the costly multi-step Transformer Decoder to a single-pass inference, boosting inference throughput while mitigating information loss. Extensive experiments on three public datasets, one proprietary industrial dataset, and online A/B tests demonstrate CHAP's superiority, validating the effectiveness and practical value of our approach. The code is publicly available at https://github.com/zzzgm/CHAP.

1 Introduction

Generative retrieval directly generates item identifiers, but static item-only Semantic IDs do not align well with dynamic personalized intents, while hierarchical decoding creates latency bottlenecks. CHAP addresses these issues through hierarchical semantic alignment, personalized dual-view modeling, and single-pass residual cascading generation, with results reported across four datasets and online A/B tests.

  • Motivation: Static item-only Semantic IDs leave a semantic gap with user intents, limiting generalization on complex queries and complicating low-latency deployment.Existing hierarchical decoding repeatedly executes costly Transformer Decoder components, especially Cross-Attention, during online serving.
  • Approach: CHAP aligns query representations with item quantization paths and synchronizes coarse-to-fine semantics through a Hierarchical Semantic Alignment paradigm.Its modules use cross-sample alignment, hierarchical-aware contrastive learning, and soft probability distillation.
  • Approach: CHAP models personalized queries and behavior sequences by combining discrete SIDs for structural guidance with dense vectors for fine-grained semantic refinement.The Personalized Cascading Generative Sequence Model crosscascades aligned sparse SIDs with raw dense representations.
  • Approach: Residual Cascading Generation restricts the costly multi-step Transformer Decoder to a single pass while offloading layer-wise generation to lightweight residual blocks.The mechanism is described as mitigating information loss and boosting inference throughput.
  • Evaluation: Extensive experiments cover three public datasets, one proprietary industrial dataset, and online A/B tests, demonstrating CHAP’s superiority.The reported evaluation includes both offline datasets and an industrial online test.

2 Related Works

Related work spans traditional sparse and dense retrieval, personalized retrieval, and generative retrieval. Generative approaches evolve from direct ItemID generation toward semantic IDs, multi-level correlations, and architectures combining sparse and dense representations.

  • Traditional and personalized retrieval: Sparse retrieval relies on exact term matching but suffers from vocabulary mismatch, whereas dense retrieval improves semantic matching but requires costly external indices.BM25 and TF-IDF exemplify sparse retrieval; DPR and ANCE exemplify dense retrieval with indices such as HNSW.
  • Generative retrieval: Generative retrieval internalizes corpus knowledge to generate ItemIDs directly, beginning with DSI and later incorporating semantic enhancements.SE-DSI is cited as a semantic extension of the original paradigm.
  • Generative retrieval: RQ-VAE-based Semantic IDs, tree structures, dense-retrieval connections, reinforcement learning, and ranking objectives represent major generative-retrieval optimization directions.TIGER introduced RQ-VAE SIDs, while SEATER, Tied-Atomic, Gen-RRL, and LTRGR pursue distinct optimizations.
  • Generative retrieval: Recent architectures exploit sparse-dense integration and multi-level correlations to optimize identifier generation.COBRA integrates sparse and dense vectors, while CAT-ID2 and MERGE use multi-level correlations.

3 Preliminaries

Personalized generative retrieval generates hierarchical Semantic IDs from user history and the current query. RQ-VAE constructs these IDs through coarse-to-fine residual quantization, while its training objective combines reconstruction, codebook, and commitment terms.

  • Problem Definition: Personalized generative retrieval models the probability of a hierarchical SID conditioned on a user’s interaction history and current query.The target SID is a sequence of tokens derived from a codebook, and inference selects the path with highest joint probability.
  • Semantic ID Generation via RQ-VAE: RQ-VAE recursively quantizes item embeddings in a coarse-to-fine manner to provide the foundational representation for CHAP.Figure 1 schematically depicts SID generation via RQ-VAE.
  • Semantic ID Generation via RQ-VAE: A language encoder and neural encoder map item text to a latent representation, which residual codebook layers quantize by selecting the nearest code word at each step.The process begins with r0 = z and uses L residual layers, each containing K code words.
  • Semantic ID Generation via RQ-VAE: After L quantization iterations, the selected tokens form a coarse-to-fine SID and their codebook vectors approximate the latent representation.The quantized representation is formed from the sum of selected codebook vectors.
  • Semantic ID Generation via RQ-VAE: RQ-VAE training reconstructs the original item embedding while separately optimizing codebook vectors and constraining encoder outputs through commitment loss.Stop-gradient operations control which terms update codebook vectors or the encoder, with β balancing commitment loss.

4 Methodology

CHAP aligns query representations with hierarchical item quantization, then combines discrete semantic structure with dense representations for personalized retrieval. Its cascading generation design separates structural guidance from expensive decoding to support coarse-to-fine retrieval with lower inference cost.

  • Hierarchical Semantic Alignment: CHAP freezes the item-side RQ-VAE codebook and aligns query representations with the stable item quantization manifold.This addresses the query–item semantic gap while preventing semantic drift and catastrophic forgetting.
  • Hierarchical Semantic Alignment: Cross-Sample Alignment uses cross-commitment and cross-reconstruction losses to bring query latents into the target item’s quantization space.The total CSA loss combines L_CC and L_CR.
  • Hierarchical Semantic Alignment: Hierarchical-Aware Contrastive Learning aligns query and target partial quantized representations across successive SID depths.The contrastive formulation uses batches containing positive targets and negatives, with temperature τ.
  • Hierarchical Semantic Alignment: Soft Probability Distillation transfers the target’s soft code-assignment distribution to the query encoder through KL divergence.This regularizes latent alignment and reduces drastic shifts caused by hard quantization assignments.
  • Personalized Sequence Modeling: Dual-view sequence modeling represents each interaction with an aligned discrete SID and a continuous dense vector.The decoder is initialized with the query’s dual representation, encouraging cross-attention to retrieve personalized results.
  • Personalized Sequence Modeling: Hybrid optimization jointly trains hierarchical SID classification and InfoNCE-based dense contrastive learning.The joint objective combines sparse structural constraints with dense ranking optimization.
  • Residual Cascading Generation and Retrieval: Residual cascading generation runs the Transformer Decoder once and updates level-specific states through lightweight residual blocks.The design uses a global sparse structural anchor while coordinating coarse-to-fine structural correctness with fine-grained semantic precision.
  • Residual Cascading Generation and Retrieval: Inference samples M candidate SIDs in parallel and fuses hierarchical path log-probability with dense cosine similarity for final scoring.Both signals are normalized over the candidate pool before fusion.

5 Experiments

CHAP is evaluated across four real-world datasets, diverse retrieval baselines, ablations, intent segments, efficiency tests, and an online A/B test. Results report consistent state-of-the-art performance, semantic alignment benefits, component-specific strengths, improved decoding efficiency, and production gains.

  • Overall Results: CHAP consistently establishes a new state-of-the-art across four datasets, outperforming strong baselines including COBRA and MERGE.The comparison spans sparse, dense, personalized, and generative retrieval methods.
  • Ablation Study: Removing Hierarchical Semantic Alignment significantly degrades performance, while removing dense refinement also reduces accuracy.The ablation supports aligned SIDs for semantic-gap reduction and dense vectors for fine-grained semantics.
  • Generated SIDs: CHAP tightly clusters relevant items around queries and pushes irrelevant items away, unlike vanilla RQ-VAE’s loosely scattered relevant items.The PCA visualization provides qualitative evidence that alignment brings dynamic intents and static item semantics into closer correspondence.
  • Intent-wise Query Analysis: Dual-view sequence modeling gives the largest boost on Ambiguous queries, HSA helps ColdStart and LongTail segments, and Residual Generation improves all slices.These components exhibit complementary strengths: history disambiguates generic queries, alignment supports sparse segments, and residual decoding supplies structural uplift.
  • Efficiency Analysis: Removing Residual Generation harms accuracy and halves inference throughput, while removing self-attention improves QPS with minimal precision loss.Removing the decoder maximizes QPS but collapses MRR, indicating that cross-attention remains important for routing long user histories.

6 Conclusion

CHAP addresses semantic-gap and inference-latency challenges in personalized generative retrieval through hierarchical alignment and single-pass cascading generation. Experiments and an online A/B test support its effectiveness and industrial throughput.

  • CHAP maps dynamic user intents into static item quantization space through hierarchical semantic alignment.
  • Dual-view sequence modeling combines aligned sparse SIDs with dense representations to model historical intent and fine-grained semantics.
  • Residual Cascading Generation separates historical intent routing from layer-wise decoding through single-pass inference.
  • Experiments across four real-world datasets and an online A/B test establish CHAP as a new state-of-the-art with industrial-grade inference throughput.

Limitations

The paper identifies limitations involving frozen item-side codebooks, dependence on RQ-VAE-style identifiers, and unresolved scaling behavior of acceleration.

  • Freezing the item-side codebook stabilizes quantization but may limit updates from collaborative filtering signals or query-side feedback.
  • CHAP’s generalization to lexical, learned atomic, and tree-based ItemIDs remains insufficiently investigated.
  • The acceleration ratio under changes in model size, codebook depth, and candidate budget remains an open question.
  • Model distillation for residual generation blocks is identified as a possible route toward more efficient large-scale deployment.

Ethical Considerations

The research received ethics approval and used authorized, anonymized user-related data in experiments based on aggregated statistics and anonymized behavioral signals.

  • The research was reviewed and approved by the relevant technical ethics committee.
  • Data collection followed platform compliance requirements and used user permission or authorization where applicable.
  • User-related data were anonymized or de-identified before model training, evaluation, and analysis.
  • Experiments relied on aggregated statistics and anonymized behavioral signals rather than personally identifiable information.

C Dataset Details

CHAP is evaluated on four datasets spanning authentic queries, long-term behavior, personalized instructions, and industrial search scenarios. The evaluation compares the framework with representative sparse, dense, personalized, and generative retrieval baselines.

  • Datasets: Four datasets are used: ESCI-us, KuaiSearch, Amazon, and Local-Life.Their settings cover authentic Amazon queries, large-scale Kuaishou search interactions, personalized Amazon instructions, and industrial Local-Life logs.
  • Datasets: ESCI-us uses authentic Amazon search queries with relevance-labeled products and official train-test splits on raw query and item texts.
  • Datasets: KuaiSearch contains approximately 330,000 users, 18,000,000 products, and 2,500,000 real search queries, retaining long-term behavior and long-tail distributions.
  • Baselines: Generative-retrieval baselines include DSI, NCI, TIGER, LTRGR, MERGE, and COBRA, covering identifier generation, ranking, and sparse-dense cascading.

E Evaluation Metrics Details

CHAP is evaluated with four retrieval metrics chosen to capture coverage, top-K success, first-hit ranking, and graded multi-item ranking quality. The section also documents implementation and reproducibility considerations.

  • Evaluation Metrics: Recall (R@K) measures whether the model retrieves all relevant items, providing a macroscopic view of coverage.It is used to assess overall semantic alignment.
  • Evaluation Metrics: Hit Ratio (HR@K) measures whether a target item appears in the top-K retrieved list.It is particularly suitable for focused single-item intents in industrial scenarios.
  • Evaluation Metrics: Mean Reciprocal Rank (MRR@K) evaluates the rank of the first relevant item retrieved.If no relevant item appears in the top K, the score is 0; the metric suits single-target or exact-intent datasets.
  • Evaluation Metrics: Normalized Discounted Cumulative Gain (NDCG@K) evaluates both the presence and positions of multiple relevant items using graded relevance scores.For ESCI-us, it distinguishes Exact, Substitute, and Complement relevance levels.
  • Reproducibility: The section provides implementation and optimization details to support reproducibility and fair comparison across CHAP and evaluated baselines.These details specifically address candidate-generation budgets and ranking protocols.

F.2 Implementation Settings of Baselines

The evaluation standardizes model capacity, textual inputs, sequence lengths, candidate budgets, and ranking protocols across baselines. It also defines ablations, intent-wise Local-Life subsets, and diagnostics for semantic alignment.

  • Baseline Settings: All deep learning baselines use comparable-capacity backbones, identical raw textual features, and matched historical sequence lengths.English datasets use BERT-base and T5-base, while Chinese datasets use Chinese-BERT-base and mT5-base.
  • Baseline Settings: Sparse retrieval baselines use Pyserini BM25, Doc2Query with 40 synthetic queries per document, and DeepCT with BERT-based term weighting.BM25 uses k1 = 1.2 and b = 0.75.
  • Baseline Settings: Personalized dense baselines receive equally long historical interaction sequences and use exhaustive FAISS search over the full corpus.This gives them an unbounded candidate budget for comparison with CHAP.
  • Baseline Settings: All generative-retrieval baselines share the T5-base or mT5-base backbone, while DSI and NCI construct hierarchical IDs with K-Means and TIGER and COBRA use RQ-VAE.The setup controls for model scale and SID-construction differences across baseline families.
  • Fairness and Ablations: Inference fairness is enforced by matching candidate budgets at M = 50, avoiding external heavy rerankers, and preserving baseline decoding integrity.The authors attribute CHAP’s ranking accuracy and throughput to its architectural innovations under this protocol.
  • Fairness and Ablations: Ablations remove or replace residual generation, self-attention, the decoder, HSA components, dense refinement, and other sequence-modeling components.The variants isolate contributions from hierarchical alignment, residual cascading, and continuous representations.
  • Intent-wise Evaluation: Local-Life is partitioned into General, Ambiguous, ColdStart, and LongTail segments to test performance across heterogeneous intent conditions.Ambiguous queries test preference-based disambiguation, ColdStart tests zero-shot semantic matching, and LongTail tests complex low-frequency queries.
  • Semantic Diagnostics: The semantic-alignment diagnostics combine clustering metrics with similarity-matrix analysis to assess compactness, separability, and cross-category similarity.CHAP shows a sharper local-neighborhood diagonal and lighter off-diagonal background than RQ-VAE and MERGE in Fig. 7.

J SID-Level Visualization Examples

The visual analyses examine how CHAP organizes relevant items across hierarchical SID layers and whether those layers express coarse-to-fine semantics. They report more concentrated assignments and progressively more specific product concepts.

  • Layer-wise SID Analysis: The visualization uses real item-to-SID mappings from RQ-VAE, MERGE, and CHAP for representative ESCI-us queries with multiple relevance levels.This setup compares how each method organizes query-relevant items across hierarchical layers.
  • Layer-wise SID Analysis: CHAP produces more concentrated layer-wise SID distributions than RQ-VAE and MERGE, especially at deeper layers.Lower curves indicate that relevant items occupy fewer hierarchical branches.
  • Coarse-to-Fine Semantics: Along a camera-related path, the first SID layer captures a broad visual-device cluster including cameras, webcams, and security-camera attributes.The first-layer prefix is 133.
  • Coarse-to-Fine Semantics: The second and third SID layers narrow the camera cluster toward USB webcams and then webcam products with microphone, desktop, laptop, and streaming attributes.This keyword evolution demonstrates progressively finer semantic organization along the path.
  • Item-level SID Examples: Table 9 compares generated SIDs for “crustless sandwich maker,” with RQ-VAE and CHAP sharing a codebook while MERGE uses a separate codebook.MERGE’s numeric SIDs therefore should not be compared token by token with RQ-VAE or CHAP.
  • Item-level SID Examples: The item-level examples illustrate how aligned SIDs improve the consistency of generated ItemIDs for query-relevant products.The example focuses on exact sandwich cutter and sealer products.
Loading 2608.30553v1…