Source-linked AI summary
STAR: Semantic Table Representation with Header-Aware Clustering and Adaptive Weighted Fusion
Shui-Hsiang Hsu, Tsung-Hsiang Chou, Chen-Jui Yu, Yao-Chung Fan
TL;DR
Table retrieval must align natural-language queries with structured tables despite semantic and structural discrepancies, while QGpT remains limited by coarse sampling and fusion. STAR uses header-aware clustering, cluster-specific synthetic queries, and weighted fusion, and improves average R@1 over QGpT by 6.39 percentage points across five benchmark datasets.
Problem
QGpT enriches table representations with synthetic queries but relies on coarse partial-table sampling and simple fusion, limiting semantic diversity and fine-grained query–table alignment.
Method
STAR applies header-aware clustering to select representative instances, generates cluster-specific synthetic queries, and combines table and query embeddings through weighted fusion.
Results
6.39 percentage points average R@1 improvement over QGpT across five benchmark datasets.
Takeaways & Limitations
Combining semantic clustering with adaptive fusion provides a robust table-representation approach for table retrieval.
Takeaways & Limitations
STAR relies on informative table headers and adds computational overhead from generating cluster-specific synthetic queries.
Abstract
from arXiv · showhide
Table retrieval is the task of retrieving the most relevant tables from large-scale corpora given natural language queries. However, structural and semantic discrepancies between unstructured text and structured tables make embedding alignment particularly challenging. Recent methods such as QGpT attempt to enrich table semantics by generating synthetic queries, yet they still rely on coarse partial-table sampling and simple fusion strategies, which limit semantic diversity and hinder effective query-table alignment. We propose STAR (Semantic Table Representation), a lightweight framework that improves semantic table representation through semantic clustering and weighted fusion. STAR first applies header-aware K-means clustering to group semantically similar rows and selects representative centroid instances to construct a diverse partial table. It then generates cluster-specific synthetic queries to comprehensively cover the table's semantic space. Finally, STAR employs weighted fusion strategies to integrate table and query embeddings, enabling fine-grained semantic alignment. This design enables STAR to capture complementary information from structured and textual sources, improving the expressiveness of table representations. Experiments on five benchmarks show that STAR achieves consistently higher Recall than QGpT on all datasets, demonstrating the effectiveness of semantic clustering and adaptive weighted fusion for robust table representation. Our code is available at https://github.com/adsl135789/STAR.
1 Introduction
Table retrieval faces a semantic gap between natural-language queries and structured tables, compounded by table-length limits. STAR addresses QGpT’s sampling and fusion limitations with representative clustering, synthetic queries, and weighted integration.
- Table retrieval must bridge semantic differences between unstructured queries and structured tables under encoder token limits.
- QGpT selects the first k rows and generates synthetic queries, but its top-k sampling may not represent the entire table and its concatenation fusion is coarse.
- STAR improves table representations without modifying the underlying retriever architecture.
- STAR uses clustering to select representative content and fusion to integrate tables with synthetic queries for query–table alignment.
- STAR replaces QGpT’s top-k sampling with header-aware clustering and simple concatenation with weighted fusion.
2 Related Works
Table retrieval has progressed from lexical and dense retrieval toward structure-aware encoding and query generation. QGpT established LLM-based synthetic-query generation, while STAR targets its limited diversity and coarse fusion.
- Sparse retrieval uses lexical matching but struggles to capture deeper semantic relationships.
- Dense retrieval uses vector representations for semantic matching and has been applied to table retrieval.
- Recent approaches use generative models to produce pseudo-queries that enhance table semantic representations.
- QGpT applies LLMs to generate synthetic queries for enriching table semantic representations.
- STAR addresses QGpT’s limited semantic diversity and simple concatenation through header-aware clustering and weighted fusion.
3 Methodology
STAR comprises Semantic Clustering and Query Generation followed by Weighted Fusion. It combines header-aware representative selection, cluster-specific queries, and separately encoded table and query signals.
- STAR’s methodology consists of Semantic Clustering and Query Generation and Weighted Fusion.
- Semantic Clustering and Query Generation: SCQG replaces top-k selection by combining semantic clustering with cluster-guided query generation.
- Semantic Clustering and Query Generation: Header-aware clustering encodes the header and each instance so clustering incorporates table schema and global context.
- Semantic Clustering and Query Generation: Header and instance embeddings are weighted using α, which controls their relative importance.
- Semantic Clustering and Query Generation: K-means groups instances into k clusters and selects the instance nearest each center, preserving semantic diversity and representativeness.
- Semantic Clustering and Query Generation: For each cluster, STAR constructs a clustered table and uses an LLM to generate a cluster-specific synthetic query.
- Semantic Clustering and Query Generation: The resulting k queries correspond to different semantic clusters and cover diverse table aspects.
- Weighted Fusion: Weighted Fusion encodes the partial table and synthetic queries separately before combining them with weights.
4 Experiments
STAR is evaluated on five table-retrieval benchmarks against QGpT using multiple fusion strategies. It consistently outperforms QGpT, while dataset-specific fusion preferences and DWF adaptability are also reported.
- Experimental Setup: STAR is evaluated on five benchmark datasets against QGpT, using Recall@K and multiple fusion strategies.The datasets are Mimo (ch), Mimo (en), OTTQA, FetaQA, and E2E-WTQ.
- Fusion Analysis: Mimo (ch) and Mimo (en) perform best with FWF λ between 0.4 and 0.5, whereas OTTQA, FetaQA, and E2E-WTQ prefer lower λ values such as 0.3 or 0.4.The reported differences are associated with structural complexity and the relative clarity of table semantics.
- Fusion Analysis: DWF performs best on OTTQA, FetaQA, and E2E-WTQ while achieving competitive results on both Mimo datasets.The authors describe DWF as robust across datasets and able to balance semantic sources without manual tuning.
5 Analysis
Ablation results show that SCQG contributes most to STAR’s performance, followed by weighted fusion and header-aware clustering. Each component improves average R@1 relative to its corresponding ablation.
- Experimental Design: The ablation study reports average Recall metrics across five datasets for the full model and three component-removal variants.The variants are STAR, w/o SCQG, w/o WF, and w/o Header-aware.
- SCQG: SCQG removal reduces average R@1 from 51.86% to 47.07% (−4.79), the largest reported ablation drop.The ablation removes semantic clustering and uses top-k sampling with direct query generation.
- Weighted Fusion: Replacing weighted fusion with simple concatenation decreases average R@1 by 2.78 points.The comparison encodes tables and synthetic queries through concatenation rather than weighted fusion.
- Header-aware Clustering: Removing header information from clustering causes a 1.29-point drop in R@1.Header-aware embeddings provide global schema context during clustering.
6 Conclusion
The paper concludes that STAR addresses heuristic sampling and coarse fusion through header-aware semantic clustering, cluster-specific query generation, and weighted fusion. Across five benchmarks, STAR improves average R@1 over QGpT, while its benefits are bounded by header quality and added query-generation cost.
- Conclusion: STAR combines header-aware K-means clustering, cluster-specific synthetic query generation, and weighted fusion for table representation.The framework is designed without modifying the underlying retriever architecture.
- Conclusion: 6.39 percentage points: STAR’s average R@1 improvement over QGpT across five benchmark datasets.The conclusion reports this as evidence for combining semantic clustering with adaptive fusion.
- Limitations: Tables with vague or missing headers may benefit less from header-aware clustering.The limitation concerns the framework’s reliance on informative table headers.
- Limitations: Cluster-specific synthetic query generation adds computational overhead compared with QGpT in large-scale retrieval settings.The authors identify efficiency as a practical limitation and propose more efficient selection strategies as future work.
A Prompt for Pseudo Query Generation
The pseudo-query prompt asks an LLM to generate one natural-language query for each clustered table. It conditions the query on the table’s content and structure and offers several query types and specificity requirements.
- Prompt Objective: For each cluster C_j, the clustered table T_j is formatted and passed to an LLM to generate one synthetic query q_j.The prompt operates separately on each clustered table.
- Prompt Objective: The prompt instructs the model to generate one query based on the actual content and structure of the input table.The input is represented as a clustered table.
- Query Types: The prompt supports numerical, list, count, and select query types.Examples include averages, filtered lists, status counts, and selecting an employee with the highest revenue.
- Requirements: Generated queries should use natural conversational language and, when possible, reference real values, names, or entities in the table.The prompt also requests specificity to the table’s actual content.
- Requirements: For fact-verification tables, the prompt favors entity-specific and temporal queries; for reasoning-oriented tables, it requests multi-step or conditional queries.These requirements tailor query generation to table type.