[Paper] Semantic Triplet Restoration: A Novel Protocol for Hierarchical Table Understanding in Large Language Models

Source: arXiv - 2605.31550v1

Overview

The paper introduces Semantic Triplet Restoration (STR), a new way to represent tables for question‑answering systems that use large language models (LLMs). Instead of feeding the model a raw HTML or Markdown rendering of a table, STR rewrites every cell as a concise “triplet” – item path, feature path, value – making the table’s hierarchical structure explicit and dramatically cutting down on token usage.

Key Contributions

• Semantic Triplet Representation: A compact, fact‑like encoding of each table cell that captures row entities, hierarchical column attributes, and cell values.
• TripletQL Router: A lightweight, query‑aware component that selects the most relevant subset of triplets (or a suitable rendering) for a given question, reducing unnecessary context.
• Empirical Validation: Demonstrated that STR matches or outperforms traditional HTML/Markdown pipelines on four bilingual (Chinese & English) table‑QA benchmarks.
• Efficiency Gains: Shows larger relative improvements for smaller LLMs and for tables with many rows/columns, highlighting benefits under tight inference budgets.
• Open‑source Release: Code, data, and reproducible scripts are publicly available, encouraging community adoption and further research.

Methodology

1. Triplet Construction

   • Item Path: Traverses the row hierarchy (e.g., Country → State).
   • Feature Path: Traverses the column hierarchy (e.g., Year → Revenue).
   • Value: The actual cell content (numeric, textual, etc.).
     The authors parse the original table (HTML/Markdown) to extract these hierarchical paths and generate a flat list of triplets.

2. TripletQL (Query‑aware Router)

   • Takes the user question, encodes it with a small transformer, and scores each triplet for relevance.
   • Returns either a filtered set of top‑k triplets or a fallback rendering (e.g., the original HTML) if the question is too broad.

3. Integration with LLMs

   • The selected triplets are concatenated with the question and fed to a downstream LLM (e.g., LLaMA‑2, GPT‑3.5).
   • Because each triplet is a short, self‑contained fact, the model can reason directly over the semantic structure without learning implicit layout cues.

Results & Findings

• Accuracy: STR consistently matches or slightly improves exact‑match scores over HTML‑based pipelines.
• Token Efficiency: Average input length drops by roughly one‑third, freeing up context windows for longer reasoning chains.
• Model Size Sensitivity: Gains are most pronounced for 7B‑parameter models, where the relative improvement can exceed 2 % absolute EM.
• Scalability: For tables exceeding 200 cells, STR’s token savings become critical, preventing context‑window overflow.

Practical Implications

• Cost‑Effective Deployment: Companies can run table‑QA services on cheaper, smaller LLMs without sacrificing accuracy, lowering inference costs on cloud platforms.
• Faster Response Times: Fewer tokens mean reduced latency, which is valuable for real‑time analytics dashboards or conversational assistants that need to answer data‑driven queries instantly.
• Simplified Prompt Engineering: Developers no longer need to craft elaborate prompts to teach the model about row/column spans; the triplet format is self‑describing.
• Better Interoperability: The triplet schema can be generated from any tabular source (CSV, Excel, SQL results), making it easy to plug into existing data pipelines.
• Enhanced Explainability: Since each fact is explicit, debugging wrong answers becomes a matter of inspecting the selected triplets rather than deciphering hidden layout cues.

Limitations & Future Work

• Complex Cell Content: The current approach treats cell values as atomic strings; handling rich content (lists, images, nested tables) requires extensions.
• Header Ambiguity: In extremely deep hierarchies, the feature path can become long, potentially re‑inflating token counts; smarter path compression is an open problem.
• Cross‑Table Reasoning: STR focuses on single‑table QA; integrating multiple tables or external knowledge bases remains unexplored.
• Language Coverage: Experiments are limited to Chinese and English; adapting the parser to languages with different script directions or table conventions may need additional work.

Overall, Semantic Triplet Restoration offers a pragmatic, token‑efficient bridge between tabular data and LLMs, opening the door for more scalable and developer‑friendly table‑question answering systems.

Authors

• Yibin Zhao
• Fangxin Shang
• Dingrui Yang
• Yuqi Wang

Paper Information

• arXiv ID: 2605.31550v1
• Categories: cs.CL
• Published: May 29, 2026
• PDF: Download PDF