acceptodds
Under review as a conference paper at ICLR 2027

HiR-MKR: Hierarchical Reconstruction and Multi-Source Knowledge Retrieval for Semi-Structured Table Question Answering

Abstract

Semi-structured tables are widely used in real-world applications, but complex layouts such as multi-level headers, merged cells, and nested subtables pose significant challenges to table question answering. ST-Raptor provides structured representations through Hierarchical Orthogonal Trees (HO-Trees), but their limited explicit hierarchical semantics and lack of complementary knowledge representations constrain effective retrieval and reasoning. We propose HiR-MKR, a semi-structured table question answering framework that integrates hierarchical structure reconstruction with multi-source knowledge retrieval. Table Adaptive Hierarchical Reconstruction Algorithm (TAHRA) reconstructs explicit hierarchical structures from HO-Trees and provides a unified basis for a knowledge graph, a Record vector store, and a Path vector store. HiR-MKR further performs query decomposition and type-aware, dependency-driven retrieval and reasoning, with structure reliability assessment to control reliance on reconstructed knowledge. Experiments on WikiTQ-ST, TempTabQA-ST, and SSTQA show that HiR-MKR consistently outperforms existing methods, achieving 82.72% accuracy on SSTQA and outperforming ST-Raptor by 10.33 percentage points. It also reduces average response time and total token consumption by 59.29% and 73.00%, respectively.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Semi-structured tables, Table question answering, Hierarchical structure reconstruction, Multi-source knowledge retrieval, Knowledge graphs

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AjXAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiR-MKR for question answering over complex semi-structured tables. Starting from ST-Raptor’s HO-Tree, the method uses TAHRA to reconstruct explicit parent–child hierarchy from field statistics, then builds a knowledge graph, a Record vector store, and a Path vector store from the hierarchy-enhanced tree. For each question, an LLM decomposes it into subquestions, predicts their types and dependencies, and selects knowledge sources accordingly for retrieval and reasoning. The authors report improved accuracy over the compared methods on WikiTQ-ST, TempTabQA-ST, and SSTQA, as well as lower online response time and token consumption than ST-Raptor on SSTQA.

Critical strengths

S1.The paper addresses an important problem in semi-structured table question answering, where hierarchical layouts and complex structural relationships introduce challenges for retrieval and reasoning. S2.The methodological distinction from the closest HO-Tree baseline is clear. The paper specifically moves from implicit path/topological semantics to explicit hierarchical nodes and edges, and uses them as a shared basis for KG, Record, and Path representations, rather than merely adding a generic retriever on top of an HO-Tree S3.The results show consistent empirical improvements. In Table 1, HiR-MKR exceeds ST-Raptor on all three benchmarks. Table 2 further shows lower SSTQA accuracy after removing TAHRA, KG retrieval, query decomposition, or reliability assessment. These findings provide positive evidence for the full pipeline and several major components.

Critical weaknesses

W1. The text states that the original HO-Tree is retained when all candidate fields have the same unique-value count and delta_S < 0.05. However, Algorithm 1 returns S, F, although its second output is defined as hierarchy order L. The manuscript does not specify whether this returned value is a special no-hierarchy indicator, whether Algorithm 2 is skipped, or how it is mapped to Algorithm 2’s input L. It is therefore unclear whether aggregation is actually performed after the conservative test is triggered. The paper should also clarify whether the candidate-field set F in Algorithm 1 and the complete field set F used to preserve non-hierarchical attributes in Algorithm 2 are the same object. W2. TAHRA is a core component of the paper. While the end-to-end ablation in Table 2 indicates that it improves final accuracy on SSTQA, the paper does not report agreement between the reconstructed hierarchical fields and human annotations or the semantic structure of the original tables. Therefore, the current experiments support TAHRA’s empirical contribution to the overall pipeline, but do not directly validate whether the claimed explicit hierarchy reconstruction is accurate or clarify its applicability across different table structures. This limits assessment of the proposed mechanism and its ability to recover structural semantics. W3. Table 2 reports only an ablation without KG retrieval; it does not separately evaluate the Record or Path vector store, nor compare type-aware retrieval with a unified or simpler retrieval strategy. Thus, the experiments support the contribution of KG retrieval to the full system, but do not quantify the independent benefits of Record retrieval, Path retrieval, or their coordinated routing, nor establish whether the three-source design is necessary relative to simpler alternatives. W4. Table 3 measures only the online QA stage and treats KG/vector-store construction as offline cost, without reporting preprocessing time, token consumption, or hardware/service conditions. W5. Appendix A.5.1 notes that HiR-MKR underperforms ST-Raptor on simple questions (87.68% vs. 93.97%) possibly because additional retrieval and reasoning introduce unnecessary processing, but this explanation is not supported by concrete error cases or quantitative attribution. The paper would benefit from a small number of representative success and failure cases showing reconstructed structures, retrieved evidence, and error sources, which would better clarify the mechanism and applicability boundary.

Questions for authors

1.When the conservative test in Algorithm 1 is triggered, what exact value is passed as L to Algorithm 2, and is Algorithm 2 executed? Please also clarify whether F denotes the same field set in Algorithms 1 and 2. 2.Can the authors provide a direct evaluation of TAHRA’s reconstructed hierarchy, e.g., hierarchical-field ordering or parent–child relation accuracy on a manually checked subset? 3.What are the individual contributions of the Record vector store, Path vector store, and type-aware source routing? Could the authors report ablations or comparisons against simpler retrieval configurations? 4. What are the preprocessing time, token/API cost, and hardware or service settings for HO-Tree reconstruction, TAHRA, and KG/vector-store construction?

Reviewer_X4LeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The author propose HiR-MKR which integrates hierarchical structure reconstruction with multi-source knowledge collaboration. Specifically, they reconstruct explicit hierarchical semantics from HO-tree and using the structure as a unified basis for complementary knowledge representations and query-adaptive retrieval. Experiments on four datasets show the effectiveness of the proposed method.

Critical strengths
  1. The problem the paper studied is practical and meaningful.
  2. The improvements over ST-Raptor are significant, and the author also show that the proposed method is much efficiency to ST-Raptor.
  3. The method achieves consistent performance across different models, showing that its effectiveness is not limited to a single LLM backbone.
Critical weaknesses
  1. My main concern is the intuition behind the design, which seems to cover only a simplified case. I don’t think three statistical indicators can identify the actual hierarchy, even though two of them are highly related (unique value count and repetition ratio). And how often do these cases appear in real datasets? How do you deal with other cases, for example, two fields have similar numbers of unique values but don't have a hierarchical relationship? Then, I think the reconstruction is the basis of the pipeline and when it is wrong, it will introduce wrong relations into the KG and affect the final performance.
  2. The paper introduces a binary structure reliability judgment ρ(H) ∈ {True, False} but does not explain in detail how it is computed.
  3. The contribution of some design is unclear. Are all these three sources necessary? Also, I’m curious about the contribution of the entropy-weighted score, how much does it contribute to the performance compared with using simply the unique value count?
  4. The efficiency results need to be clarified. The results show lower inference costs, but the method requires additional KG construction and vector indexing. The comparison does not account for the offline costs. The author should clarify these and show the end-to-end cost.
Questions for authors

Please see the weaknesses.

Reviewer_LDzWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies question answering over semi-structured tables, which contain complex layouts such as multi-level headers, merged cells, and nested subtables. The HO-Tree proposed by ST-Raptor represents a table as two linked trees: a header tree that captures the hierarchy of multi-level headers, and a data tree that holds the records. Arguing that the HO-Tree leaves some hierarchical semantics only implicit, the authors propose HiR-MKR, whose core component TAHRA reorganizes the records into an explicit hierarchy. After filtering out unsuitable columns, TAHRA orders the remaining columns mainly by their number of distinct values, treating columns with fewer distinct values as higher-level categories, and groups the records along this order to form a tree. Based on this tree, the authors build three knowledge sources: a knowledge graph of parent-child relations and record values, a vector store that embeds each record’s hierarchical path together with its field values, and a vector store that embeds only the hierarchical paths.

Given a question, HiR-MKR decomposes it into subquestions, classifies each according to whether it asks for facts about an entity, hierarchical relations, or aggregation and comparison, and identifies dependencies among them. It then retrieves evidence for each subquestion from the knowledge sources assigned to its type, answers the subquestions in order, and generates the final answer. The authors evaluate HiR-MKR on three benchmarks -- WikiTQ-ST, TempTabQA-ST, and SSTQA -- and report 82.14%, 87.43%, and 82.72% accuracy, respectively, outperforming ST-Raptor by around 10 percentage points on each benchmark.

Critical strengths

S1. HiR-MKR consistently outperforms ST-Raptor by 9.84 to 10.97 percentage points on all three benchmarks, and improves over directly prompting the same backbone LLM, DeepSeek-V3, by 12.50 to 23.62 percentage points.

Critical weaknesses

W1. The motivation and central claim of the paper are not sufficiently substantiated.

W1.1. The paper does not explain how the hierarchy introduced by TAHRA differs from the header hierarchy already represented in the HO-Tree. It provides no example showing an original table, its HO-Tree, and the reconstructed tree, nor a question whose answer changes due to the reconstruction. The authors should provide such an example and report how often TAHRA actually reconstructs a hierarchy.

W1.2. The paper does not show which types of questions benefit from TAHRA. Appendix A.5.1 reports results only by difficulty level and omits the w/o TAHRA setting. The authors should report the accuracy of the full model and the w/o TAHRA setting for each question type.

W1.3. There is no evidence that TAHRA recovers the actual hierarchy of the table. TAHRA places columns with fewer distinct values higher, but does not check whether each value of one column falls under a single value of another. For example, a region–branch pair forms a true hierarchy, but a gender–job title pair, in which the same job title appears under multiple genders, is turned into a tree in the same way. The authors should verify whether the reconstructed hierarchies satisfy this nesting relationship.

W2. The empirical comparison and the novelty of HiR-MKR are not clearly established.

W2.1. Tables 1 and 8 are inconsistent. In Table 8, DeepSeek-V3 outperforms GPT-4o at every difficulty level of SSTQA (92.94 vs. 86.64, 61.83 vs. 59.75, and 47.19 vs. 43.26), whereas in Table 1 its overall SSTQA accuracy is lower (63.22 vs. 66.45). If both tables are based on the same questions and scoring, this is impossible, since the overall accuracy is a weighted average of the per-level accuracies. The authors should clarify the source of the baseline results and the evaluation setup used for each table, and explain this discrepancy.

W2.2. The boundary of the novelty of HiR-MKR is unclear. Question decomposition [1], query-adaptive retrieval [2], and hybrid graph–vector retrieval [3] have clear precedents, and ST-Raptor itself already decomposes questions into sub-questions and retrieves data over the tree. These precedents do not mean that the specific combination in HiR-MKR is identical to prior work, but the paper does not explain how its design differs from them or from closely related table reasoning methods such as Binder [4], Dater [5], Chain-of-Table [6], and TableRAG [7]. In particular, Dater also decomposes questions for table-based reasoning.

[1] D. Zhou et al. “Least-to-Most Prompting Enables Complex Reasoning in Large Language Models.” ICLR 2023. [2] S. Jeong et al. “Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity.” NAACL 2024. [3] B. Sarmah et al. “HybridRAG: Integrating Knowledge Graphs and Vector Retrieval Augmented Generation for Efficient Information Extraction.” In Proceedings of ICAIF 2024, pp. 608–616. [4] Z. Cheng et al. “Binding Language Models in Symbolic Languages.” ICLR 2023. [5] Y. Ye et al. “Large Language Models are Versatile Decomposers: Decomposing Evidence and Questions for Table-based Reasoning.” SIGIR 2023. [6] Z. Wang et al. “Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding.” ICLR 2024. [7] S. Chen et al. “TableRAG: Million-Token Table Understanding with Language Models.” NeurIPS 2024.

Questions for authors

Q1. How can Tables 1 and 8 be reconciled? Please provide the number of questions in each difficulty level, the scoring method (e.g., exact match, normalized matching, or an LLM-based judge), the aggregation rule, and the source of each baseline result. Were the baselines evaluated under the same setup as HiR-MKR?

Q2. How is the reliability assessment ρ(Hh)\rho(H_h) computed, and what exactly is the fallback “question answering over the original table”? How often is the fallback triggered, and what is the accuracy on each branch?

Q3. Is there any mechanism to detect or recover from errors in question decomposition? For example, what happens when a subquestion is misclassified, or when an intermediate answer passed through a Derive or Condition dependency is incorrect or empty? How accurate are the decomposition and type classification steps?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.