ASTRA: Adaptive Semantic Tree Reasoning Architecture for Complex Table Question Answering
Summary: arXiv:2604.08999v1 Announce Type: cross
Abstract: Table serialization remains a critical bottleneck for Large Language Models (LLMs) in complex table question answering, hindered by challenges such as structural neglect, representation gaps, and reasoning opacity. Existing serialization methods fail to capture explicit hierarchies and lack schema flexibility, while current tree-based approaches suffer from limited semantic adaptability. To address these limitations, we propose ASTRA (Adaptive Semantic Tree Reasoning Architecture) including two main modules, AdaSTR and DuTR.
Introduction
In recent years, the field of Natural Language Processing (NLP) has witnessed significant advancements, particularly with the rise of Large Language Models (LLMs). However, when it comes to complex table question answering, these models face substantial hurdles. Traditional methods struggle with table serialization, resulting in various issues that affect the accuracy and efficiency of question answering.
Challenges in Table Serialization
Several key challenges plague existing table serialization techniques:
- Structural Neglect: Many existing models overlook the inherent structure of tables, leading to misinterpretations.
- Representation Gaps: Current approaches often fail to represent the data contained in tables adequately.
- Reasoning Opacity: The reasoning processes behind question answering remain unclear, hindering user trust and model transparency.
Introducing ASTRA
To tackle these challenges, we introduce ASTRA, which comprises two main components designed to enhance table serialization and reasoning capabilities:
- AdaSTR: This module harnesses the global semantic awareness of LLMs to reconstruct tables into Logical Semantic Trees. By explicitly modeling hierarchical dependencies, AdaSTR employs an adaptive mechanism tailored to optimize construction strategies based on the scale of the table.
- DuTR: The second component is a dual-mode reasoning framework. DuTR combines tree-search-based textual navigation for improved linguistic alignment with symbolic code execution, ensuring precise verification of answers derived from the tables.
Experimental Results
We conducted extensive experiments utilizing complex table benchmarks to validate the effectiveness of ASTRA. The results showed that our method consistently achieved state-of-the-art (SOTA) performance, outperforming existing models significantly. This improvement is attributed to the enhanced semantic representation and reasoning capabilities provided by the AdaSTR and DuTR modules.
Conclusion
The introduction of ASTRA marks a significant advancement in the domain of complex table question answering. By addressing critical limitations in existing serialization methods, ASTRA not only enhances the performance of LLMs but also increases the interpretability of their reasoning processes. This work paves the way for future research and development in the intersection of table data and natural language processing.