On Linearizing Structured Data in Encoder-Decoder Language Models: Insights from Text-to-SQL

TL;DR

This work interrogates how linearization-based structured data representations are processed by encoder-decoder LMs in text-to-SQL, using a prefix-tuned T5 model on the Spider dataset. Through probing and causal tracing, the authors show that linearized inputs preserve crucial low-level textual information and that node relationships are encoded in an ego-centric manner, with structure-node encodings largely dedicated to their own nodes. They reveal duplicative robustness between encoder self-attention and decoder cross-attention for modality fusion and identify a pipeline-like inner process that mirrors schema linking, syntax prediction, and node selection. The findings suggest opportunities for model compression and more informed design of SDR systems, while providing a deeper mechanistic understanding of how linearization-based approaches handle inherently non-linear structured data. Overall, the study advances interpretability for SDR in encoder-decoder LMs and offers guidance for future research and optimization.

Abstract

Structured data, prevalent in tables, databases, and knowledge graphs, poses a significant challenge in its representation. With the advent of large language models (LLMs), there has been a shift towards linearization-based methods, which process structured data as sequential token streams, diverging from approaches that explicitly model structure, often as a graph. Crucially, there remains a gap in our understanding of how these linearization-based methods handle structured data, which is inherently non-linear. This work investigates the linear handling of structured data in encoder-decoder language models, specifically T5. Our findings reveal the model's ability to mimic human-designed processes such as schema linking and syntax prediction, indicating a deep, meaningful learning of structure beyond simple token sequencing. We also uncover insights into the model's internal mechanisms, including the ego-centric nature of structure node encodings and the potential for model compression due to modality fusion redundancy. Overall, this work sheds light on the inner workings of linearization-based methods and could potentially provide guidance for future research.

Figures (7)

• Figure 1: The input to the text-to-SQL parser consists of the query in natural language text (blue), and the relevant structured data (red), other tokens (gray). "self-node," refers to the input tokens corresponding to the expected output node where a node refers to both column and table names, and "structure-context," represents all the structured input tokens excluding the self-node. The output is the predicted SQL query (top).
• Figure 2: An illustrative sample showing the restoring effect of each encoder intermediate state. The decoder prompt: SELECT song_name FROM singer WHERE ==> age. Restoring the self-node hidden state on any layer can recover the correct prediction, while almost all other states do not have such an effect. More samples are available in Figure \ref{['fig:exp1-appendix']}.
• Figure 3: Error type analysis on decoder cross-attention corruption on the text or structure part.
• Figure 4: Error type analysis on decoder self-attention corruption on various layer ranges.
• Figure 5: Encoder state restoration effectiveness. Multi-token nodes are usually harder to recover by restoring a single state. Supplementary for Figure \ref{['fig:exp1']}.