Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReSyn: A Generalized Recursive Regular Expression Synthesis Framework

Seongmin Kim, Hyunjoon Cheon, Su-Hyeon Kim, Yo-Sub Han, Sang-Ki Ko

Abstract

Existing Programming-By-Example (PBE) systems often rely on simplified benchmarks that fail to capture the high structural complexity-such as deeper nesting and frequent Unions-of real-world regexes. To overcome the resulting performance drop, we propose ReSyn, a synthesizer-agnostic divide-and-conquer framework that decomposes complex synthesis problems into manageable sub-problems. We also introduce Set2Regex, a parameter-efficient synthesizer capturing the permutation invariance of examples. Experimental results demonstrate that ReSyn significantly boosts accuracy across various synthesizers, and its combination with Set2Regex establishes a new state-of-the-art on challenging real-world benchmark.

ReSyn: A Generalized Recursive Regular Expression Synthesis Framework

Abstract

Existing Programming-By-Example (PBE) systems often rely on simplified benchmarks that fail to capture the high structural complexity-such as deeper nesting and frequent Unions-of real-world regexes. To overcome the resulting performance drop, we propose ReSyn, a synthesizer-agnostic divide-and-conquer framework that decomposes complex synthesis problems into manageable sub-problems. We also introduce Set2Regex, a parameter-efficient synthesizer capturing the permutation invariance of examples. Experimental results demonstrate that ReSyn significantly boosts accuracy across various synthesizers, and its combination with Set2Regex establishes a new state-of-the-art on challenging real-world benchmark.

Paper Structure

This paper contains 57 sections, 8 theorems, 2 equations, 6 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Example-Based and Neural Regex Synthesis
  4. Divide-and-Conquer Synthesis
  5. Positioning of Our Work
  6. Proposed Method
  7. Definitions and Notation
  8. Three-Stage End-to-End Neural Recursive Synthesis Framework: Re-Syn
  9. Theoretical Hardness of Optimal Decomposition of Examples
  10. Experiments
  11. Experimental Setup
  12. Data Collection and Preprocessing
  13. Baselines
  14. Implementation Details
  15. Evaluation Metrics
  16. ...and 42 more sections

Key Result

Lemma 1

For any finite language $S$, the language expression cost $c_{E}(S)$ is equal to the optimal alignment cost $c(S)$.

Figures (6)

  • Figure 1: The architecture of Set2Regex. The Hierarchical Set Encoder aggregates character features into string embeddings ($h_i$) and subsequently into a global context ($c$) via PMA to ensure permutation invariance. The decoder employs a dual-attention mechanism, attending first to $c$ for global structure and then to $\{h'_i\}$ for local details.
  • Figure 2: Structural Complexity Comparison across Benchmarks. The top row displays the distribution of Top-level Operators; while Structured-Regex and Snort are dominated by Concatenations, RegExLib exhibits a diverse structural composition with significant use of Union operators. The bottom row illustrates the AST Depth distribution using a sequential color gradient (darker indicates deeper nesting). Notably, RegExLib contains a significant proportion of high-complexity instances (Depth $\ge$ 5), whereas the other benchmarks are concentrated in shallower regions.
  • Figure 3: Performance comparison across regex AST depth. Non-recursive methods ( Forest, Split-Regex) show sharp degradation with increasing depth, while Recursive maintains robust performance. The gap widens beyond depth 4, highlighting the necessity of recursive decomposition for complex real-world regexes.
  • Figure 4: Example alignment of two strings 'http' and 'ftps'
  • Figure 5: A running example of the Re-Syn framework process. The input positive set $P$ is first decomposed via Segmentation into three logical components ($P^{(1)}$: User, $P^{(2)}$: Domain, $P^{(3)}$: TLD), mirroring a Concatenation structure. Subsequently, the Router recursively determines the strategy for each subset. For instance, $P^{(1)}$ and $P^{(3)}$ are further decomposed via Partitioning (Union), while non-decomposable components are synthesized directly. Finally, the partial regexes are reconstructed bottom-up to form the complete regex. Note that this example is simplified for illustrative purposes.
  • ...and 1 more figures

Theorems & Definitions (20)

  • Lemma 1
  • Lemma 2
  • proof : Proof (Sketch)
  • Theorem 1
  • Definition 1: Expression cost
  • Example 1
  • Definition 2: Language expression cost
  • Definition 3: String decomposition cost
  • Example 2
  • Definition 4: Alignment cost
  • ...and 10 more