Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SParC: Cross-Domain Semantic Parsing in Context

Tao Yu, Rui Zhang, Michihiro Yasunaga, Yi Chern Tan, Xi Victoria Lin, Suyi Li, Heyang Er, Irene Li, Bo Pang, Tao Chen, Emily Ji, Shreya Dixit, David Proctor, Sungrok Shim, Jonathan Kraft, Vincent Zhang, Caiming Xiong, Richard Socher, Dragomir Radev

TL;DR

SParC addresses the gap in context-aware semantic parsing by introducing a large, cross-domain dataset of 4,298 question sequences (12k+ questions) over 200 databases in 138 domains, annotated with executable SQL. The authors adapt two state-of-the-art context-dependent models (CD-Seq2Seq and SyntaxSQL-con) and show that, despite modeling history, the best approach achieves only 20.2% exact-match on all questions and under 10% on interaction-level sequences, underscoring significant modeling challenges for cross-domain, context-dependent text-to-SQL. The work provides a comprehensive data collection pipeline, in-depth analysis of contextual dependencies and linguistic phenomena, and rigorous benchmarking, highlighting the need for advances in context fusion and cross-domain generalization. The dataset and baselines, along with a public leaderboard, are intended to spur progress toward robust, multi-turn SQL understanding across diverse domains.

Abstract

We present SParC, a dataset for cross-domainSemanticParsing inContext that consists of 4,298 coherent question sequences (12k+ individual questions annotated with SQL queries). It is obtained from controlled user interactions with 200 complex databases over 138 domains. We provide an in-depth analysis of SParC and show that it introduces new challenges compared to existing datasets. SParC demonstrates complex contextual dependencies, (2) has greater semantic diversity, and (3) requires generalization to unseen domains due to its cross-domain nature and the unseen databases at test time. We experiment with two state-of-the-art text-to-SQL models adapted to the context-dependent, cross-domain setup. The best model obtains an exact match accuracy of 20.2% over all questions and less than10% over all interaction sequences, indicating that the cross-domain setting and the con-textual phenomena of the dataset present significant challenges for future research. The dataset, baselines, and leaderboard are released at https://yale-lily.github.io/sparc.

SParC: Cross-Domain Semantic Parsing in Context

TL;DR

SParC addresses the gap in context-aware semantic parsing by introducing a large, cross-domain dataset of 4,298 question sequences (12k+ questions) over 200 databases in 138 domains, annotated with executable SQL. The authors adapt two state-of-the-art context-dependent models (CD-Seq2Seq and SyntaxSQL-con) and show that, despite modeling history, the best approach achieves only 20.2% exact-match on all questions and under 10% on interaction-level sequences, underscoring significant modeling challenges for cross-domain, context-dependent text-to-SQL. The work provides a comprehensive data collection pipeline, in-depth analysis of contextual dependencies and linguistic phenomena, and rigorous benchmarking, highlighting the need for advances in context fusion and cross-domain generalization. The dataset and baselines, along with a public leaderboard, are intended to spur progress toward robust, multi-turn SQL understanding across diverse domains.

Abstract

We present SParC, a dataset for cross-domainSemanticParsing inContext that consists of 4,298 coherent question sequences (12k+ individual questions annotated with SQL queries). It is obtained from controlled user interactions with 200 complex databases over 138 domains. We provide an in-depth analysis of SParC and show that it introduces new challenges compared to existing datasets. SParC demonstrates complex contextual dependencies, (2) has greater semantic diversity, and (3) requires generalization to unseen domains due to its cross-domain nature and the unseen databases at test time. We experiment with two state-of-the-art text-to-SQL models adapted to the context-dependent, cross-domain setup. The best model obtains an exact match accuracy of 20.2% over all questions and less than10% over all interaction sequences, indicating that the cross-domain setting and the con-textual phenomena of the dataset present significant challenges for future research. The dataset, baselines, and leaderboard are released at https://yale-lily.github.io/sparc.

Paper Structure

This paper contains 35 sections, 7 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Context-independent semantic parsing
  4. Context-dependent semantic parsing with SQL labels
  5. Context-dependent semantic parsing with denotations
  6. Conversational QA and dialogue system
  7. Data Collection
  8. Interaction goal selection
  9. Question creation
  10. SQL annotation
  11. Data review and post-process
  12. Data Statistics and Analysis
  13. Data statistics
  14. Contextual dependencies of questions
  15. Contextual linguistic phenomena
  16. ...and 20 more sections

Figures (5)

  • Figure 1: Two question sequences from the SParC dataset. Questions ($Q_i$) in each sequence query a database ($D_m$), obtaining information sufficient to complete the interaction goal ($C_m$). Each question is annotated with a corresponding SQL query ($S_i$). SQL segments from the interaction context are underlined.
  • Figure 2: The heatmap shows the percentage of SQL token overlap between questions in different turns. Token overlap is greater between questions that are closer to each other and the degree of overlap increases as interaction proceeds. Most questions have dependencies that span 3 or fewer turns.
  • Figure 3: Percentage of question sequences that contain a particular SQL keyword at a given turn. The complexity of questions increases as interaction proceeds on SParC as more SQL keywords are triggered. The same trend was not observed on ATIS.
  • Figure 4: More examples in SParC.
  • Figure 5: Additional example in SParC annotated with different thematic relations. Entities (purple), properties (magenta), constraints (red), and answers (orange) are colored.