Back

RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL

Haoyang LiJing ZhangCuiping LiHong Chen

article2023AAAI408 citations

Proposes a text-to-SQL framework that decouples schema linking from query generation by pre-filtering relevant database elements with a ranking cross-encoder and generating SQL skeletons before full queries to guide decoding.

Listen

Modern data management systems rely heavily on relational databases, but non-technical users often struggle to query them effectively because they lack expertise in Structured Query Language (SQL). Automated text-to-SQL systems address this gap by translating everyday natural language questions into database queries. However, standard sequence-to-sequence language models face significant performance hurdles because they attempt to simultaneously determine which database tables and columns to reference (schema linking) and assemble the structural query keywords (skeleton parsing). This coupling creates excessive complexity and leaves models vulnerable to variations in phrasing and database structures.

The article demonstrates and evaluates a novel framework called RESDSQL, which decouples schema linking from skeleton parsing to improve both query generation accuracy and overall system robustness. The authors evaluate this approach across standard cross-domain benchmarks and challenging stress-test scenarios that simulate real-world usage.

To achieve this separation, the approach implements a two-stage process. First, an independent cross-encoder model identifies, ranks, and filters database tables and columns, feeding only the most relevant items into the main translation model. This cross-encoder incorporates a column-enhancement mechanism to infer table names when users mention only column names, along with an imbalanced-data loss function. Second, the sequence-to-sequence decoder generates the overall SQL skeleton before generating the final, complete SQL query, allowing the high-level structural framework to guide the detailed output.

Empirical evaluations demonstrate substantial performance gains across key metrics. First, on the challenging Spider benchmark test set, RESDSQL set a new state-of-the-art result by achieving 79.9% execution accuracy, outperforming the previous top baseline of 75.5% by 4.4 percentage points. Second, the framework proves highly parameter-efficient: its base configuration outmatched standard baseline models that were roughly 13 times larger. Third, ablation analyses show that filtering and ranking schema items is the primary performance driver, improving exact match accuracy by 4.5 percentage points and execution accuracy by 7.8 percentage points. Finally, across three robustness benchmarks featuring synonyms, domain knowledge paraphrases, and omitted column names, RESDSQL consistently surpassed leading competitors by wide margins, achieving up to 81.9% execution accuracy on realistic perturbation tests.

These findings indicate that separating structural query planning from schema item selection substantially reduces model training difficulty and operational noise. For organizations deploying natural language interfaces to databases, this translates directly to lower computational infrastructure costs, higher query reliability, and greater resilience against ambiguous or imperfect user inputs without requiring complex graph architectures.

Organizations developing or deploying natural language query interfaces should consider adopting decoupled pre-filtering and two-step skeleton decoding architectures over monolithic models. Teams should also select filtering thresholds (such as the number of top tables and columns retained) based on their specific schema complexity, balancing the trade-off between missing required fields and introducing unnecessary noise.

Confidence in these findings is high given the extensive validation across multiple public benchmark datasets. However, decision-makers should note that the system relies on fixed selection thresholds for database elements and requires normalized SQL training conventions. In operational settings with highly specialized enterprise jargon or vastly larger enterprise schemas, targeted pilot testing remains advisable.

Renmin University of Chinadatabase schemasparsing algorithmrelational schemasRESDSQLtext-to-SQLschema linking
arXiv: 2302.05965RUCKBReasoning/RESDSQL

Abstract

One of the recent best attempts at Text-to-SQL is the pre-trained language model. Due to the structural property of the SQL queries, the seq2seq model takes the responsibility of parsing both the schema items (i.e., tables and columns) and the skeleton (i.e., SQL keywords). Such coupled targets increase the difficulty of parsing the correct SQL queries especially when they involve many schema items and logic operators. This paper proposes a ranking-enhanced encoding and skeleton-aware decoding framework to decouple the schema linking and the skeleton parsing. Specifically, for a seq2seq encoder-decode model, its encoder is injected by the most relevant schema items instead of the whole unordered ones, which could alleviate the schema linking effort during SQL parsing, and its decoder first generates the skeleton and then the actual SQL query, which could implicitly constrain the SQL parsing. We evaluate our proposed framework on Spider and its three robustness variants: Spider-DK, Spider-Syn, and Spider-Realistic. The experimental results show that our framework delivers promising performance and robustness. Our code is available at https://github.com/RUCKBReasoning/RESDSQL.

Topics

database schemasparsing algorithmrelational schemassql

Keywords

RESDSQLtext-to-SQLschema linkingskeleton parsingsemantic parsingexecution-guided SQL selectorskeleton-aware decoderdatabase schema

Table of Contents

  • Introduction
  • Problem Definition
  • Methodology
  • Model Overview
  • Ranking-Enhanced Encoder
  • Skeleton-Aware Decoder
  • Experiments
  • Experimental Setup
  • Results on Spider
  • Results on Robustness Settings
  • Ablation Studies
  • Related Work
  • Encoder-Decoder Architecture
  • Schema Item Classification
  • Intermediate Representation
  • Conclusion
  • Acknowledgments
  • References

Knowls

  1. Knowl 1 — RESDSQL Framework Architecture

    model/method

    RESDSQL (Ranking-Enhanced Encoding and Skeleton-Aware Decoding for Text-to-SQL) is an encoder-decoder framework designed to decouple schema linking from SQL skeleton parsing in natural language database querying. In standard sequence-to-sequence Text-to-SQL architectures, the model must simultaneously align question entities with database schema items and parse the complex operator syntax of the SQL query. RESDSQL separates these concerns into two components:

    1. Ranking-Enhanced Encoder: An auxiliary RoBERTa cross-encoder evaluates the natural language question against candidate schema elements (tables and columns, represented with semantic names) to compute item relevance scores. The top-k1k_1 tables and top-k2k_2 columns per table are ranked and serialized using their original database names, concatenated with foreign key constraints, and supplied as input to the seq2seq encoder (e.g., T5).
    2. Skeleton-Aware Decoder: Rather than generating the final SQL string directly, the autoregressive transformer decoder first generates the abstract SQL skeleton (composed of SQL keywords and operator placeholders) and then generates the complete executable SQL query. The generated skeleton acts as a prefix that constrains and guides subsequent query generation via the decoder's masked self-attention mechanism. An execution-guided decoding selector retains the first executable SQL candidate within beam search.

  2. Knowl 2 — Cross-Encoder Schema Item Classification and Ranking

    model/method

    To filter and prioritize database schema items prior to sequence-to-sequence translation, RESDSQL employs an auxiliary cross-encoder that jointly classifies tables and columns based on the input question.

    • Input Representation: Given a natural language question qq and a database schema SS containing tables T={t1,t2,…,tN}T = \{t_1, t_2, \dots, t_N\} with columns {c1i,…,cnii}\{c_1^i, \dots, c_{n_i}^i\} for table tit_i, the input sequence is serialized using semantic names: X=q∣t1:c11,…,cn11∣⋯∣tN:c1N,…,cnNNX = q \mid t_1 : c_1^1, \dots, c_{n_1}^1 \mid \dots \mid t_N : c_1^N, \dots, c_{n_N}^N where ∣\mid denotes the delimiter token.
    • Encoding and Pooling: The serialized sequence XX is processed through a RoBERTa encoder. Because a single table or column name may be tokenized into multiple subwords, a pooling module consisting of a two-layer Bidirectional LSTM (BiLSTM) and a non-linear fully-connected layer aggregates token vectors into a single representation Ti∈R1×dT_i \in \mathbb{R}^{1 \times d} for each table tit_i and Cki∈R1×dC_k^i \in \mathbb{R}^{1 \times d} for each column ckic_k^i, where dd denotes the hidden dimension.
    • Top-kk Filtering: At inference, schema items are sorted by their classification probabilities. The top-k1k_1 tables (e.g., k1=4k_1 = 4) and top-k2k_2 columns per retained table (e.g., k2=5k_2 = 5) are selected using original database names to construct the ranked input sequence for the generator, eliminating irrelevant schema elements.

  3. Knowl 3 — Column-Enhanced Attention Layer for Table Representation

    model/method

    In Text-to-SQL tasks, questions often mention specific column names without explicitly citing their parent table names, resulting in misclassified or omitted tables. To alleviate this issue, RESDSQL injects column representations into table representations via a column-enhanced multi-head attention layer.

    For table tit_i with initial pooled embedding Ti∈R1×dT_i \in \mathbb{R}^{1 \times d} and column embedding matrix C:i∈Rni×dC_{:}^i \in \mathbb{R}^{n_i \times d} (where nin_i is the number of columns belonging to tit_i and dd is the hidden size), a column-attentive table embedding TiC∈R1×dT_i^C \in \mathbb{R}^{1 \times d} is computed as: TiC=MultiHeadAttn(Ti,C:i,C:i,h)T_i^C = \text{MultiHeadAttn}\left(T_i, C_{:}^i, C_{:}^i, h\right) where TiT_i serves as the query, C:iC_{:}^i serves as both key and value, and hh is the number of attention heads (set to h=8h = 8).

    The final column-enhanced table embedding T^i∈R1×d\hat{T}_i \in \mathbb{R}^{1 \times d} is obtained by combining the original table embedding with TiCT_i^C via a residual connection and row-wise L2L_2 normalization: T^i=Norm(Ti+TiC)\hat{T}_i = \text{Norm}\left(T_i + T_i^C\right)

  4. Knowl 4 — Multi-Task Focal Loss for Schema Item Classification

    equation

    Because target SQL queries only reference a small subset of database tables and columns, the schema classification objective suffers from severe class imbalance between positive and negative labels. To address this, the cross-encoder in RESDSQL is optimized using focal loss in a multi-task learning formulation:

    L1=1N∑i=1NFL(yi,y^i)+1M∑i=1N∑k=1niFL(yki,y^ki)\mathcal{L}_1 = \frac{1}{N} \sum_{i=1}^N FL\left(y_i, \hat{y}_i\right) + \frac{1}{M} \sum_{i=1}^N \sum_{k=1}^{n_i} FL\left(y_k^i, \hat{y}_k^i\right)

    where:

    • NN is the number of tables in the database schema.
    • nin_i is the number of columns in the ii-th table, and M=∑i=1NniM = \sum_{i=1}^N n_i is the total number of columns in the database.
    • yi∈{0,1}y_i \in \{0, 1\} is the ground-truth binary label indicating whether table ii is referenced in the target SQL query, and yki∈{0,1}y_k^i \in \{0, 1\} is the ground-truth binary label for the kk-th column of table ii.
    • FLFL is the focal loss function with focusing parameter γ=2\gamma = 2 and balancing factor α=0.75\alpha = 0.75.
    • y^i\hat{y}_i and y^ki\hat{y}_k^i are predicted classification probabilities computed by multi-layer perceptrons: y^i=σ((T^iU1t+b1t)U2t+b2t)\hat{y}_i = \sigma\left(\left(\hat{T}_i U_1^t + b_1^t\right) U_2^t + b_2^t\right) y^ki=σ((CkiU1c+b1c)U2c+b2c)\hat{y}_k^i = \sigma\left(\left(C_k^i U_1^c + b_1^c\right) U_2^c + b_2^c\right) where U1t,U1c∈Rd×wU_1^t, U_1^c \in \mathbb{R}^{d \times w}, b1t,b1c∈Rwb_1^t, b_1^c \in \mathbb{R}^w, U2t,U2c∈Rw×2U_2^t, U_2^c \in \mathbb{R}^{w \times 2}, and b2t,b2c∈R2b_2^t, b_2^c \in \mathbb{R}^2 are trainable parameters, σ(⋅)\sigma(\cdot) is the softmax function, T^i\hat{T}_i is the column-enhanced table embedding, and CkiC_k^i is the column embedding.

  5. Knowl 5 — Skeleton-Aware Decoding Objective

    equation

    To bridge the semantic gap between natural language questions and structured SQL queries, the seq2seq decoder in RESDSQL is trained to generate the SQL skeleton prior to emitting the full SQL query. The optimization objective across GG training instances is:

    L2=1G∑i=1Gp(lis,li∣Si)\mathcal{L}_2 = \frac{1}{G} \sum_{i=1}^G p\left(l_i^s, l_i \mid S_i\right)

    where:

    • SiS_i is the serialized input sequence containing the question, the ranked top-kk schema items, and foreign key relations.
    • lisl_i^s is the SQL skeleton consisting of SQL keywords and slot placeholders (e.g., select _ from _ where _ order by _ asc).
    • lil_i is the target executable SQL query.

    Due to the autoregressive masked self-attention of the transformer decoder, the generated skeleton lisl_i^s acts as a structural scaffold that conditions each subsequent token in lil_i, enabling the model to copy structural keywords directly or focus on copying schema items and literal values into the corresponding skeleton slots.

  6. Knowl 6 — SQL Query Normalization and Skeleton Extraction

    algorithm

    To eliminate annotator style variations and extract training targets for the skeleton-aware decoder, RESDSQL normalizes ground-truth SQL queries and derives their structural skeletons according to the following procedure:

    Input: Original SQL query SQLoSQL_o
    Output: Normalized SQL query SQLnSQL_n, SQL skeleton SQLsSQL_s
    Procedure NormalizeSQL(SQLoSQL_o):
        Convert all keywords and schema item names to lowercase
        Add spaces around parentheses and replace double quotes with single quotes
        If an "order by" clause does not specify direction:
            Append "asc" to the "order by" clause
        Remove "as" alias clauses and replace table aliases with original table names
        Return normalized query SQLnSQL_n
    Procedure ExtractSkeleton(SQLnSQL_n):
        Identify all SQL keywords in SQLnSQL_n (excluding "join on" keywords, which lack explicit natural language question counterparts)
        Replace all table names, column names, constants, and value expressions with placeholder slots "_"
        Extract the sequence of preserved keywords and placeholder slots to form SQLsSQL_s
        Return SQLsSQL_s

    For example, given the original query: SELECT T1.duration, T1.file_size FROM files AS T1 JOIN song AS T2 ON T1.f_id = T2.f_id WHERE T2.genre = "pop" ORDER BY T2.song_name The normalization step produces: select files.duration, files.file_size from files join song on files.f_id = song.f_id where song.genre = 'pop' order by song.song_name asc The extracted skeleton is: select _ from _ where _ order by _ asc

  7. Knowl 7 — Text-to-SQL Performance on Spider Benchmark

    empirical result

    RESDSQL was evaluated on the cross-domain Spider benchmark (7,000 training, 1,034 development, and 2,147 test queries) using Exact-set-Match accuracy (EM, %) and EXecution accuracy (EX, %).

    Approach Dev Set Test Set
    EM EX EM EX
    Non-seq2seq methods
    RAT-SQL + GRAPPA 73.4 - 69.6 -
    RAT-SQL + GAP + NatSQL 73.7 75.0 68.7 73.3
    SMBOP + GRAPPA 74.7 75.0 69.5 71.1
    DT-Fixup SQL-SP + RoBERTa 75.0 - 70.9 -
    LGESQL + ELECTRA 75.1 - 72.0 -
    S2S^2SQL + ELECTRA 76.4 - 72.1 -
    Seq2seq methods
    T5-3B 71.5 74.4 68.0 70.1
    T5-3B + PICARD 75.5 79.3 71.9 75.1
    RASAT + PICARD 75.3 80.5 70.9 75.5
    Proposed RESDSQL
    RESDSQL-Base 71.7 77.9 - -
    RESDSQL-Base + NatSQL 74.1 80.2 - -
    RESDSQL-Large 75.8 80.1 - -
    RESDSQL-Large + NatSQL 76.7 81.9 - -
    RESDSQL-3B 78.0 81.8 - -
    RESDSQL-3B + NatSQL 80.5 84.1 72.0 79.9

    RESDSQL-Base (71.7% EM / 77.9% EX on the dev set) outperforms standard T5-3B (71.5% EM / 74.4% EX) despite having substantially fewer parameters. When combined with NatSQL intermediate representation, RESDSQL-3B achieves 80.5% EM / 84.1% EX on the development set and 72.0% EM / 79.9% EX on the hidden test set, setting a new state-of-the-art result and improving test execution accuracy by 4.4% over T5-3B + PICARD (75.5%).

  8. Knowl 8 — Robustness on Perturbed Spider Benchmarks

    empirical result

    To measure generalization and robustness under natural language perturbations, RESDSQL-3B + NatSQL (trained exclusively on Spider's training set) was evaluated on three Spider variants: Spider-DK (535 queries with domain-knowledge paraphrasing), Spider-Syn (1,034 queries with synonym substitutions), and Spider-Realistic (508 queries with explicit column references removed).

    Approach Spider-DK Spider-Syn Spider-Realistic
    EM (%) EX (%) EM (%) EX (%) EM (%) EX (%)
    RAT-SQL + BERT 40.9 - 48.2 - 58.1 62.1
    RAT-SQL + GRAPPA 38.5 - 49.1 - 59.3 -
    T5-3B - - 59.4 65.3 63.2 65.0
    LGESQL + ELECTRA 48.4 - 64.6 - 69.2 -
    TKK-3B - - 63.0 68.2 68.5 71.1
    T5-3B + PICARD - - - - 68.7 71.4
    RASAT + PICARD - - - - 69.7 71.9
    LGESQL + ELECTRA + SUN 52.7 - 66.9 - 70.9 -
    RESDSQL-3B + NatSQL 53.3 66.0 69.1 76.9 77.4 81.9

    RESDSQL-3B + NatSQL establishes superior performance across all three benchmarks, achieving 53.3% EM / 66.0% EX on Spider-DK, 69.1% EM / 76.9% EX on Spider-Syn, and 77.4% EM / 81.9% EX on Spider-Realistic. This represents a +6.5% EM absolute improvement over the previous state of the art (70.9%) on Spider-Realistic, showing that cross-encoder schema filtering effectively mitigates entity-linking sensitivity to question phrasing.

  9. Knowl 9 — Ablation Studies on Schema Ranking, Skeleton Parsing, Column Enhancement, and Focal Loss

    empirical result

    Ablation studies on the Spider development set quantify the performance impact of each individual design element in RESDSQL:

    1. Cross-Encoder Architecture and Loss (measured by AUC):
    Model Variant Table AUC Column AUC Total AUC
    Cross-encoder (Full) 0.9973 0.9957 1.9930
    - w/o Column-Enhanced Layer 0.9965 0.9939 1.9904
    - w/o Focal Loss (with Cross-Entropy) 0.9958 0.9943 1.9901

    Removing the column-enhanced layer degrades table AUC from 0.9973 to 0.9965 and column AUC from 0.9957 to 0.9939. Replacing focal loss with standard cross-entropy loss causes the total AUC score to drop from 1.9930 to 1.9901 due to severe positive/negative label imbalance.

    1. Seq2Seq Decoding and Ranking (measured on RESDSQL-Base):
    Model Variant EM (%) EX (%)
    RESDSQL-Base 71.7 77.9
    - w/o ranking schema items (unordered full schema) 67.2 70.1
    - w/o skeleton parsing (direct SQL generation) 71.0 77.1

    Replacing the top-kk ranked schema items with the full unordered schema sequence leads to a major degradation of 4.5% in EM (71.7% to 67.2%) and 7.8% in EX (77.9% to 70.1%). Removing the intermediate skeleton generation leads to drops of 0.7% EM and 0.8% EX.

Coverage note — Hyperparameter optimization grids and standard database content extraction heuristics were omitted as standard training configurations.

References

  1. 1.Bogin, B.; Gardner, M.; and Berant, J. 2019. Global Reasoning over Database Structures for Text-to-SQL Parsing. In EMNLP-IJCNLP 2019, 3657–3662.
  2. 2.Cai, R.; Xu, B.; Zhang, Z.; Yang, X.; Li, Z.; and Liang, Z. 2018. An Encoder-Decoder Framework Translating Natural Language to Database Queries. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, 3977–3983.
  3. 3.Cai, R.; Yuan, J.; Xu, B.; and Hao, Z. 2021. SADGA: Structure-Aware Dual Graph Aggregation Network for Text-to-SQL. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, 7664–7676.
  4. 4.Cao, R.; Chen, L.; Chen, Z.; Zhao, Y.; Zhu, S.; and Yu, K. 2021. LGESQL: Line Graph Enhanced Text-to-SQL Model with Mixed Local and Non-Local Relations. In ACL/IJCNLP 2021, 2541–2555.
  5. 5.Chen, D.; Lin, Y.; Li, W.; Li, P.; Zhou, J.; and Sun, X. 2020. Measuring and Relieving the Over-Smoothing Problem for Graph Neural Networks from the Topological View. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, 3438–3445.
  6. 6.Deng, X.; Awadallah, A. H.; Meek, C.; Polozov, O.; Sun, H.; and Richardson, M. 2021. Structure-Grounded Pretraining for Text-to-SQL. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, 1337–1350.
  7. 7.Devlin, J.; Chang, M.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, 4171–4186.
  8. 8.Gan, Y.; Chen, X.; Huang, Q.; Purver, M.; Woodward, J. R.; Xie, J.; and Huang, P. 2021a. Towards Robustness of Text-to-SQL Models against Synonym Substitution. In ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, 2505–2515.
  9. 9.Gan, Y.; Chen, X.; and Purver, M. 2021. Exploring Underexplored Limitations of Cross-Domain Text-to-SQL Generalization. In EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, 8926–8931.
  10. 10.Gan, Y.; Chen, X.; Xie, J.; Purver, M.; Woodward, J. R.; Drake, J. H.; and Zhang, Q. 2021b. Natural SQL: Making SQL Easier to Infer from Natural Language Specifications. In Findings of EMNLP 2021, 2030–2042.
  11. 11.Gao, C.; Li, B.; Zhang, W.; Lam, W.; Li, B.; Huang, F.; Si, L.; and Li, Y. 2022. Towards Generalizable and Robust Text-to-SQL Parsing. In Findings of EMNLP 2022.
  12. 12.Giordani, A.; and Moschitti, A. 2012. Automatic Generation and Reranking of SQL-derived Answers to NL Questions. In Proceedings of the Second International Conference on Trustworthy Eternal Systems via Evolving Software, Data and Knowledge, 59–76.
  13. 13.Guo, J.; Zhan, Z.; Gao, Y.; Xiao, Y.; Lou, J.; Liu, T.; and Zhang, D. 2019. Towards Complex Text-to-SQL in Cross-Domain Database with Intermediate Representation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28-August 2, 2019, Volume 1: Long Papers, 4524–4535.
  14. 14.Hochreiter, S.; and Schmidhuber, J. 1997. Long Short-Term Memory. Neural Comput., 1735–1780.
  15. 15.Hui, B.; Geng, R.; Wang, L.; Qin, B.; Li, Y.; Li, B.; Sun, J.; and Li, Y. 2022. S2SQL: Injecting Syntax to Question-Schema Interaction Graph Encoder for Text-to-SQL Parsers. In Findings of ACL 2022, 1254–1262.
  16. 16.Iyer, S.; Konstas, I.; Cheung, A.; Krishnamurthy, J.; and Zettlemoyer, L. 2017. Learning a Neural Semantic Parser from User Feedback. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, 963–973.
  17. 17.Krishnamurthy, J.; Dasigi, P.; and Gardner, M. 2017. Neural Semantic Parsing with Type Constraints for Semi-Structured Tables. In EMNLP 2017, 1516–1526.
  18. 18.Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed, A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In ACL 2020, 7871–7880.
  19. 19.Lin, T.; Goyal, P.; Girshick, R. B.; He, K.; and Dollár, P. 2017. Focal Loss for Dense Object Detection. In IEEE International Conference on Computer Vision, ICCV 2017, 2999–3007.
  20. 20.Lin, X. V.; Socher, R.; and Xiong, C. 2020. Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing. In Findings of EMNLP 2020, 4870–4888.
  21. 21.Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692.
  22. 22.Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019.
  23. 23.Qi, J.; Tang, J.; He, Z.; Wan, X.; Cheng, Y.; Zhou, C.; Wang, X.; Zhang, Q.; and Lin, Z. 2022. RASAT: Integrating Relational Structures into Pretrained Seq2Seq Model for Text-to-SQL. In EMNLP 2022.
  24. 24.Qin, B.; Wang, L.; Hui, B.; Li, B.; Wei, X.; Li, B.; Huang, F.; Si, L.; Yang, M.; and Li, Y. 2022. SUN: Exploring Intrinsic Uncertainties in Text-to-SQL Parsers. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, 5298–5308.
  25. 25.Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. J. Mach. Learn. Res., 140:1–140:67.
  26. 26.Rubin, O.; and Berant, J. 2021. SmBoP: Semi-autoregressive Bottom-up Semantic Parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, 311–324.
  27. 27.Schlichtkrull, M. S.; Kipf, T. N.; Bloem, P.; van den Berg, R.; Titov, I.; and Welling, M. 2018. Modeling Relational Data with Graph Convolutional Networks. In The Semantic Web - 15th International Conference, ESWC 2018, 593–607.
  28. 28.Scholak, T.; Schucher, N.; and Bahdanau, D. 2021. PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding from Language Models. In EMNLP 2021, 9895–9901.
  29. 29.Shaw, P.; Chang, M.; Pasupat, P.; and Toutanova, K. 2021. Compositional Generalization and Natural Language Variation: Can a Semantic Parsing Approach Handle Both? In ACL/IJCNLP 2021, 922–938.
  30. 30.Shaw, P.; Uszkoreit, J.; and Vaswani, A. 2018. Self-Attention with Relative Position Representations. In NAACL-HLT, 464–468.
  31. 31.Shazeer, N.; and Stern, M. 2018. Adafactor: Adaptive Learning Rates with Sublinear Memory Cost. In Proceedings of the 35th International Conference on Machine Learning, ICML 2018, 4603–4611.
  32. 32.Shi, P.; Ng, P.; Wang, Z.; Zhu, H.; Li, A. H.; Wang, J.; dos Santos, C. N.; and Xiang, B. 2021. Learning Contextual Representations for Semantic Parsing with Generation-Augmented Pre-Training. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, 13806–13814.
  33. 33.Suhr, A.; Chang, M.; Shaw, P.; and Lee, K. 2020. Exploring Unexplored Generalization Challenges for Cross-Database Semantic Parsing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, 8372–8388.
  34. 34.Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 5998–6008.
  35. 35.Wang, B.; Shin, R.; Liu, X.; Polozov, O.; and Richardson, M. 2020a. RAT-SQL: Relation-Aware Schema Encoding and Linking for Text-to-SQL Parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, 7567–7578.
  36. 36.Wang, C.; Tatwawadi, K.; Brockschmidt, M.; Huang, P.-S.; Mao, Y.; Polozov, O.; and Singh, R. 2018. Robust Text-to-SQL Generation with Execution-Guided Decoding. arXiv preprint arXiv:1807.03100.
  37. 37.Wang, K.; Shen, W.; Yang, Y.; Quan, X.; and Wang, R. 2020b. Relational Graph Attention Network for Aspect-based Sentiment Analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, 3229–3238.
  38. 38.Xu, P.; Kumar, D.; Yang, W.; Zi, W.; Tang, K.; Huang, C.; Cheung, J. C. K.; Prince, S. J. D.; and Cao, Y. 2021. Optimizing Deeper Transformers on Small Datasets. In ACL/IJCNLP 2021, 2089–2102.
  39. 39.Yin, P.; and Neubig, G. 2017. A Syntactic Neural Model for General-Purpose Code Generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, 440–450.
  40. 40.Yu, T.; Li, Z.; Zhang, Z.; Zhang, R.; and Radev, D. R. 2018a. TypeSQL: Knowledge-Based Type-Aware Neural Text-to-SQL Generation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, 588–594.
  41. 41.Yu, T.; Wu, C.; Lin, X. V.; Wang, B.; Tan, Y. C.; Yang, X.; Radev, D. R.; Socher, R.; and Xiong, C. 2021. GraPPa: Grammar-Augmented Pre-Training for Table Semantic Parsing. In 9th International Conference on Learning Representations, ICLR 2021.
  42. 42.Yu, T.; Yasunaga, M.; Yang, K.; Zhang, R.; Wang, D.; Li, Z.; and Radev, D. R. 2018b. SyntaxSQLNet: Syntax Tree Networks for Complex and Cross-DomainText-to-SQL Task. arXiv preprint arXiv:1810.05237.
  43. 43.Yu, T.; Zhang, R.; Yang, K.; Yasunaga, M.; Wang, D.; Li, Z.; Ma, J.; Li, I.; Yao, Q.; Roman, S.; Zhang, Z.; and Radev, D. R. 2018c. Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 3911–3921.
  44. 44.Zelle, J. M.; and Mooney, R. J. 1996. Learning to Parse Database Queries Using Inductive Logic Programming. In Proceedings of the Thirteenth National Conference on Artificial Intelligence and Eighth Innovative Applications of Artificial Intelligence Conference, AAAI 96, IAAI 96, 1050–1055.
  45. 45.Zhang, Y.; Warstadt, A.; Li, X.; and Bowman, S. R. 2021. When Do You Need Billions of Words of Pretraining Data? In ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, 1112–1125.
  46. 46.Zhong, R.; Yu, T.; and Klein, D. 2020. Semantic Evaluation for Text-to-SQL with Distilled Test Suites. In EMNLP 2020, 396–411.
  47. 47.Zhong, V.; Xiong, C.; and Socher, R. 2017. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning. arXiv preprint arXiv:1709.00103.

Citation

MLA
Li, H., et al. “RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL”. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 11, 2023, pp. 13067–75, https://doi.org/10.1609/AAAI.V37I11.26535.
APA
Li, H., Zhang, J., Li, C., & Chen, H. (2023). RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11), 13067–13075. https://doi.org/10.1609/AAAI.V37I11.26535
Chicago
Li, H., J. Zhang, C. Li, and H. Chen. 2023. “RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL”. Proceedings of the AAAI Conference on Artificial Intelligence 37 (11): 13067–75. https://doi.org/10.1609/AAAI.V37I11.26535.
Harvard
Li, H. et al. (2023) “RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL”, Proceedings of the AAAI Conference on Artificial Intelligence, 37(11), pp. 13067–13075. Available at: https://doi.org/10.1609/AAAI.V37I11.26535.
Vancouver
1. Li H, Zhang J, Li C, Chen H (2023) RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL. Proceedings of the AAAI Conference on Artificial Intelligence 37:13067–13075

BibTeX

@article{Li_2023, title={RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL}, volume={37}, ISSN={2159-5399}, url={http://dx.doi.org/10.1609/AAAI.V37I11.26535}, DOI={10.1609/aaai.v37i11.26535}, number={11}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, publisher={Association for the Advancement of Artificial Intelligence (AAAI)}, author={Li, Haoyang and Zhang, Jing and Li, Cuiping and Chen, Hong}, year={2023}, month=June, pages={13067–13075} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF