LGESQL: Line Graph Enhanced Text-to-SQL Model with Mixed Local and Non-Local Relations

Abstract

This work aims to tackle the challenging heterogeneous graph encoding problem in the text-to-SQL task. Previous methods are typically node-centric and merely utilize different weight matrices to parameterize edge types, which 1) ignore the rich semantics embedded in the topological structure of edges, and 2) fail to distinguish local and nonlocal relations for each node. To this end, we propose a Line Graph Enhanced Text-to-SQL (LGESQL) model to mine the underlying relational features without constructing metapaths. By virtue of the line graph, messages propagate more efficiently through not only connections between nodes, but also the topology of directed edges. Furthermore, both local and non-local relations are integrated distinctively during the graph iteration. We also design an auxiliary task called graph pruning to improve the discriminative capability of the encoder. Our framework achieves state-of-theart results (62.8% with GLOVE, 72.0% with ELECTRA) on the cross-domain text-to-SQL benchmark Spider at the time of writing.

TIE: Topological Information Enhanced Structural Reading Comprehension on Web Pages

“…On the other hand, Glass et al (2021) introduces row and column interactions into their models and determines the final answers based on the top-ranked relevant rows and columns. In addition, Text-to-SQL is another group of methods to tackle Table QA problems and has been widely studied recently (Yu et al, 2018;Bogin et al, 2019;Wang et al, 2020;Cao et al, 2021;Chen et al, 2021d,e;Hui et al, 2022). They use databases to store the source tables and translate natural language queries into Structured Query Language (SQL) to retrieve answers from the databases.…”

SADGA: Structure-Aware Dual Graph Aggregation Network for Text-to-SQL

“…ShadowGNN [8] presents a graph project neural network to abstract the representation of the question and schema. LGESQL [7] utilizes the line graph to update the edge features in the heterogeneous graph for Text-to-SQL, which further considers both local and non-local, dynamic and static edge features. Differently, our SADGA not only adapts a unified dual graph framework for both the question and database schema, but also devises a structure-aware graph aggregation mechanism to sufficiently utilize the global and local structure information across the dual graph on the question-schema linking.…”

“…However, it is hard for database developers to build the Text-to-SQL model for each specific database from scratch because of the high annotation cost. Therefore, cross-domain Text-to-SQL, aiming to generalize the trained model to the unseen database schema, is proposed as a more promising solution [13,4,5,29,8,24,21,7]. The core issue of cross-domain Text-to-SQL lies in building the linking between the natural language question and database schema, well-known as the question-schema linking problem [13,29,21,19,37].…”

“…There are two categories of efforts to solve the aforementioned question-schema linking problem -matching-based method [13] and learning-based method [29,21,19,7]. IRNet [13] is a typical matching-based method, which uses a simple yet effective string matching approach to link question words and tables/columns in the schema.…”

“…However, both of the above two categories of methods still suffer from the problem of insufficient generalization ability. There are two main reasons for the above problem: (1) the structure gap between the encoding process of the question and database schema: as shown in Figure 1, most of the existing approaches treat the question as a sequence and learn the representation of the question by sequential encoders [13,4,5] or transformers [29,30,21], while the database schema is the structured data whose representation is learned based on graph encoders [7,4,5] or transformers with predefined relations [29,30]. Such the structure gap leads to difficulty in adapting the trained model to the unseen schema.…”