XL-HQL: A HQL query generation method via XLNet and column attention

Abstract

Context:

Object-relational mapping (ORM) tools, like Hibernate, are widely used to facilitate the development of database applications by bridging the gap between object-oriented programming (OOP) and relational database management systems (DBMS). These ORM tools simplify the process of mapping OOP objects to relational tables, addressing issues of data inconsistency and performance. However, they also introduce the need to write queries in specific languages, such as Hibernate Query Language (HQL), to manage data interactions within the database.

Objective:

These query languages can be difficult to write and error-prone due to the complexities of accurately mapping object models to relational schema with intricate relationships and inheritance hierarchies. To mitigate this issue, a recent study introduced the task of automated HQL query generation, i.e., automatically generating HQL from program context (target method’s signature, properties, and optional method comments and call context). However, the existing solution, HQLgen, has shown limited performance, with an accuracy of 34.52%.

Method:

In this paper, we propose a novel HQL query generation approach named XL-HQL. XL-HQL aims to address two main challenges in HQL query generation: limited context information and large search space. Specifically, XL-HQL contains a pre-trained model-based encoder, rules defined to reduce search space, and a column-attention-enabled decoder, which is shown to be effective in SQL generation approaches.

Result:

To evaluate the effectiveness of XL-HQL, we designed and conducted experiments on an existing HQL query generation benchmark, which contains 24,118 HQL queries extracted from 3,481 open-source projects. The experimental results show that our approach achieves 66.93% and 64.47% accuracy on mixed and cross-project datasets, respectively, nearly doubling the performance of the state-of-the-art (SOTA) baseline.

Conclusions:

The application of pre-trained models that are suitable for handling long sequences for the HQL query generation task shows great potential. Moreover, the defined rules based on OOP knowledge are effective for reducing search space and improving the performance of the task.