面向数据库自然语言查询的语义理解研究

近年来发展起来的基于自然语言对话技术的数据库查询方法为用户通过人机对话形式进行数据分析提供了可能。语义理解是其中的核心技术，其作用是将自然语言查询转换为结构化查询语言SQL语句。由于SQL表达形式及数据库结构的复杂多变，当前该语义理解任务面临着三个方面的挑战：标注数据难度大、效率低、成本高；准确的语义理解需要深度利用复杂多变的数据库结构知识；半监督和无监督范式的算法研究不足，缺乏提升训练数据稀疏情况下模型鲁棒性的有效机制。为此，在数据标注方面，本项目拟构建可视化标注平台，提出基于主动学习的数据标注框架，提升标注效率、降低标注难度；在模型结构方面，提出针对异构图的高阶图注意力神经网络，提升模型对异构信息的编码能力；在优化算法方面，提出基于对偶学习的闭环优化框架，提高模型在训练数据稀疏场景下的鲁棒性。通过上述系列研究，本项目力求为构建准确度高、鲁棒性好的语义理解模型提出完整的解决方案。

In recent years, the database query method based on natural language dialogue technology has been developed, which makes it possible for users to analyze data using natural language. Here semantic parsing, i.e. Text2SQL, plays a core role, and its role is to convert natural language query into the structured query language SQL. Due to the complexity and changeability of SQL expression form and database structure, the semantic parsing task is faced with three challenges: 1) Labelling large-scale training data is difficult, low efficiency, and high cost. 2) How to use the complex database structure knowledge to assist natural language understanding is a question. 3) The research of semi-supervised and unsupervised learning algorisms is insufficient, and there is no effective mechanism to improve the model robustness under sparse training data. Therefore, 1) for data annotation, this project plans to build a web-based annotation platform and propose a data annotation framework based on an active learning framework to improve annotation efficiency and reduce annotation difficulty. 2) For model structure, a high-order graph attention neural network for heterogeneous graphs is proposed to improve the ability to encode heterogeneous information. 3) For optimization algorithms, a closed-loop optimization framework based on dual learning is proposed to improve the robustness of the model when training data is sparse. Through the above series of studies, this project tries to put forward a complete solution for constructing a semantic parsing model with high accuracy and good robustness.