Natural Language To SQL(NL2SQL)任务的目标是将自然语言查询转化为结构化查询语言.现有的大多数模型所使用的方法是将NL2SQL任务分解为多个子任务,为每个子任务构建一个专用的全连接神经网络解码器.这些方法存在一些问题,如模型设计与模型结构较为简单,在学习不同子任务之间的依赖关系的能力有限.为了解决这些问题,将多通道并行LSTM模型引入到NL2SQL任务中,并采用稀疏连接层联合不同的子任务解码器,提升神经网络表现能力和计算资源的使用效率.在WikiSQL数据集上的评估结果表明,与基线模型相比,文中提出的模型计算精度较好.
Research on NL2SQL based on sparse connection and multichannel LSTM
The goal of the Natural Language to SQL(NL2SQL)task is to tronsform natural language query statements to equivalent structured query languages.Most existing models use the method of breaking down NL2SQL tasks into multiple sub-tasks and build a dedicated fully connected neural network decoder for each sub-task.These methods have some problems,such as simple model design and model structure,limited a-bility to learn dependencies between different sub-tasks.To solve these problems,this paper creatively in-troduces multichannel parallel LSTM model into NL2SQL tasks,and uses sparse connection layer to com-bine different sub-task decoders to increase the performance of neural networks and the efficiency of compu-ting resources.The evaluation results on the WikiSQL data set show that the proposed model is more accu-rate than the baseline model.
Natural Language To SQLnatural language interfacepre-trained modelmulti-channel LSTMsparse connection
NETL
NSTL
万方数据
