Scholarlink

<p>Semantic parsing (SP) aims to convert natural language into SQL, which is a complex task. Traditional SP research typically depends on supervised learning, which requires extensive, accurately labeled databases. This study introduces a semi-supervised generative adversarial network (SS-GAN) model, enhanced by a spatial attention-based transductive long-short-term memory (TLSTM), to improve classifier effectiveness. The semi-supervised aspect of SS-GAN is essential in situations where labeled data is limited or unevenly distributed. The TLSTM model applies transductive learning methods, focusing on samples close to the test data to refine the adaptation process and outperform traditional LSTM models. Our methodology includes a regularization tactic to mitigate mode collapse and stabilize SS-GAN training. We guide the generator with feature vectors from the discriminator to generate precise outputs. Furthermore, we integrate a reconstruction loss into SS-GAN’s loss function, forcing the generator to mimic the discriminator’s features, which aligns the outputs more closely with real data configurations. This elevates the effectiveness of SS-GAN and ensures high-quality output. We evaluate the recommended framework utilizing the WikiSQL and SparC databases. Our experimental outcomes highlight the model's superiority over existing models, achieving execution accuracies of 0.902 and 0.844 on the WikiSQL and SparC databases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Spatial Attention-Based Transductive Long Short-Term Memory for Semantic Parsing Trained with a Semi-Supervised Method

  • 期刊论文

  • 发表日期：

    2025年2月14日

查看全文
  • Shichao Chen

摘要

Semantic parsing (SP) aims to convert natural language into SQL, which is a complex task. Traditional SP research typically depends on supervised learning, which requires extensive, accurately labeled databases. This study introduces a semi-supervised generative adversarial network (SS-GAN) model, enhanced by a spatial attention-based transductive long-short-term memory (TLSTM), to improve classifier effectiveness. The semi-supervised aspect of SS-GAN is essential in situations where labeled data is limited or unevenly distributed. The TLSTM model applies transductive learning methods, focusing on samples close to the test data to refine the adaptation process and outperform traditional LSTM models. Our methodology includes a regularization tactic to mitigate mode collapse and stabilize SS-GAN training. We guide the generator with feature vectors from the discriminator to generate precise outputs. Furthermore, we integrate a reconstruction loss into SS-GAN’s loss function, forcing the generator to mimic the discriminator’s features, which aligns the outputs more closely with real data configurations. This elevates the effectiveness of SS-GAN and ensures high-quality output. We evaluate the recommended framework utilizing the WikiSQL and SparC databases. Our experimental outcomes highlight the model's superiority over existing models, achieving execution accuracies of 0.902 and 0.844 on the WikiSQL and SparC databases.