Abstract:

Associating form widgets with an urban knowledge system to transform form filling based data into a standardized urban knowledge graph automatically presents a novel solution to the challenge of large-scale data element production. However， the form system integrates numerous widgets named using domain-specific terms such as entities， relationships， and their attributes， making it difficult for users to find the desired widget quickly， thereby giving rise to the task of form widget recommendation. To tackle three key challenges： significant variability in application scenarios， complex contextual dependencies in configuration， and sparse form configuration data， a Tree-structured Long Short-Term Memory （Tree-LSTM） network based Recommendation model for Form Widget （TRFW） was proposed. Firstly， a scenario multi-classification model was trained on rich and general form configuration data to learn dependencies between form texts and scenario features. Then， a network structure of Tree-LSTM was employed to extract structural and thematic features from form contexts， and the user configuration intent features were constructed on this basis. Concurrently， an AutoEncoder （AE）-based widget naming encoder was trained to establish loosely-coupled connections between semantically similar widget names， thereby enhancing the model’s ability to recommend semantically relevant widgets while maintaining robustness for newly added widgets. Experimental results on a publicly sourced dataset demonstrate that the proposed model has superior performance compared to all baseline models， achieving a HitRatio@1 of 84.62% and a HitRatio@10 of 97.31%， with a minimum improvement of 87.0% in HitRatio@1 over unsupervised methods， while outperforming graph convolution-based recommendation methods by at least 32.4% and 12.4% in HitRatio@1 and HitRatio@10， respectively. It can be seen that the proposed model can recommend widget lists that align with user intent effectively， thereby reducing user interaction costs during widget selection significantly.

Key words: intent-awareness, Tree-structured Long Short-Term Memory (Tree-LSTM), recommendation model, e-form, urban knowledge system

摘要：

通过将表单中的控件与城市知识体系相关联，基于表单填报的数据能自动成为标准规范的城市知识图谱，这是解决数据要素规模化生产难题的新思路。然而，表单系统中内置的众多的以实体、关系及其属性的业务术语命名的控件使得用户难以快速找到所需控件，进而催生了表单控件推荐这一全新任务。针对表单应用场景差异大、配置上下文依赖复杂和表单配置数据稀疏这3个挑战，提出一种基于树形长短期记忆（Tree-LSTM）网络的表单控件推荐模型（TRFW）。首先，从丰富的通用表单配置数据中训练场景多分类模型，学习表单文本与场景特征之间的依赖关系；其次，使用基于Tree-LSTM的网络结构，捕捉表单上下文中的结构与主题特征，并基于此构造用户的配置意图特征；同时，训练基于自编码器（AE）的控件命名编码器，建立相似控件命名之间的松耦合连接关系，使得模型能够在推荐相似语义控件的同时，对新增控件具有良好的鲁棒性。在公开数据构建的数据集上的实验结果表明，所提模型与所有基准模型相比具有最优性能，HitRatio@1为84.62%，HitRatio@10为97.31%，与无监督的推荐方法相比，HitRatio@1至少提升了87.0%，与基于图卷积的推荐方法相比，HitRatio@1和HitRatio@10分别至少提升了32.4%与12.4%。可见，所提模型能够有效推荐出符合用户意图的控件列表，从而降低用户在选择控件时的交互成本。

关键词: 意图感知, 树形长短期记忆网络, 推荐模型, 电子表单, 城市知识体系