Deep Cascade Multi-Task Learning for Slot Filling in Online Shopping Assistant

• DOI: 10.1609/aaai.v33i01.33016465
• Publication Date: 2019-08-27T07:31:48Z
• Authors: Yu Gong; Xusheng Luo; Yu Zhu; Wenwu Ou; Zhao Li; Muhua Zhu; Kenny Q. Zhu; Lu Duan; Xi Chen
• Keywords: Benchmark (surveying); Baseline (sea); Natural language understanding; Sequence labeling; Named Entity Recognition

Slot filling is a critical task in natural language understanding (NLU) for dialog systems. State-of-the-art approaches treat it as sequence labeling problem and adopt such models BiLSTM-CRF. While these work relatively well on standard benchmark datasets, they face challenges the context of E-commerce where slot labels are more informative carry richer expressions. In this work, inspired by unique structure knowledge base, we propose novel multi-task model with cascade residual connections, which jointly learns segment tagging, named entity tagging filling. Experiments show effectiveness proposed structures. Our has 14.6% advantage F1 score over strong baseline methods new Chinese shopping assistant dataset, while achieving competitive accuracies dataset. Furthermore, online test deployed dominant platform shows 130% improvement accuracy user utterances. already gone into production platform.