Word-Constrained Response Generation for International Chinese Language Education based on Decoder Backward Attention.

Dialogue systems are a valuable technology in the field of natural language processing to improve work, learning, and daily life. Currently, dialogue systems are employed as an educational technology for mentoring, evaluation, and personalized learning. To make dialogue teaching achieve the purpose of training vocabulary at the primary level of international Chinese learning education, we first collect the entire dialogue corpus from textbooks to create the dataset, and then we propose a dialogue response generation model, Seq2BF-Attention, containing a specific word based on the sequence to backward and forward sequences framework by adding an attention to enhance the modeling of dialogue posts. We also provide two decoder connection strategies, backward hidden connection and backward attention, to handle the problems of not sharing parameters and incoherent generation separately. It has been experimentally proven that our suggested models perform well in both the ICLE and Weibo datasets across all metrics.