End-to-end Knowledge Graph-enabled Multi-round Dialogue Model Based on Transformer

As an important branch of artificial intelligence, natural language processing can complete many complex tasks for humans. Among them, intelligent dialogue is to allow users to communicate with a dialogue model and let it complete specific tasks. In order to better propose an end-to-end dialogue model, this work adopts a structural model based on Transformer and a pointer generation network, which seeks an appropriate alternative through the pointer generation network. Moreover, the end-to-end model has the ability to directly copy unknown words from the knowledge base. It makes innovations in many aspects such as the input part of the encoder, the encoding of position information, the ONE-HOT matrix, and the task layering of the decoding layer. The study compares the proposed model with benchmark models including Seq2Seq, Seq2Seq+Attn, and Mem2Seq through experiments under two datasets, i.e.,bAbI and In-Car. On the bAbI dataset, the proposed model significantly outperforms Seq2Seq and Seq2Seq+Attn models in terms of word accuracy, dialogue accuracy, and BLUE score indicators. On the In-Car dataset, it is found that the proposed model is better than the Mem2Seq model in handling long corpus.