Improving Response Quality in a Knowledge-Grounded Chat System Based on a Sequence-to-Sequence Neural Network
Sihyung Kim, Harksoo Kim, Oh‐Woog Kwon, Young-Gil Kim · 2019
A knowledge-grounded chat model based on a sequence-to-sequence neural network is proposed to perform more natural question answering. The previous chat model suffered from the same words being repeated during decoding. To reduce this word repetition, we propose a new decoding mechanism that considers the generation distribution of previously generated words. In experiments with question-answering sentences that are semi-automatically constructed, the proposed model outperformed a representative knowledge-grounded chat model, with a better accuracy of 4.35% p in finding answer phases. In addition, it demonstrated 1.63-2.05%p better results in all evaluation measures such as BLEU and ROUGE for evaluating the quality of responses and 2.83%p better results in evaluation measures such as distinct-1 for evaluating the diversity of responses.