Multi-Turn Response Selection Using Business Sequential Relations in Traffic Field

BERT-based models play an essential role and achieve significant results in many tasks of natural language processing (NLP), including dialog tasks. However, there is a limitation of BERT to handle long dialogs. The paper studies how pre-trained model handles long conversations as the input in the multi-turn response selection task. Given the lack of data with long conversations and current dialog tasks are based on engineering applications, it is necessary to collect large dataset with long text sequences. Meanwhile, collecting dialog dataset is a time-consuming work, practical industrial settings usually need more accurate. In this paper, we cite the Chinese Multi-Intention Dialogue (The CMID-Transportation) dataset of transportation customer service, and change it to adapt to response selection task in the dialog systems. To this end, we propose a business-level strategy and use truncation methods to address this problem on the corpus. The experimental results on this corpus show that our proposed approach is fit for the BERT-based model and brings better performance.