Automatic Generation of Website-Based Multi-Turn Question-Answering Dialog Systems

Authors: Stefan Hillmann, Philine Görzig, Sebastian Möller

Abstract:

We present an overview of our chatbot system which uses a two-step NLU approach as well as a LaBSE-based multilingual NLU. We show, how can generate most parts of the chatbot in the university domain automatically from textual information on web pages. The evaluation of the 2-step NLU shows that the manually edited NLU (F1 ~ 0.8 outperforms the automatically trained (F1 ~ 0.42), but the latter needs nearly no manual efforts for training. Also, the multilingual NLU provides good performance (F1 ~ 0.71), taking into account that it needs no additional training material.