Microsoft, Columbia University | Godel: large scale pre training of goal oriented dialogue

title ： Microsoft 、 Columbia University |GODEL: Large-Scale Pre-Training for Goal-Directed Dialog（GODEL： Large scale pre training of goal oriented dialogue ）

author ：Baolin Peng, Michel Galley, Pengcheng He,  Jianfeng Gao etc.

brief introduction ： This paper introduces a large pre training language model for dialogue . In contrast, earlier models are used , for example DialoGPT,GODEL The new phase of pre training with grounding aims to better support GODEL Adapt to a wide range of downstream dialogue tasks that require external information to the current dialogue （ for example , Database or file ） To produce a good response . Experiments on a range of benchmarks include task oriented dialogues 、 Conversational QA And rooted open domain conversations show GODEL Better than the most advanced technology in the small sample fine-tuning setting of the pre trained dialogue model , In person and automatic evaluation . A new characteristic evaluation method of the author is to introduce the concept of utility response to evaluate the usefulness （ External evaluation ） In addition, their communicative characteristics （ Internal evaluation ）. The authors demonstrate that external evaluation provides improved annotator agreement and relevance to automated metrics .

The code download ：https://github.com/Microsoft/GODEL

Paper download ：https://arxiv.org/pdf/2206.11309v1.pdf