CC BY 4.0 （除特别声明或转载文章外）
如果这篇博客帮助到你，可以请我喝一杯咖啡~
XDAI: A Tuning-free Framework for Exploiting Pre-trained
Language Models in Knowledge Grounded Dialogue Generation
Abstract
大规模预训练语言模型 (PLM) 已显示出前景在各种下游任务上取得进展，其中对话是最关心的问题之一。然而，仍然存在挑战.供个人开发人员创建以知识为基础的对话由于收集成本昂贵，因此在如此大的模型上进行系统支持系统的知识资源为任务调整这些大型模型。为了解决这些障碍，我们提出 XDAI，一个基于知识的对话系统，配备即时感知的免调优 PLM 开发,并由现成的开放域外部知识支持,资源加上易于更改的特定领域机制。借助 XDAI，开发人员无需任何微调即可利用 PLM,快速创建开放域对话系统的成本以及轻松定制自己的特定领域系统。广泛的实验包括人工评估、图灵测试和在线评估展示了竞争性能XDAI 与最先进的通用 PLM 和特定的用于对话的 PLM。 XDAI 试点研究利用PLM 并取得了有趣的发现，这些发现可能会启发其他基于 PLM 的应用程序的未来研究。
INTRODUCTION
main challenge
- Obstracles to High-quality Data Curation
- Trade-off between Effectiveness and Efficiency
XDAI unique features:
- Quick Start
- Efficient Inference
- Customized Deployment
- Incremental Modification
Inpact & Beneficial Groups
- developers with limited resources
- practitioners from other fields
- newcomers to the machine learning domanin to easily and use PLMs to accomplish their creative ideas.
PRELIMINARAIES
Problem Formulation
- Dialogue History
- External Knowledge Pool
- Knowledge-grounded Dialogue Generation task
Background Techniques
Pre-trained Model Exploitation approachers
- prompt searching
- controlled generation
components of building XDAI’s basic
- open-domain dialogue service (XDAI-open)
- domain-specific dialogue service (XDAI-domain)
XDAI
Overview Framework
- offline knowledge curation system
- online dialogue generation system Presented Work
offline system:knowledge curation
- system provides the functions of data collection and resource integration
- offer an optional concept expansion function for discovering abundant domain-specific knowledge pieces from a few seeds for XDAI-domian
- data collection:Xlore2
- Resource Integration: Description-formatted and QA-formatted
- concept expansion: performs candidate extraction for recalling the relevant concepts as much as possible and then concept ranking for selecting the most relevant ones
Framework
Online system :dialogue Generation
- dialogue history selection
- dialogue knowledge injection
- background knowledge addition
Availability
- SessionManager
- ChatbotAgent
Experimental Setting
Human Evaluation Protocol. gap
Evaluation Metrics
1) Conherence 2) Informativeness 3) Inconsistency 4) Hallucination 5) Engagingness
Result
maintains a competitive performance,especially leading in several metrics under the Travel topic.
1) still maintains an advantage in these dimensions 2) XDAI still maintains a good performance with its automatically explored knowledge. 3) XDAI’s performance in Domain remains relatively competitive.
Online Evaluation
- User Involvement
- Case Study
Conclusion
In this paper,we present a tuning-free framework, XDAI, for exploiting language models in knowledge-grounded dialogue generation.