An approach to intelligent Chinese dialogue system

Fang Chen, Baozong YuanInterspeech 2000 · DOI

TL;DR

This paper introduces the Intelligent Chinese Dialogue System (ICDS), a novel approach to natural human-computer interaction using Chinese speech. ICDS integrates dialogue management, task processing, natural language interface, user modeling, and a knowledge base, enabling users to query and retrieve information in natural speech. Its key innovation lies in its domain-independent architecture and its ability to learn task-specific and linguistic knowledge, allowing for adaptation to new domains.

What Problem Does This Paper Address?

The paper addresses the challenge of creating natural and intuitive human-computer interaction systems, specifically for the Chinese language. Traditional systems often require users to interact through rigid command structures or specialized interfaces. The authors aim to bridge this gap by developing a system that can understand and respond to users in natural spoken Chinese, making computer interaction more accessible and user-friendly. This involves tackling the complexities of Chinese speech recognition, natural language understanding, dialogue management, and speech synthesis within a unified framework.

Methods

The ICDS is built upon a modular architecture comprising five key components:

  • Dialogue Processor: This module acts as the central coordinator, managing the flow of conversation. It translates task demands into semantic forms, plans dialogue strategies, and updates the user model based on interactions.
  • Task Processor: This component is responsible for managing domain-specific knowledge. It analyzes user tasks, decomposes them into sub-tasks, and queries the knowledge base to retrieve relevant information. It represents tasks as trees, where each node is a sub-task.
  • Natural Language Interface: This module handles the interaction with the user. It includes:
  • Speech Recognition and Understanding: Utilizes HMM-based real-time Chinese speech recognition (achieving 96% accuracy in domain-specific questions) and a language understanding module with grammar and semantic analysis to convert spoken input into a machine-understandable format.
  • Text Generation and Speech Synthesis: Employs a bi-directional grammar to generate diverse natural Chinese sentences (2-15 variations per question) and synthesizes them into speech using a pre-built speech database.
  • Learning Ability: The system can learn new sentence patterns and vocabulary through user-provided editing forms, enhancing its adaptability.
  • User Model: This module dynamically describes the user's knowledge level, interests, and intentions, allowing the system to tailor its responses and suggestions for more effective dialogue.
  • Knowledge Base: Stores domain-specific information, including graphical representations (maps), concepts, and rules, which can be learned and edited by users.

A novel aspect is the system's ability to learn knowledge for both task domains and linguistic domains, making it adaptable to various applications without being hardcoded for a specific one.

Datasets

Not explicitly discussed in the paper. The system is designed to learn knowledge from user input and pre-defined tables, suggesting it can be trained on domain-specific data provided by the user.

Key Insights and Ideas

  • Domain Independence through Learning: The ICDS's ability to learn task-specific and linguistic knowledge is a significant contribution, allowing it to be adapted to new domains (e.g., museum guidance, city tours) without requiring a complete system redesign.
  • Flexible Knowledge Representation: The use of tables for learning task domain knowledge, including graphical representations of maps and scene features, provides a user-friendly way to populate the knowledge base.
  • Diverse Response Generation: The system's capability to generate multiple (2-15) correct answers for a single query enhances the naturalness and richness of the interaction.
  • User Model for Enhanced Dialogue: Incorporating a user model allows the system to personalize the dialogue, making it more efficient and responsive to individual user needs and knowledge levels.

Key Results

The paper reports a speech recognition accuracy of 96% for domain-related questions. The system can generate 2-15 different correct answers for each query within a domain. The ICDS has been demonstrated with a "Park Tour Guide" application.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the system is designed to be domain-independent, it is currently structured for domain-specific applications. The accuracy of speech recognition might be lower for out-of-domain queries.
  • Future Work: The authors suggest that by improving its learning capability, the ICDS can be applied to more domains, further enhancing its role as a natural human-machine interactive system. They also imply that further research into more robust language understanding and generation techniques would be beneficial.

Summary updated

BibTeX
@inproceedings{chen2000approach,
  author = {Fang Chen and Baozong Yuan},
  title = {{An approach to intelligent Chinese dialogue system}},
  booktitle = {Interspeech},
  year = {2000},
  doi = {10.21437/icslp.2000-339},
}
Figure