Conversations With a Computer - An Example of Network Programming in Relational Data Base Environment

M. Nalbach, K. Studzinski, S. WaligorskiCOLING 1982

TL;DR

This paper presents a novel approach to building conversational AI systems by representing conversation flow as networks stored in a relational database. This method allows for flexible, modular, and easily modifiable conversation logic, which can be compiled into efficient Fortran code for execution.

What Problem Does This Paper Address?

Developing efficient human-computer conversations requires significant information, including conversation scripts, language grammar and vocabulary, and semantic understanding. Managing and modifying this complex information is challenging. The authors aim to address this by proposing a structured and standardized way to represent and manage conversational logic, making it more accessible for development and debugging.

Methods

The core of the proposed method is to represent conversational logic as networks. These networks are composed of "atomic units," each representing an exchange between a human and a computer, along with accompanying actions. These networks can be recursive, drawing parallels to Augmented Transition Networks (ATNs) used in syntactic analysis. The key innovation is storing these networks as relations within a relational database. This allows for easy modification and management. For execution, these networks can be interpreted or compiled into Fortran code, offering a trade-off between flexibility and speed. The paper also mentions the possibility of using these networks for bootstrapping, where a created network can control the process of creating or modifying other networks.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Relational Database for Conversation Logic: Storing conversation flow as relations in a database provides a standardized and easily modifiable structure, simplifying development and debugging.
  • Network Representation of Dialogues: Viewing conversations as networks of atomic units and transitions offers a powerful way to model complex conversational flows.
  • Recursion and ATN Parallels: The ability to introduce recursion into these networks makes them conceptually similar to ATNs, enabling more sophisticated conversational patterns.
  • Interpreted vs. Compiled Networks: The paper highlights the flexibility of interpreting networks for conversation control and the performance benefits of compiling them into Fortran for complex syntactic grammars.
  • Bootstrapping Conversational AI: The concept of using existing networks to create or modify new ones suggests a self-improving or generative approach to conversational AI development.

Key Results

The paper demonstrates an implementation of these ideas using Fortran for the data base management programs. User access to the data base is conversational. The networks are stored as relations, and the paper provides an example of such a relation in Figure 1, illustrating states, arguments, conditions, actions, and transitions. The compiled networks achieved good speeds, even for complicated syntactic grammars. Conversation control networks are always interpreted, as speedup was not deemed necessary in those cases.

Limitations and Future Work

The authors acknowledge that their implementation was made in Fortran, despite its disadvantages, primarily for portability reasons. They suggest that Fortran is still easier to transfer to other computers or mini-software compared to other languages like LISP. The paper implies that further exploration of the bootstrapping mechanism for conversational AI development could be a future direction.

Summary updated

BibTeX
@inproceedings{nalbach1982conversations,
  author = {M. Nalbach and K. Studzinski and S. Waligorski},
  title = {{Conversations With a Computer - An Example of Network Programming in Relational Data Base Environment}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1982},
}

Loading related content...

Figure