Understanding of Japanese in an Interactive Programming System

Kenji Sugiyama, Masayuki Kameda, Kouji AkiyamaCOLING 1984

TL;DR

This paper introduces KIPS, an interactive programming system that understands Japanese natural language to generate business application programs. It models both the program and user statements as dynamic objects, enabling a more flexible and robust understanding of Japanese linguistic nuances.

What Problem Does This Paper Address?

Developing natural language programming systems (NLPS) that can understand user specifications is a significant challenge. Traditional NLPS often rely on separate models for programs and user statements, leading to rigidity. Furthermore, understanding natural language input, especially languages like Japanese with unique grammatical structures (e.g., predicate at the end of a sentence, flexible word order), poses specific hurdles. This paper addresses the need for a more adaptable NLPS capable of handling the complexities of Japanese input for program generation.

Methods

KIPS employs an "object-centered" approach, modeling both the program under discussion and the user's statements as organizations of dynamic objects, akin to object-oriented programming.

  • Statement Model: This model acts as an intermediary representation of the user's input. It's built by parsing Japanese sentences and categorizing semantic components into operations, parameters, and item-conditions. This model is then used to check for consistency and augment information before being assimilated into the program model.
  • Program Model: This model represents the program being built. It's divided into a "model of files" and a "model of processes." These models are populated with instances of objects that possess both static relations (like class hierarchies) and dynamic behaviors (how to instantiate themselves).
  • TEC (Transformational-Expressive Grammar): A specialized parsing system is used for sentence analysis. TEC is designed to handle Japanese sentence structures by using look-back buffers to store noun phrases that appear early in the sentence, allowing them to be associated with the predicate that appears at the end.
  • Message Passing: Objects within the statement and program models communicate via message passing, facilitating dynamic updates and modifications to the program model based on user input.
  • Subgoaling: To handle the fragmented and sometimes arbitrary nature of user specifications, KIPS uses subgoaling methods. These methods define different ways to acquire specific pieces of program information, allowing for flexible input.

Datasets

The paper mentions a corpus of 60 program descriptions, comprising about 300 sentences, sampled from a Hyper COBOL user's manual. This corpus was used to build the program model. No specific datasets for training or evaluation are detailed beyond this.

Key Insights and Ideas

  • Object-Oriented Modeling for NLPS: Representing both program logic and user input as dynamic objects offers a more flexible and extensible architecture for NLPS compared to traditional, separate models.
  • Handling Japanese Linguistic Nuances: The TEC parser, with its look-back buffers, is a crucial innovation for effectively processing Japanese sentences where the predicate appears at the end and word order can be variable.
  • Statement Model as a Crucial Intermediary: Using a statement model as a buffer between raw input and the program model allows for error checking, consistency verification, and augmentation of information, leading to more reliable program generation.
  • Subgoaling for Flexible Specification: The subgoaling mechanism allows the system to gracefully handle incomplete or out-of-order user specifications, making the interaction more natural.

Key Results

The paper states that a prototype system is operational and has successfully produced several dozen programs from Japanese specifications. However, specific quantitative results (e.g., accuracy metrics, success rates) are not detailed. The qualitative finding is that the "object-centered" approach is effective in making the system flexible.

Limitations and Future Work

  • Limitations: The paper acknowledges that the current model depends on the kind of operations and their sequence. The discourse model for acquiring information is also noted as being dependent on these factors. The system currently focuses on understanding process specifications and may not handle other types of Japanese sentences.
  • Future Work: The authors plan to extend their research to understanding Japanese sentences that contain information beyond just process specifications.

Summary updated

BibTeX
@inproceedings{sugiyama1984understanding,
  author = {Kenji Sugiyama and Masayuki Kameda and Kouji Akiyama},
  title = {{Understanding of Japanese in an Interactive Programming System}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure