Author: Shun Ishizaki, Electrotechnical Laboratory
Abstract
This paper presents a computational framework for generating natural Japanese text from language-independent conceptual representations. By utilizing Conceptual Dependency (CD) theory and Memory Organization Packets (MOPs), the system transforms abstract logic trees—originally parsed from Spanish news reports—into grammatically correct and stylistically appropriate Japanese. The research specifically addresses the structural and pragmatic challenges of the Japanese language, including its verb-final (SOV) word order, strict chronological narrative requirements, and the frequent use of ellipsis.
Introduction
In the early 1980s, machine translation research was often constrained by direct mapping techniques that struggled to bridge the gap between structurally disparate languages. This paper explores a more sophisticated alternative: generating text from a deep, conceptual level of representation. By focusing on the "meaning" of a story rather than its surface-level source text, the system developed at the Electrotechnical Laboratory demonstrates how abstract logic can be realized as natural-sounding Japanese.
The Challenge of Japanese Generation
Japanese presents several unique hurdles for automated generation systems that are not typically encountered in Romance or Germanic languages:
- Syntactic Structure: Unlike the Subject-Verb-Object (SVO) structure of Spanish or English, Japanese is strictly Verb-final (SOV).
- Temporal Logic: Japanese narrative structure is inherently chronological. If an event sequence involves a cause and an effect, the cause must almost universally precede the effect in the sentence structure to remain coherent.
- Case Markers: The distinction between particles such as wa (topic) and ga (subject) is subtle and context-dependent, making it difficult to automate.
- Ellipsis: Natural Japanese frequently employs "zero pronouns," where the subject is omitted if it can be inferred from the context—a behavior that contrasts sharply with the explicit subject requirements of Spanish.
Methodology
The system leverages two foundational pillars of 1980s Artificial Intelligence: Conceptual Dependency (CD) and Memory Organization Packets (MOPs). These frameworks allow the system to represent events as abstract graphs that are independent of any specific natural language.
Temporal Inference and Sorting
Before generating text, the system performs a "temporal inference" phase. It analyzes the conceptual graph to determine the real-world chronological order of events. Using causal links (e.g., "LEAD-TO" or "PRECONDITION") and MOPs—which act as pre-defined scripts for common scenarios like a "Police Capture"—the generator sorts the event nodes to ensure the resulting Japanese sentences follow a logical flow.
The Generation Pipeline
The realization of text follows a structured pipeline:
- GEN-SUBJECT: Identifies the primary actor. If the actor is unspecified in the conceptual representation, the system triggers a passive voice construction.
- GEN-VERB-MODIFIERS: This module places objects, locations, and temporal markers in the correct sequence preceding the verb.
- GEN-VERB: The final step selects the appropriate verb and applies the necessary inflections for tense and mood.
Pragmatic Heuristics
To ensure the output feels natural, the system implements specific rules for particles and ellipsis. For instance, it uses ga for subordinate clauses expressing reason or time, while reserving wa for the main clause. Furthermore, it performs ellipsis by deleting subjects in main clauses if they match the subject of the preceding subordinate clause, avoiding the repetitive and clunky feel of "translation-ese."
Experimental Results
The generator was evaluated using conceptual representations derived from 15 Spanish newspaper stories concerning terrorism. These inputs were generated by a parser developed by Lytinen and Schank (1982).
The system successfully produced coherent, multi-clause Japanese sentences. For example, when processing a complex node involving a "Mock Trial" and an "Execution," the generator correctly ordered the trial before the execution, inserted the causal conjunctive node ("because"), and placed the verb shokeishita ("executed") at the end of the sentence.
Discussion and Insights
The most significant finding of this research is that CD theory and MOPs—though developed within an English-centric linguistic context—are robust enough to serve as a foundation for Japanese generation. This suggests that the underlying logic of events is largely universal, even if the grammatical realization of those events varies wildly between languages.
However, the system has notable limitations. It does not currently handle "focus"—the rhetorical emphasis a speaker might place on specific parts of a sentence. Additionally, the reliance on hand-crafted MOPs limits the system's scalability; expanding the generator to handle general-purpose conversation or diverse news topics would require significant manual knowledge engineering.
Conclusion
Shun Ishizaki’s work demonstrates that high-quality Japanese generation is possible when the system is grounded in a deep understanding of event logic and temporal flow. By prioritizing the chronological and pragmatic rules of the target language over a literal translation of the source, the system achieves a level of naturalness that was pioneering for its time.
Summary updated