Introduction
In Natural Language Generation (NLG), the primary goal is to transform abstract concepts into coherent linguistic strings. While early systems focused heavily on "ideational meaning"—the factual content of who did what to whom—languages like Japanese necessitate a sophisticated handling of "interpersonal meaning." Politeness in Japanese is not merely an additive feature; it often dictates a complete structural reorganization of the clause, involving humble versus respectful verb forms, causative constructions, and specific honorific prefixes.
The Challenge of Non-Local Mapping
The technical hurdle in modeling Japanese politeness lies in its "non-local" nature. Traditional frameworks often rely on a 1-to-1 mapping where a single semantic choice (e.g., plurality) triggers a single grammatical feature. In contrast, a single social decision in Japanese—such as the intent to express humility—simultaneously affects multiple disparate parts of the grammar. This paper addresses the mismatch between high-level social intent and the resulting complex grammatical structures.
Extended Inquiry Semantics
To bridge this gap, the author extends the "inquiry semantics" framework used in the Nigel computational grammar, which is rooted in Systemic Functional Grammar (SFG). The core of this system consists of "inquiries" (questions posed to a knowledge base) and "choosers" (decision trees that select grammatical features based on inquiry responses).
The innovation presented is Extended Inquiry Semantics. In this model, a "politeness chooser" acts as a high-level manager. Rather than selecting a single local feature, it traverses a complex decision tree based on social context—such as social distance and officiality—and pre-selects a "bundle" of features across various grammatical systems.
As shown in the decision tree above, the system evaluates whether an action is independent of the audience and whether specific lexical verbs for humility exist before arriving at a set of grammatical requirements.
Implementation and Results
The system was tested using a corpus of telephone conversations regarding hotel reservations and hand-crafted examples within the Kyoto University Japanese grammar project. The primary result is a functional Japanese clause generator capable of producing socially nuanced text.
The "politeness chooser," comprising approximately 15 decision points, successfully constrained the grammar to produce varying levels of respect. For instance, it could generate complex strings like “kyou hassou sasete-itadaki-tai to omoimasu” ("I think I want to humbly receive from you your allowing me to send it today"), coordinating causality, favor-receiving, and humility systems.
The trace above illustrates the system's logic in real-time, showing how environmental responses to social inquiries lead to the pre-selection of multiple features like WISHFULNESS, FAVOURS, and MODIFIED-PROCESS.
Conclusion and Future Work
This research demonstrates that the interpersonal dimension of language is as systematic as the ideational dimension and can be modeled computationally. By defining the specific inquiries the grammar must make, the author identifies the social information a knowledge base must contain to support natural-sounding Japanese.
While successful at the clause level, future work is required to extend this "non-local" control to discourse-level planning. Additionally, the author suggests that this mechanism for handling interpersonal nuances may be equally necessary for refining NLG systems in other languages, including English.
Summary updated