Introduction
Mastering the Japanese language presents unique challenges for second language learners, extending beyond basic syntax into the realm of social and situational pragmatics. A primary difficulty lies in "empathy relationships"—the social hierarchy and perspective between the speaker, giver, and receiver—which dictate the correct choice of verbs such as ageru, kureru, and morau. Traditional Natural Language Processing (NLP) tools often overlook these "socially" incorrect sentences if they are grammatically sound in isolation. Researchers at the Science University of Tokyo have addressed this by developing a diagnostic parser based on Lexicalized Tree Adjoining Grammar (LTAG) to identify both grammatical and situational errors in student compositions.
Methodology: Lexicalized Tree Adjoining Grammar (LTAG)
The system utilizes LTAG because its lexicalized nature allows grammatical rules to be anchored directly to specific words, facilitating precise error localization. In this framework, the grammar consists of a set of elementary trees: Initial Trees, representing basic predicate-argument structures, and Auxiliary Trees, representing recursive structures like modifiers.
Japanese phrases are categorized into Yougen (inflected words like verbs/adjectives) and Taigen (uninflected words like nouns). To handle the flexible word order of Japanese, the parser employs a stack-based mechanism. Candidate modifier phrases are held in stack memory until they can be unified with a head word to form a "Saturated Initial Tree" (SIT).
Situational Constraints and Empathy
A key innovation is the embedding of "situational constraints" within the feature structures of lexical items. Drawing on Susumu Kuno’s linguistic theories, the system tracks the "camera angle" of the speaker. Social distance is formalized as mathematical inequalities (e.g., $E(nurse|locutor) > E(son|locutor)$), allowing the parser to programmatically detect when a student's choice of verb implies an incorrect social perspective.
The system also addresses the complexity of the particle wa, distinguishing between its "thematic" and "discriminatory" (contrastive) uses based on the saturation of the verb's case slots.
Diagnostic Capabilities and Results
When a sentence fails to parse, the system applies diagnostic rules rather than simply returning an error. It can identify "modification crossover" (semantic confusion due to word order), "wrong inflection," and "false modification."
The prototype demonstrates the ability to detect multiple overlapping errors. For instance, it can flag the use of ageru (to give) when the context requires kureru (to give to me) while simultaneously identifying a misplaced possessive marker or an incorrect case marker like ga.
Conclusion and Future Work
The LTAG-based diagnostic parser offers a robust framework for Computer-Assisted Language Learning (CALL) systems by bridging the gap between syntax and social logic. However, scaling the system requires a more comprehensive ontology for semantic relationships and more efficient methods for manual lexical preparation. Future developments will focus on handling anaphora and integrating electronic dictionaries to provide more detailed feedback to learners.
Summary updated