Introduction
In the early 1980s, natural language processing (NLP) for Japanese faced significant hurdles when moving beyond "toy" sentences to real-world technical text. Scientific and engineering articles are characterized by frequent ellipses (omitted subjects or objects), complex modalities (layers of intent or attitude in verb suffixes), and structural ambiguities in long sentences. This paper introduces a semantic analyzer built on an Extended Case Frame Model designed to address these complexities.
The Extended Case Frame Model
The core of the system is a move away from simplistic case frames toward a more robust model. While traditional models use a small set of universal primitives, the authors argue for a larger, more redundant collection of relations to capture the "richness of language phenomena."
Analysis Pipeline
The analyzer employs a four-stage process to transform raw text into a Frame Type Semantic Structure (FTSS):
- Morphological Stage: Segmenting the text into individual words.
- Case Element Stage: Identifying noun and adverb phrases.
- Simple Sentence Stage: Matching these phrases to specific predicate slots within the case frames.
- Whole Sentence Stage: Connecting sub-sentences and resolving embedded clauses.
To manage ambiguity, the system generates multiple possible semantic structures and assigns plausibility scores based on how well they align with the knowledge base. The highest-scoring structure is prioritized for final output.
Handling Ellipses with Local Scene Frames (LSF)
A key innovation is the use of Local Scene Frames. Because Japanese often omits contextually "obvious" information, the system maintains a "local scene" of recent participants and events. When a required slot in a case frame is empty, the system searches the LSF to find a likely candidate from previous discourse, effectively "seeing" the missing actors.
System Integration and Results
The analyzer was integrated into experimental systems like JMACS (a multi-window editor) and LUTE (Language Understander, Translator & Editor).
The system's capabilities are demonstrated through the analysis of complex sentences, such as:
"When we see the signboard of a passing station from an express train's window, we can read it if the name is simple. However, it is difficult in the case of a lot of characters."
As shown in the LUTE interface (Figure 1), the analyzer successfully identifies the source ("from an express train's window"), the object ("signboard"), and the conditional relationship between seeing and reading. Crucially, it resolves the ellipsis in the second sentence, identifying that "difficulty" refers back to the act of "reading" mentioned in the first.
Conclusion
The research emphasizes that practical language understanding requires a rich, redundant knowledge base rather than minimal primitives. While the system effectively handles complex technical Japanese, it faces a "knowledge bottleneck," as scaling the hand-crafted frames to general-purpose language is a significant manual undertaking. Nevertheless, the iterative growth of semantic structures and the use of context-aware frames represent a major step forward in processing real-world Japanese text.
Summary updated