Abstract
The parsing of complex technical documents, such as patents, is often hindered by extreme structural ambiguity and long, dense sentences. This paper presents an object-oriented parser that treats grammatical constituents as active, message-passing objects. By utilizing a pseudo-parallel left-corner algorithm integrated with a "delay mechanism," the system applies syntactic and semantic constraints early in the process. This approach effectively prunes meaningless parse paths, preventing combinatorial explosion while remaining robust to the linguistic irregularities common in real-world technical text.
Introduction
The primary challenge in automated text understanding for technical domains is structural ambiguity. In documents like Texas Instruments patent descriptions, a single sentence can be exceptionally long and dense, leading to a "combinatorial explosion" of potential syntactic interpretations. Traditional parsers often fail by either delaying semantic application too long—resulting in an unmanageable number of junk parses—or by applying it too rigidly, which causes failure when encountering slightly non-standard grammar. This research proposes a middle ground: a system that handles the "messiness" of real-world text by using semantic context to guide the parsing process in real-time through an object-oriented framework.
Methodology
The parser is implemented using an Object-Oriented (OO) approach, specifically leveraging the "flavor" system in ZLISP. This architecture shifts the paradigm from static data structures to active, autonomous entities.
Active Constituents and Local Grammar
Each grammatical category, such as a Noun Phrase (NP) or a Clause, is modeled as an object. These objects maintain internal state, attempt to satisfy their own rules, and communicate with one another via message-passing. The system employs a context-free "Local Grammar" augmented with percolation rules to pass features (e.g., number or verb type) up the tree.
Pseudo-Parallel Left-Corner Algorithm
The parser utilizes a pseudo-parallel left-corner algorithm that combines bottom-up evidence with top-down filtering. This allows the system to track multiple potential interpretations simultaneously. To manage the flow of information, a delay mechanism is implemented. Instead of broadcasting every possible sub-phrase immediately, the parser queues constituents and only "transmits" them for semantic processing once local ambiguity is resolved.
Key Insights: Handling Complexity
A significant innovation of this system is the use of "levellers" and message-passing to manage deeply nested or self-embedding structures.
By treating grammar rules as actors, the system facilitates a bidirectional flow of information. Syntax suggests meaning, while semantic predictions help resolve syntactic choices. The "levellers" allow the parser to process sub-parts of a sentence without waiting for the entire top-level structure to complete, which is essential for the long, rambling sentences found in legal and technical writing.
Results and Discussion
The system was tested on a corpus of Texas Instruments patent descriptions, known for their structural complexity. Qualitative results indicate that the parser successfully handles the "explosive" ambiguity that typically cripples simpler systems. The integration of the left-corner algorithm with a reachability matrix provides an efficient means of parsing long sentences typical of the target domain.
However, several limitations remain. The computational overhead of managing numerous active objects and messages is a concern. Furthermore, the current grammar does not fully address ellipses (null realizations), as doing so without sufficient constraints would degrade performance. Future work will focus on improving semantic integration across paragraph boundaries and scaling the "Local Grammar" to more general, less structured text.
Conclusion
The object-oriented parser offers a flexible and robust architecture for understanding complex technical text. By combining active constituents with a strategic delay mechanism, it balances the need for early semantic pruning with the flexibility required to navigate the structural intricacies of real-world language.
Summary updated