Abstract
This paper presents a novel approach to natural language parsing that leverages "pragmatic expectations"—contextual information regarding a speaker's likely communicative intent—to guide a Head-driven Phrase Structure Grammar (HPSG) parser. By integrating high-level discourse goals with low-level grammatical constraints, the proposed method achieves a ten-fold increase in processing speed while significantly improving disambiguation accuracy. The system effectively prunes the search space of possible sentence meanings by treating pragmatics as a filter during the parsing process rather than a post-processing step.
Introduction
Computational natural language understanding is frequently hindered by ambiguity. Purely syntactic parsers, which rely solely on grammatical rules, face two primary challenges: local ambiguity, where grammatically possible but semantically nonsensical sub-structures are explored, and global ambiguity, where a single sentence yields multiple valid grammatical interpretations.
Consider the sentence, "Robin promised to come at noon." As shown in Figures 1 and 2, a standard parser must decide whether "at noon" modifies the act of promising or the act of arriving. While human listeners resolve this using conversational context, a standard parser often wastes significant computational resources exploring both possibilities. This research aims to use conversational context to guide the parser toward the intended interpretation immediately.
Methodology
The proposed system bridges the gap between social context and technical grammar by combining HPSG with Systemic Grammar and pragmatic modeling. HPSG provides a detailed, lexicalized framework where words carry complex feature structures, while pragmatic modeling captures discourse goals such as politeness and register.
Feature Mapping and Clausal Flattening
The core innovation lies in mapping "flat" pragmatic expectations into the deep, hierarchical structures of HPSG. Because HPSG structures are often deeply nested, the author introduced a "clausally flattened" structure. This allows the parser to share a "clausal value" across different sentence segments, facilitating the comparison of proposed grammatical structures against predicted pragmatic roles.
Best-First Search
The parser employs a best-first search algorithm. Instead of an exhaustive search, it uses a heuristic evaluation function to rank partial parses (edges) based on their alignment with predicted pragmatic roles. This process utilizes abductive reasoning to hypothesize a constituent's role early in the parse, allowing guidance to influence the search as soon as possible.
Results and Discussion
The integration of pragmatic guidance yielded dramatic improvements in efficiency and accuracy. In complex test cases involving conversational exchanges, the search space was significantly reduced. For instance, an unguided parser considered 252 distinct constituents and produced 24 possible parses. In contrast, the guided parser considered only 39 constituents and converged on the single correct parse immediately. This reduction resulted in a greater than ten-fold speedup in CPU time.
Conclusion
By treating pragmatics as an active guide during parsing, this system maintains grammatical modularity while achieving high efficiency. Future work will focus on scaling the system to broader grammar coverage and extending the "clausal flattening" technique to non-verbal constituents such as noun phrases.
Summary updated