Abstract
This paper addresses the long-standing debate in natural language processing between modular and integrated parsing strategies. Modular systems often suffer from combinatorial explosions due to late semantic resolution, while integrated systems suffer from representational inefficiency. The author proposes MOPTRANS, a system that maintains separate syntactic and semantic knowledge bases but combines them dynamically during processing. By using semantics to propose connections and syntax to filter them, MOPTRANS achieves efficient ambiguity resolution without sacrificing the generality of syntactic rules.
Introduction
In the mid-1980s, Natural Language Processing (NLP) was divided into two primary camps. Modular systems performed syntactic analysis first to create a parse tree, then handed it to a semantic component. This often led to a "combinatorial explosion" because syntax alone cannot resolve many ambiguities (such as prepositional phrase attachment), forcing the parser to track dozens of unlikely interpretations.
Integrated systems attempted to solve this by mixing syntax and semantics into a single rule base. While this resolved ambiguity early, it was representationally inefficient. Syntactic rules were hard-coded into specific semantic rules for every individual verb, resulting in massive, redundant rule sets that lacked global syntactic awareness. Lytinen’s MOPTRANS seeks to build a system that gets the best of both worlds: the ambiguity-killing power of semantics and the clean, generalized rules of formal syntax.
Methods
The author introduces MOPTRANS, a machine translation system that utilizes a novel "dynamic combination" approach. Unlike previous integrated parsers, MOPTRANS maintains a largely autonomous set of syntactic rules (inspired by Marcus’s PARSIFAL) and a separate semantic analyzer.
As the system reads a sentence, the semantic analyzer looks at the concepts in "active memory" and proposes potential connections based on what "makes sense" (e.g., a person is a good candidate for the "actor" of a "giving" action). The system then consults the syntactic grammar. A syntactic rule is only applied if it matches the semantic connection currently being proposed.
Because it maintains a separate syntactic representation, the system can use "global" syntactic information (like whether a main verb has already been found) to resolve ambiguities that "local" semantic rules might miss. In this model, syntax acts as a gatekeeper: semantics might think two concepts "fit" together, but the syntactic rules ensure that the word order and grammatical structure actually allow that connection.
Evaluation and Datasets
The system was evaluated using a specialized dataset developed for the MOPTRANS project, consisting of short (1–3 sentence) newspaper stories concerning terrorism and crime. The parser was tested on five languages: English, Spanish, French, German, and Chinese. The system contained enough linguistic and semantic knowledge to parse approximately 25–50 stories for each input language, demonstrating robustness across typologically different structures.
Key Insights
- The Cost of Delay: The paper illustrates how a single unresolved ambiguity can "cascade," creating an exponentially larger number of interpretations. Resolving these via semantics immediately is computationally essential.
- Representational Efficiency: By separating the "how" of syntax (word order patterns) from the "what" of semantics (thematic roles), the author avoids duplicating the same syntactic logic across thousands of different verb definitions.
- Search Space Reduction: In the sentence "John gave Mary a book," the system never considers "Mary" as the direct object of "gave" because the semantic analyzer prefers "Mary" as the recipient. Since a valid syntactic rule (Dative Movement) supports this semantic preference, the incorrect syntactic path is pruned instantly.
Results and Discussion
MOPTRANS successfully handles sentences that were historically difficult for integrated parsers, such as "I saw the soldier called to his sergeant." It correctly identifies "called" as a passive participle because its global syntactic rules recognize that "saw" has already filled the "main verb" slot.
However, the system has limitations. It relies on heavily hand-coded semantic and syntactic knowledge, limiting its use to the specific domain of crime and terrorism stories. While 25–50 stories per language was significant for 1986, the manual labor required to scale this to general-purpose language remains a massive hurdle. This "propose-and-filter" loop represents a significant step toward efficient, knowledge-based natural language understanding.
Summary updated