Introduction
The relationship between meaning and text is one of the most significant hurdles in natural language processing. In general language, the sheer volume of synonyms and ambiguities makes computational modeling nearly impossible. Researchers Richard Kittredge and Igor Mel’Cuk propose that the solution lies in the study of sublanguages—specialized domains like stock market reports or technical manuals. These domains possess a restricted vocabulary and narrow semantic scope while maintaining the paraphrastic richness of natural language. Their work focuses on the "Stock Market Report Automation Device" (SMRAD), a system designed to map meanings to varied, natural-sounding prose.
The Meaning-Text Model (MTM)
The core of the SMRAD system is the Meaning-Text Model (MTM) theory, which decomposes the transition from abstract meaning to surface text into several distinct levels.
Semantic Representation
At the highest level, meaning is represented as an oriented semantic network. In this graph-based structure, nodes represent "predicates" (actions or relations) or "entities" (objects like stocks and prices). Arcs are numbered to indicate specific argument slots. For instance, a price change is modeled using a CHANGE predicate with four arguments: the stock, the initial price, the final price, and the net change.
As seen in the semantic network above, the model also incorporates subjective qualifiers like LARGE or SMALL. These qualifiers are essential for the system to choose appropriate verbs—such as "jump" for a large increase or "edge up" for a small one—during the synthesis process.
Deep Syntactic Level
The semantic network serves as a blueprint for the deep syntactic level. Here, the graph is mapped onto an unordered dependency tree where nodes are lexical items (words) and arcs represent syntactic relations such as "subject of" or "modifier."
The transition between these levels is governed by lexical and syntactic correspondence rules. Lexical rules associate a semantic subgraph with a specific word, while syntactic rules dictate how these words are assembled into a coherent tree structure. Because the underlying semantic network is not "consumed" during this process, the system can generate multiple synonymous sentences from a single meaning, ensuring stylistic variety.
Key Insights and Implementation
The SMRAD system, implemented in PROLOG, leverages several key linguistic concepts to achieve natural output:
- Bidirectionality: The model is designed to be reversible. The same rules used for text generation (synthesis) are intended to eventually facilitate text interpretation (analysis).
- Lexical Functions: To handle natural word pairings (collocations), the authors utilize "lexical functions." For example, a function might map a "large" change to the intensifier "sharply" in the phrase "up sharply."
- Domain Constraints: By focusing on a sublanguage, the authors eliminate much of the ambiguity found in general English, allowing for a more precise "semantic calculus."
Conclusion and Future Work
This research provides a robust theoretical foundation for automating the generation of specialized reports. While the initial implementation focused on synthesis within the stock market domain, the authors envision extending the model to full text analysis. Although the reliance on hand-crafted linguistic rules presents a challenge for scalability compared to modern statistical methods, the SMRAD project demonstrates that sublanguages offer a viable path toward truly computable models of natural language meaning.
Summary updated