THEMES FROM 1972

Robert F. SimmonsACL 1982 · DOI

Abstract

In this retrospective analysis, we examine the state of computational linguistics circa 1972—a period marked by theoretical divergence between structuralists, transformationalists, and early AI researchers. We present "Pogen," a procedural logic-based system designed to generate variations of John Keats’ poetry. By utilizing phrase structure grammars within a logic programming framework, Pogen demonstrates that classical linguistic structures remain essential for maintaining thematic consistency and syntactic integrity in controlled natural language generation (NLG).

1. Introduction: The 1972 Landscape

The year 1972 represented a "time of confusion" in the study of language. The field was divided into silos: structuralists focused on surface patterns, transformationalists sought deep-seated rules, and "computerniks" (early AI researchers) prioritized functional logic. Robert F. Simmons likens this era to the Sufi parable of the blind men and the elephant, where each researcher perceived only a fragment of the linguistic whole. This paper argues that the integration of these perspectives—specifically through procedural logic—offers a path toward more coherent machine-generated text.

2. The Challenge of Controlled Generation

The primary technical challenge addressed is controlled natural language generation (NLG). While generating syntactically correct sentences was becoming feasible, ensuring that a machine could produce text that was semantically "on-theme" remained difficult. Simmons explores how phrase structure grammars, often dismissed as "old-school," can be repurposed to guide the flow of computation in language tasks, ensuring that generated output adheres to specific semantic constraints.

3. Methodology: The Pogen System

The "Pogen" (Poem Generator) system was implemented in a LISP-based version of Prolog. It operates through three primary mechanisms:

3.1 Substitution Classes (SCLASS)

SCLASS represents sets of semantically related terms. For instance, a class representing "Beauty" might contain tokens such as life, knowledge, wisdom, and love. These classes provide the vocabulary from which the system draws.

3.2 Line Rules (KLINE)

KLINE serves as the structural skeleton of the generation process. These rules define the templates for lines, specifying where fixed terminal symbols (e.g., "is") and variables from the SCLASS should be placed. This mimics the role of phrase structure rules in traditional grammar.

3.3 Procedural Logic and Memory

The system employs a CHOOSE function for pseudo-random selection. To prevent the output from becoming a "word salad," Pogen includes a memory mechanism. Once a specific term is selected for a semantic variable, the system can record and reuse that choice, ensuring thematic consistency across multiple lines of a verse.

4. Data and Implementation

Rather than utilizing large-scale corpora, the system relies on a hand-crafted knowledge base derived from the concluding lines of John Keats’ "Ode on a Grecian Urn":

"Beauty is truth, truth beauty, that is all / Ye know on earth and all ye need to know."

The "dataset" consists of substitution classes and line rules manually extracted and expanded from this specific literary context to allow for variations on Keats' original themes.

5. Key Insights

5.1 Grammar Symmetry

A significant finding is the bidirectionality of linguistic systems. The same grammar used to parse English into semantic representations can be inverted to generate English from those same representations. This symmetry simplifies the architecture of translation and question-answering systems.

5.2 Semantic Dependency

By respecting the dependencies of words as they occur in a text, researchers can control the "flavor" of the output. This moves NLG from random generation toward meaningful communication by ensuring that word choices are contextually linked.

6. Results: "Robot-Poetry"

Pogen successfully generated verses that, while perhaps lacking human artistic nuance, maintained strict syntactic and thematic coherence. Examples include:

  • (LOVE IS LOVE -- LOVE LOVE) / (ITS ALL YE HAVE TIL HEAVEN) / (ITS ALL YE NEED TO HAVE)
  • (WISDOM IS RAPTURE -- RAPTURE WISDOM) / (ITS WHAT YOU MEET FOR LIVING) / (ITS WHAT YOU WANT TO MEET)

While Simmons humorously admits these verses might be "those Keats wisely rejected," he highlights that the program successfully demonstrates thematic consistency and syntactic correctness through simple recursive rules.

7. Limitations and Conclusion

The Pogen system is limited by its reliance on manually defined templates and classes; it lacks a deep conceptual understanding of the language it manipulates. However, the work points toward a future where these methods are integrated into broader systems for machine translation and complex story generation. As Simmons suggests, procedural logic provides the "friendly environment" necessary for linguistic theories to reach their full potential in artificial intelligence.

Summary updated

BibTeX
@inproceedings{simmons1982themes,
  author = {Robert F. Simmons},
  title = {{THEMES FROM 1972}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1982},
  doi = {10.3115/981251.981278},
}
Figure