Semantically Governed Machine Translation of BE-clauses with Adverbs and Prepositional Phrases. Demonstration of a Ring-Model for German/English, English/German including Analysis-Synthesis, Transformation, Transition, and Generation

Klaus Gunther SchweisthalCOLING 1967

TL;DR

This paper introduces a "Ring-Model" for bidirectional German-English machine translation, focusing on BE-clauses with adverbs and prepositional phrases. It proposes a novel approach to semantic equivalence and stylistic acceptability by integrating analysis, synthesis, transformation, transition, and generation through a system of factor formulas and matrices.

What Problem Does This Paper Address?

Traditional machine translation often struggles with semantic equivalence and stylistic naturalness, especially when dealing with complex grammatical structures like adverbial and prepositional phrases. The authors identify a need to move beyond simple word-for-word translation and address the nuances of meaning and grammatical function. They aim to create a system that can accurately translate not just the literal meaning but also the intended sense and grammatical correctness between German and English, specifically for sentences containing "to be" (BE) clauses augmented with adverbs and prepositional phrases.

Methods

The core of the proposed solution is a "Ring-Translation-Model" that operates on a "factor formula" representation. This model involves several key stages:

  • Analysis-Synthesis: Incoming phrases are analyzed into semantic factors.
  • Morpho-nomo-transformation: This is a crucial step where phrases are converted into factor formulas.
  • Nomo-nomo-transformation/transition: Factor formulas are transformed or transitioned between languages. This process is guided by a "coordinator's system of maximum agreement" to ensure semantic equivalence. If an identical factor formula exists in the target language, a "transition" is used; otherwise, a "transformation" is applied.
  • Nomo-morpho-transformation: The factor formula is then converted back into an expressible phrase in the target language.
  • Reductive nomo-nomo-transformation: Used in the English-to-German direction to simplify the factor formula.

The model relies heavily on a system of "association lists" and "function matrices" to manage the relationships between semantic factors, word classes, and grammatical structures. These matrices define the permissible combinations of prepositions and noun classes, and the coordination of subject and complement noun classes.

Datasets

The paper mentions the creation of a German and English lexicon comprising 16 prepositions, 16 adverbs, and one noun from each of five noun classes. The nomo-association chart (Figure 2) is filled with over 30 semantic factors. The criteria for "meaningful" BE-clauses were based on approximately 40,000 examples of modern German and a smaller collection of English examples, supplemented by interviews with native English speakers. No specific benchmark datasets are explicitly named or introduced as new.

Key Insights and Ideas

  • Semantic Equivalence over Lexical Equivalence: The model prioritizes translating the underlying semantic factors rather than just finding direct word-for-word equivalents, which is key to handling linguistic differences.
  • Factor Formulas as a Universal Representation: The use of factor formulas as an intermediate representation allows for a more abstract and language-independent way to capture meaning.
  • The Power of Matrices: The function matrices are central to defining the grammatical and semantic constraints, ensuring that generated phrases are both meaningful and grammatically correct in the target language. This is particularly evident in how they restrict preposition-noun combinations.
  • Handling Ambiguity through "Maximum Agreement": The system uses a principle of "maximum agreement" to resolve ambiguities and choose the most appropriate translation when multiple options exist.
  • Bidirectional Translation: The "Ring-Model" is designed to be reversible, enabling translation from German to English and vice versa.

Key Results

The paper demonstrates the satisfactory output of the ring model through examples. For instance, the German phrase "Vater ist drinnen" can be translated to "Vater ist innerhalb des Hauses" and then to "Father is inside (of) the house." The system successfully handles cases where direct translation is not possible, such as the English sentence "The candle is inside (of) the father," which is translated into German as "Die Kerze ist im Körper des Vaters eingeschlossen" via an auxiliary verbalization program. The authors claim that for cases where function matrices are identical in structure between languages, semantically equivalent translations can be executed.

Limitations and Future Work

  • Matrix Identicality: The paper notes that for cases where function matrices are not identical in structure between languages, an auxiliary program is needed to ensure proper translation. This suggests a limitation in the direct applicability of the core ring model without additional components.
  • Auxiliary Programs: The reliance on auxiliary programs for certain transformations (e.g., verbalization of semantic factors, handling German adverbs with no English counterpart) indicates that the core ring model might not be fully self-contained for all translation scenarios.
  • Future Work: The authors imply that further development would involve expanding the lexicon and refining the matrices to cover a wider range of linguistic phenomena. The need for auxiliary programs also suggests that future work might focus on integrating these more seamlessly or developing more comprehensive core matrices.

Summary updated

BibTeX
@inproceedings{schweisthal1967semantically,
  author = {Klaus Gunther Schweisthal},
  title = {{Semantically Governed Machine Translation of BE-clauses with Adverbs and Prepositional Phrases. Demonstration of a Ring-Model for German/English, English/German including Analysis-Synthesis, Transformation, Transition, and Generation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1967},
}
Figure