A computer model for Russian grammatical description and method of English synthesis in machine translation

D.m. YatesCOLING 1967

TL;DR

This paper introduces a novel computational model for representing Russian grammatical structures in a way that's independent of specific language conventions. This model, combined with a synthesis algorithm, aims to facilitate machine translation by bridging the gap between Russian analysis and English generation.

What Problem Does This Paper Address?

Machine translation systems often struggle with the nuances of grammatical structure and the differences in how languages express these structures. Specifically, this paper addresses the challenge of representing complex Russian grammatical features in a way that can be effectively used by an English synthesis algorithm. Traditional approaches might directly map structures, leading to awkward or incorrect translations. The authors aim to create an intermediate representation that captures the essential grammatical information from Russian, allowing a separate English synthesis process to generate natural-sounding English output, regardless of the specific linguistic conventions used in the source or target languages.

Methods

The paper proposes a model based on two core linguistic features: structure and system.

  • Structure: This is fundamentally a hierarchical representation of constituents (word groups). However, it deviates from traditional constituent structures in four key ways:
  • Constituents can exemplify choices within systems, meaning some linguistic units (like auxiliaries) might not have a direct place in the structure.
  • A single item can occupy multiple positions within the structure, particularly useful for relative words linking clauses.
  • Constituents are not necessarily contiguous in the text.
  • Explicit ordering of sub-constituents is not strictly required if systems are sufficiently powerful.
  • System: This refers to choices within closed sets of possibilities, such as tense or mood. These choices are encoded in "systems words" associated with elements in the structure.

The model represents grammatical structures as list structures in the computer. Each element in the list either contains addresses of other elements (representing inclusion of word groups) or is marked as a terminal element (representing single words or idioms). Elements are labeled with constituent types and roles, and addresses are also labeled with the role of the included group.

The English synthesis algorithm takes this model representation and generates English output. It makes decisions on:

  1. Re-ordering of elements.
  2. Insertion of English "function" words (auxiliaries, etc.).
  3. Selection of English equivalents from a dictionary.
  4. Inflection of English equivalents.

This process is managed by Constituent Type Procedures (CTPs), which are recursive subroutines responsible for generating specific types of constituents (e.g., nominal group CTP, verbal group CTP).

Datasets

Not discussed in the paper. The paper focuses on the model and algorithm itself, using examples derived from Russian analysis.

Key Insights and Ideas

  • Language-Independent Representation: The core idea is to create a grammatical representation that is detached from the surface conventions of either the source (Russian) or target (English) language. This allows for a more robust and flexible translation process.
  • Structure-System Duality: The model's strength lies in its ability to capture both the hierarchical arrangement of linguistic units (structure) and the choices made within grammatical categories (systems). This dual approach provides a richer description than structure alone.
  • Constituent Type Procedures (CTPs): The modular design using CTPs for each constituent type is a practical approach to managing the complexity of English synthesis. This allows for specialized handling of different grammatical units.
  • Flexibility in Incomplete Analysis: The model and algorithm are designed to be useful even when the Russian analysis is incomplete. Arbitrarily linking partial structures allows for synthesis to proceed, with new analysis passes or CTPs being easily incorporated as they become available.

Key Results

The paper describes the model and synthesis algorithm and provides examples of how they would work. While no quantitative results (like translation accuracy percentages) are presented, the authors state that the model and algorithm "proved satisfactory in practical use." The paper demonstrates the model's ability to represent complex Russian structures and outlines the steps the synthesis algorithm takes to generate English.

Limitations and Future Work

  • Incomplete Analysis Coverage: The authors acknowledge that their analysis processes could not recognize all the features described by the model. Specifically, clauses and comparative groups were not used at all.
  • Implementation Status: Some CTPs, like the clause CTP, were not fully implemented, being developed only to the flowchart stage.
  • Limited Semantic/Collocational Techniques: The synthesis decisions are based purely on grammatical data (structural and systemic), without employing semantic or collocational techniques. This implies that the generated translations might lack naturalness in certain contexts.
  • Future Work: The paper mentions that the routine for handling "government phenomena" was developed only to the flowchart stage, suggesting this as a potential area for further development. The authors also note that as new analysis passes or synthesis CTPs become available, they can be incorporated very simply, implying ongoing development and improvement.

Summary updated

BibTeX
@inproceedings{yates1967computer,
  author = {D.M. Yates},
  title = {{A computer model for Russian grammatical description and method of English synthesis in machine translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1967},
}
Figure