Data Preparation for Syntactic Translation

L. W. ToshCOLING 1965

TL;DR

This paper details the methodology for preparing linguistic data for a generalized syntactic translation system developed at the University of Texas. It outlines the challenges and strategies involved in creating machine-readable grammars and translation rules, moving beyond simple word-for-word translation to capture syntactic structures for more accurate cross-lingual communication.

What Problem Does This Paper Address?

The paper addresses the fundamental challenge of machine translation: how to represent and process the complex syntactic structures of natural language to enable accurate translation between languages. Traditional word-for-word translation methods, while simple, fail to capture nuances of grammar, word order, and idiomatic expressions, leading to nonsensical or inaccurate outputs. The authors aim to build a system that can handle these syntactic complexities, moving towards a more robust and generalizable translation capability. This involves not just the translation process itself, but the crucial, often overlooked, step of preparing the necessary linguistic data in a format that computers can understand and utilize.

Methods

The paper describes a hierarchical approach to machine translation, categorizing models into three levels of complexity: word-for-word, rule-for-rule (syntactic), and transformational-semantic. The focus is on the second level, syntactic translation, which involves analyzing the input sentence's structure and transforming it into an equivalent structure in the target language.

Key methodological aspects include:

  • Syntactic Analysis: Using context-free phrase structure grammars to parse input sentences and identify their underlying structures.
  • Transfer Grammar: A mechanism to transform the parsed structures of the source language into equivalent structures of the target language.
  • Syntactic Synthesis: Generating the output sentence in the target language from the transformed structures, ensuring grammatical well-formedness.
  • Data Preparation: A significant portion of the paper is dedicated to the meticulous process of creating these grammars and translation rules. This involves:
  • Selecting representative text corpora in both source (German) and target (English) languages.
  • Manually creating detailed phrase structure descriptions for sentences, labeling grammatical features.
  • Encoding these descriptions into a formal notation for computer processing.
  • Assigning unique identification numbers to grammar rules for tracking and management.
  • Verifying the accuracy and completeness of the grammars through computer-aided analysis.
  • Defining interlingual transformation relationships by mapping corresponding sub-structures between languages.

Datasets

The paper utilizes two primary datasets:

  • German Input Corpus: The first six paragraphs of an essay from "UNIVERSITAS" were used as the input text for German.
  • English Translation Corpus: A human-prepared English translation of the same essay was used as the target corpus.

These corpora were specifically prepared for the demonstration of the translation system, serving as test cases for verifying the morpho-syntactic descriptions and the translation rules.

Key Insights and Ideas

  • The Cruciality of Data Preparation: The paper strongly emphasizes that the success of any machine translation system hinges on the quality and comprehensiveness of the linguistic data prepared. This is not a trivial step but a complex, labor-intensive process.
  • Beyond Word-for-Word: The limitations of simple word-for-word translation are starkly illustrated, highlighting the necessity of understanding and manipulating syntactic structures for meaningful translation.
  • Hierarchical Approach: The concept of a hierarchical system, moving from lexical analysis to syntactic and potentially semantic levels, provides a structured framework for tackling translation complexity.
  • Bidirectional Grammars: The authors aim for grammars that can be used for both analysis (parsing) and synthesis (generation), facilitating a more efficient and integrated translation process.
  • The Challenge of Ambiguity: The paper acknowledges that even with syntactic analysis, ambiguities in language (e.g., pronoun reference) remain a significant challenge, requiring further development and potentially recourse to extralinguistic context.

Key Results

The paper presents a functional syntactic translation model that was demonstrated in January of the year of publication. While specific quantitative performance metrics are not detailed, the paper illustrates the process of data preparation and the resulting translation outputs (shown in Appendices C and D). The authors note that the initial output quality was "unsatisfactory" and attributed this to a combination of program errors and inadequate linguistic data. However, they also highlight that subsequent runs with improved data and rules showed promising results, with computer translations appearing "quite similar" to human translations, though still revealing differences.

Limitations and Future Work

Limitations:

  • Inadequate Linguistic Data: The primary limitation identified is the insufficient quantity and quality of linguistic data, leading to "unsatisfactory" translation output in early stages.
  • Program Errors: Early versions of the system also suffered from program errors that impacted translation quality.
  • Ambiguity in Pronoun Reference: The model struggles with pronoun-antecedent agreement, particularly when multiple antecedents are present, and when context outside the linguistic environment is needed for resolution.
  • Limited Scope of Grammars: The grammars developed so far tend to exclude features from morpho-syntactic description, limiting their ability to handle certain linguistic phenomena.
  • Model Limitations: The current model has limitations in accounting for all semantic collocational controls and certain transformational problems like pronominal reference.

Future Work:

  • Expanding Syntactic Description: The authors plan to expand the syntactic description of all languages under investigation.
  • Testing Transfer Data: Compiling sufficient transfer data between language pairs to test the general validity and adequacy of the model and system.
  • Improving Translation Quality: Investigating ways to improve translation quality by using more sophisticated models, potentially incorporating a grammar of structural semantics.
  • Addressing Semantic Equivalence: Including features for lexical collocation and transformations of semantic equivalence in a higher-order systematic description.
  • Empirical Testing: Empirically testing recent theories and notions of linguistics with a comprehensive data base.

Summary updated

BibTeX
@inproceedings{tosh1965data,
  author = {L. W. Tosh},
  title = {{Data Preparation for Syntactic Translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1965},
}
Figure