Machine Translation : What Type of Post-Editing on What Type of Documents for What Type of Users

Anne-Marie LaurianCOLING 1984

TL;DR

This paper argues that machine translation (MT) isn't a one-size-fits-all solution. It proposes a typology of documents and errors to guide post-editing efforts, distinguishing between rapid, meaning-focused post-editing and more thorough, style-conscious revision. This approach aims to optimize MT use by matching document type, user needs, and post-editing intensity.

What Problem Does This Paper Address?

The paper tackles the challenge of effectively integrating machine translation (MT) into professional translation workflows. While MT has advanced significantly, raw output is often not directly usable. This leads to the need for post-editing, but the extent and type of post-editing required can vary dramatically. The authors identify a lack of clear guidelines for deciding when MT is appropriate, what types of documents are best suited for it, and how much post-editing is necessary for different user needs and document dissemination goals. This ambiguity can lead to inefficient use of resources, frustration with MT, and potentially subpar translations.

Methods

The authors' approach is primarily analytical and empirical, drawing on observations from the European Commission's use of the SYSTRAN system in Luxembourg. Their methodology involves:

  • Error Analysis: A detailed study of errors remaining in raw MT output to develop a linguistic typology of common MT mistakes. This typology categorizes errors into three main types: isolated word errors, errors in expressing relations, and errors in structure/information display, with further sub-classifications.
  • Document Typology: Based on the error analysis, the paper proposes a framework for classifying documents according to their suitability for MT and the expected post-editing effort. This typology considers factors like subject matter, function, structure, sentence complexity, and terminology.
  • Post-Editing Typology: The core contribution is the distinction between two primary types of post-editing:
  • Conventional Post-Editing (CPE): Aims to produce a text as close as possible to a human translation, focusing on high quality and style.
  • Rapid Post-Editing (RPE): Focuses on ensuring the text is linguistically correct and conveys the meaning accurately, but without significant attention to style.
  • Empirical Study: An experiment was conducted where time was the primary criterion for differentiating CPE and RPE. This study aimed to understand the linguistic attitudes induced by time limitations and to define the boundaries between necessary, possible, and superfluous post-editing.

Datasets

Not explicitly discussed in the paper. The authors refer to their observations based on "several studies I carried out on the SYSTRAN output produced in Luxembourg within the Commission of the European Communities." This suggests the "dataset" was the corpus of translated documents processed by SYSTRAN within that organization.

Key Insights and Ideas

  • MT is not a universal solution: The paper strongly emphasizes that not all documents are suitable for MT, and improper use leads to negative perceptions.
  • Post-editing is a distinct skill: Post-editing is presented not as simple revision or correction but as a new specialization requiring specific rules and approaches.
  • Document typology is crucial for efficiency: Classifying documents based on their characteristics and expected use allows for more informed decisions about MT adoption and post-editing intensity.
  • Two distinct post-editing modes: The distinction between Rapid Post-Editing (RPE) for quick information needs and Conventional Post-Editing (CPE) for dissemination is a key takeaway. RPE prioritizes meaning and correctness, while CPE aims for stylistic polish.
  • Style in technical documents is about transparency: Unlike literary style, the "style" of technical documents should ideally be transparent, allowing information to be conveyed without the reader noticing the writing itself. Minimal post-editing can often achieve this optimal style.
  • User satisfaction is the ultimate metric: The adequacy of a translation's style should ultimately be judged by the end-user's satisfaction.

Key Results

The paper doesn't present quantitative results in terms of specific error reduction percentages or speedup metrics. Instead, its key results are qualitative and conceptual:

  • A linguistic typology of MT errors: This provides a framework for understanding common MT failures.
  • A framework for document selection: The paper outlines criteria for deciding if a document is suitable for MT.
  • A clear distinction between RPE and CPE: This provides a practical model for managing post-editing efforts based on time constraints and quality requirements.
  • Identification of three levels of post-editing necessity:
    1. Necessary post-editing: Essential for making the text understandable, clear, readable, and exact.
  • Possible post-editing: Involves stylistic refinement to adapt to the communication situation, author, and reader.
  • Superfluous post-editing: Efforts to achieve perfection on documents not requiring it, often done by individuals who struggle to accept that perfection isn't always the goal.

Limitations and Future Work

  • Limitations:
    • The study is based on observations of a specific system (SYSTRAN) within a particular organization (CEC Luxembourg), which may limit generalizability to other MT systems or contexts.
  • Quantitative data on error rates, post-editing time savings, and quality improvements is not explicitly presented.
  • The "dataset" is not formally defined or made available.
  • Future Work:
    • The authors suggest that their work can help inform the development of new training programs for translators and linguists, focusing on the nuances of MT and post-editing.
  • The paper implies that further research could explore the "future programmes that could be built up for new generations of students" in the context of evolving translation technologies.
  • The authors mention a forthcoming publication on conventional and rapid post-editing, suggesting a deeper dive into the methodologies.

Summary updated

BibTeX
@inproceedings{laurian1984machine,
  author = {Anne-Marie Laurian},
  title = {{Machine Translation : What Type of Post-Editing on What Type of Documents for What Type of Users}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1984},
}
Figure