Machine Translation: its History, Current Status, and Future Prospects

Jonathan SlocumACL 1984

TL;DR

This paper provides a comprehensive historical overview and current status report on Machine Translation (MT), arguing that despite past skepticism, MT systems are now viable and increasingly cost-effective for specific applications, particularly in technical domains. It highlights the shift from ambitious "fully-automatic high-quality" goals to pragmatic, user-driven acceptance of imperfect but useful systems, driven by the growing demand for translation and the limitations of human translators.

What Problem Does This Paper Address?

The paper addresses the persistent challenge of translating natural human languages automatically, a field that has historically been plagued by over-ambition, disillusionment, and a lack of clear understanding of what constitutes "acceptable" translation. It aims to demystify Machine Translation (MT) and Machine-Aided Translation (MAT) by tracing their evolution, explaining the underlying technologies, and assessing their current practical utility and future prospects. The author tackles the perception gap between the U.S. view of translation as a simple skill and the more professionalized, domain-specific approach taken in Europe and Canada, which is crucial for understanding the real-world acceptance and development of MT.

Methods

The paper categorizes MT systems into three broad types:

  • Machine Translation (MT): Aims for full automation, with the computer responsible for the entire translation process.
  • Machine-Aided Translation (MAT): Involves human-computer interaction, either with the computer assisting the human translator (MAHT) or the human guiding the computer's translation process (HAMT).
  • Terminology Databanks (TD): Provide access to specialized vocabulary.

It also outlines key linguistic techniques:

  • Direct vs. Indirect Translation: Direct systems translate word-for-word with minimal linguistic analysis, while indirect systems aim for a deeper understanding of meaning before synthesis.
  • Interlingua vs. Transfer: Interlingua approaches aim for a language-independent representation of meaning, while transfer approaches map between language-specific representations.
  • Local vs. Global Scope: Local scope focuses on word-level analysis, while global scope attempts a more comprehensive sentence or paragraph analysis.

The paper then reviews historical projects (GAT, CETA, METAL, TAUM, ALP) and current production systems (SYSTRAN, LOGOS, METEO, Weidner, SPANAM, CULT, ALPS), detailing their approaches, strengths, and weaknesses.

Datasets

Not explicitly discussed in terms of specific datasets used for training or evaluation. The paper refers to the vast dictionaries developed for systems like SYSTRAN (over a million entries for USAF FTD) and SPANAM (55,000 entries), and the ongoing expansion of dictionaries for systems like METAL. The "data" processed by these systems is primarily technical documentation, scientific papers, and weather reports.

Key Insights and Ideas

  • Pragmatism over Perfection: The field has shifted from the unrealistic goal of "fully-automatic high-quality" translation to accepting systems that are "good enough" for their intended purpose, often involving human post-editing.
  • The Human Translation Context Matters: The author emphasizes that the U.S. perception of translation as a low-skill activity contrasts sharply with the professionalized, domain-specific approach in Europe and Canada, influencing the acceptance and development of MT.
  • Cost-Effectiveness is Key: For businesses and governments, the primary drivers for adopting MT are cost reduction and increased speed, even if the output isn't perfect.
  • Technical Translation is the Sweet Spot: MT excels in technical domains where terminology is critical and style is less important, unlike literary translation.
  • The "Arrived" Status of MT: Despite ongoing research, production MT systems are already working and providing economic benefits to users, making them "arrived" in a practical sense.
  • The Importance of Transfer: The paper suggests that the transfer approach, which maps between language-specific representations, is often more practical than pure interlingua for current MT systems.
  • AI's Hesitation: The author notes that many AI researchers have historically ignored MT, viewing it as a separate, less sophisticated problem, which has hindered progress in developing truly multilingual AI systems.

Key Results

  • Production Systems are Viable: Systems like SYSTRAN, LOGOS, METEO, Weidner, SPANAM, CULT, and ALPS are in active use by various government agencies and private companies, demonstrating practical utility.
  • Productivity Gains: Users report significant productivity increases, with some translators doubling their output when using MT aids. For example, SPANAM reduced manpower requirements by 45% and monetary savings of 61%.
  • Cost Savings: MT systems are found to be cost-effective, especially when compared to the high cost and limited availability of human translators for technical documents.
  • SYSTRAN's Success in CEC: The European Commission's use of SYSTRAN for English-French, French-English, and English-Italian translation has shown relative success, leading to further versions.
  • METEO's Unique Automation: TAUM-METEO is highlighted as the world's only truly fully-automatic MT system, integrated into a national weather communications network.
  • Weidner's Integration: The Weidner system's tight integration with word-processing technology and its ability to produce formatted translations are noted as significant features.

Limitations and Future Work

Limitations: * Quality is Still Imperfect: No current MT system produces output indistinguishable from good human translation. * "Pure AI" Techniques Lagging: Many current production systems are based on older, less sophisticated technology, and "pure AI" techniques are not yet widely integrated into operational MT. * Scarcity of Research Systems: In the U.S., there's a lack of significant MT research systems, unlike in Japan and Europe. * Language Specificity: English is not a representative language for MT development, and systems focused solely on English may not generalize well. * GETA's Software Constraints: The GETA system is limited by its assembly language software, making modifications difficult and hindering innovation. * SUSY's Primitive Software: The SUSY system's software design, with its independent linguistic strata, makes rule management and interaction management difficult.

Future Work: * Continued R&D: There is a strong need for ongoing MT R&D to improve quality and reduce costs, driven by increasing demand. * Integration with AI: The author suggests that MT research, when integrated with AI and focused on multiple languages, could become a scientifically respectable discipline and a crucial test for AI. * Focus on Meaning Representation: Future systems will likely focus on deeper meaning representations and more sophisticated linguistic theories. * Japan's Commitment: Japan's significant investment in MT R&D, particularly within the Fifth Generation project, is expected to drive future advancements. * Multilingual Systems: Projects like EUROTRA aim to develop true multilingual systems, which will require significant effort in analysis, transfer, and synthesis modules. * Translator Collaboration: The author stresses the importance of consulting with human translators for the design and development of MT systems.

Summary updated

BibTeX
@inproceedings{slocum1984machine,
  author = {Jonathan Slocum},
  title = {{Machine Translation: its History, Current Status, and Future Prospects}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure