Elamur Research⌕
☰
home / papers / 2004 / Language Weaver Arabic-\ensuremath>English MT

Language Weaver Arabic-\ensuremath>English MT

Paper · 2004 · Daniel Marcu, Alex Fraser, William Wong et al.

Summary

This demo paper presents a commercial statistical machine translation system from Language Weaver that translates Modern Standard Arabic into English. The underlying research came from USC/ISI. The company's main contribution was engineering: laboratory SMT systems normally assume unlimited memory and compute, and Language Weaver made one run in real time on an ordinary Windows desktop. The system targets wide coverage of Arabic news and was already deployed at several U.S. Government sites. Builds exist for machines with 1 GB or 2 GB of RAM and a processor of at least 1.5 GHz, and they translate at no less than 500 words per minute. Users can set a trade-off between speed and quality, from rapid gisting output to the best achievable translation for each sentence. The paper reports no BLEU scores or other accuracy evaluation. It shows that data-driven MT had matured enough for practical, resource-limited deployment.

Key points

  • The system is a statistical MT engine for Modern Standard Arabic → English, built on USC/ISI research.
  • Hardware targets are 1 GB or 2 GB RAM variants, a CPU of 1.5 GHz or faster, and standard desktop Windows.
  • Guaranteed throughput is at least 500 words per minute.
  • A user-controlled setting trades speed against quality, ranging from fast gisting to best per-sentence quality.
  • The system covers general Arabic news, is in use at U.S. Government sites, and was demonstrated live on current news and Internet postings.

Interesting ideas

  • I1. Exposing a speed-versus-quality dial to the user lets one engine serve both quick skimming and careful reading, instead of shipping two separate systems.
  • I2. Packaging the same system in builds sized for different memory budgets (1 GB vs 2 GB) widens where it can run without changing its core design.
  • I3. Stating a guaranteed minimum throughput on commodity hardware, rather than only an average speed, turns performance into a checkable requirement.
  • I4. Demonstrating on fresh, live news instead of a curated test set is an honest check of robustness to unseen input.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tree): offer a fast perceptron-only parsing mode and a slower gparse/beam mode behind one switch, chosen per document or per sentence.
  • I2 (world): ship compiled knowledge (WordNet/FrameNet/Wikidata layers) in a compact and a full pack, and record which pack produced each inference for explainability.
  • I3 (learn): add a worst-case words-per-second floor to the evaluation suite so regressions in speed fail tests just like accuracy drops.
  • I4 (answer): regularly run the readers on newly collected texts never used in development, to catch overfitting to the school-test and fable sets.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 3Research subjects and languages the work is about.Machine translation · 1,105 worksEnglish · 2,379 worksSemitic languages · 3,066 works

Cite

Cite as Daniel Marcu, Alex Fraser, William Wong and Kevin Knight. 2004. Language Weaver Arabic-\ensuremath>English MT. Proceedings of the Workshop on Computational Approaches to Arabic Script-based Languages.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{marcu2004language,
  title = {{Language Weaver Arabic-ensuremath>English MT}},
  author = {Daniel Marcu, Alex Fraser, William Wong and Kevin Knight},
  booktitle = {Proceedings of the Workshop on Computational Approaches to Arabic Script-based Languages},
  year = {2004},
  url = {https://aclanthology.org/W04-1617/}
}

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/marcu-2004-language-weaver-arabic-ensuremath-english/