[Mova](https://md.mova.elamur.ai/free/en/) · [this page in HTML](https://mova.elamur.ai/free/en/docs/papers/toma-1976-systran/)

[home](https://md.mova.elamur.ai/free/en/) / [docs](https://md.mova.elamur.ai/free/en/docs/) / [papers](https://md.mova.elamur.ai/free/en/docs/papers/) / [Machine translation](https://md.mova.elamur.ai/free/en/docs/papers/mtrans/) / SYSTRAN

# SYSTRAN

**Authors:** Peter Toma · **Year:** 1976 · **Venue:** Foreign Broadcast Information Service Seminar on Machine Translation **Link:** [aclanthology.org/1976.earlymt-1.11](https://aclanthology.org/1976.earlymt-1.11/) · **Licence:** [CC BY-NC-SA 3.0](https://creativecommons.org/licenses/by-nc-sa/3.0/) — summary only

[Foreign Broadcast Information Service Seminar on Machine Translation](https://aclanthology.org/volumes/1976.earlymt-1/)[licence](https://creativecommons.org/licenses/by-nc-sa/3.0/)

- [Publisher's page](https://aclanthology.org/1976.earlymt-1.11/)
- [PDF](https://aclanthology.org/1976.earlymt-1.11.pdf)

**Authors (1)**

1. Toma, Peter

Citations and author metrics from OpenAlex (); authors with the same name are not always told apart.

## What it is about

A short account of the SYSTRAN rule-based system, given as an oral presentation and written up afterwards. The main claim is that the system is reliable, economic and satisfies its users. The engine is separated from the linguistic data: dictionaries and rules can be changed without touching the programs, and a new language pair needs only a new source analysis module, with the dictionary lookup and the target synthesis reused. Rules are written in a macro language for linguists. The paper gives throughput figures: about 300,000 words per CPU hour in batch, and roughly 18 seconds of processor time per 1,000 words. It gives no accuracy numbers.

## Ideas

- I1. Test the claim that a new reader changes nothing else: after adding the Hebrew reader in `tools/translate/src/hebrew.rs`, check that the English standard and the English-only output stay byte identical. Make this a fixed regression check, run twice.
- I2. Let hand-written overrides act with precedence over learned tables, as SYSTRAN's glossary entries do, and measure them as a separate layer: how many tokens each override changes and whether the change is kept by the residual.
- I3. Report time per stage as a throughput number (words per second in each of reader, encoder and stack) in `measure.rs`, so that a later change that slows one stage shows up in the same table.

## For Mova

The paper suggests that a modular design is worth testing as a property, not only assumed: reader, standard encoder and stack should be separable, and a change in one should leave the others' outputs unchanged. This fits the layer-by-layer residual in exact mode. Override precedence is already present in the author DELTA, so the new part is the stage-level throughput record in `tools/translate/src/measure.rs`.

Topics: [Machine translation](https://md.mova.elamur.ai/free/en/docs/papers/topics/machine-translation/)

## Cite

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

```
@inproceedings{toma1976systran,
  title = {{SYSTRAN}},
  author = {Toma, Peter},
  booktitle = {Foreign Broadcast Information Service Seminar on Machine Translation},
  year = {1976},
  url = {https://aclanthology.org/1976.earlymt-1.11/}
}
```

- [← previous Machine Translation Strategies: A Comparison of F-Structure Transfer and Semantically Based Interlingua](https://md.mova.elamur.ai/free/en/docs/papers/thunes-1994-machine-translation-strategies-comparison-f/)
- [next → Leveraging Rule-Based Machine Translation Knowledge for Under-Resourced Neural Machine Translation Models](https://md.mova.elamur.ai/free/en/docs/papers/torregrosa-2019-leveraging-rule-based-machine-translation/)
- [↑ back to Machine translation](https://md.mova.elamur.ai/free/en/docs/papers/mtrans/)