KANT: Knowledge-Based, Accurate Natural Language Translation

What it is about

An overview of KANT, an interlingua system from Carnegie Mellon for translating technical documents written in a controlled source language. The system has four language-independent modules (parser, interpreter, mapper and generator), so that a new language needs its own lexicon and grammar but no change to the code. The parser turns each source sentence into an interlingua expression built from domain concepts (objects, events and properties), and the generator produces a target sentence for each expression, so no bilingual dictionary or transfer grammar is needed for a language pair. The paper reports applications: a Spanish-English utility-management system, and a heavy-equipment documentation system for eleven languages. The authors claim high accuracy without post-editing, because the source is controlled and the domain is narrow. The paper is an overview with no measurements of its own.

Ideas

• I1. Check that the interlingua stays independent. Count, for each new target author, how many components of the reader and the encoder had to change. If the count is not zero, a language-specific fact has leaked into the meta-world, and the concept table needs a fix.
• I2. Carry the domain frame of a concept as part of its identity. In KANT each concept points to a domain frame used during analysis. In Mova the prag frame (event, fact or say, with its roles) can disambiguate two source words that share one concept, which the merge rule currently cannot.
• I3. Measure the coverage of the concept table per kind of text. A Bible book, a school text and a legal text should each have a coverage figure, so that gaps in the meta-world are found by domain, as the KANT authors found their gaps by application.

For Mova

The code is tools/translate (meta.rs for the concept table, and the frame anchors of the prag layer) and the independence check belongs in measure.rs. The paper's modules are a useful picture of the layering, but its interlingua is hand-built for a narrow domain, so the parts to copy are the independence check and the coverage per domain.

Topics: Machine translation

Cite

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

@inproceedings{mitamura1994kant,
  title = {{KANT: Knowledge-Based, Accurate Natural Language Translation}},
  author = {Mitamura, Teruko and Nyberg, Eric and Carbonell, Jaime},
  booktitle = {Proceedings of the First Conference of the Association for Machine Translation in the Americas},
  year = {1994},
  url = {https://aclanthology.org/1994.amta-1.36/}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/mitamura-1994-kant-knowledge-based-accurate-natural/