Interlingual Representations for Multilingual Translation

32 concepts26 papers1 proposal

Imagine you are tasked with building a machine translation system that can translate between five different languages. If you connect every language directly to every other language, you need to build and maintain $5 \times 4 = 20$ distinct translation pathways. Now imagine scaling this to 20 languages: you suddenly need 380 separate pathways.

This quadratic explosion ($O(N^2)$) is the fundamental bottleneck of direct transfer translation. As the number of languages grows, the engineering overhead becomes completely unsustainable.

The solution to this scaling crisis is a beautifully simple architectural pivot: instead of translating directly from one language to another, we translate from a source language into a universal, language-independent intermediate representation of meaning—an interlingua—and then generate the target language from that shared core.

The Power of Decoupling

By introducing an interlingua, the translation process is split into two entirely independent phases:

  1. Analysis: Mapping the source language into the interlingual representation.
  2. Synthesis: Generating the target language from the interlingual representation.
Language A (Source)  ───[ Analysis ]───►  ┌──────────────┐
                                          │  INTERLINGUA │
Language B (Target)  ◄───[ Synthesis ]───  └──────────────┘

This decoupling transforms a quadratic scaling problem into a linear one ($O(N)$). For $N$ languages, a system only requires $N$ analyzers and $N$ synthesizers. As Tomita et al. demonstrated in The Universal Parser Architecture for Knowledge-based Machine Translation, this architecture dramatically reduces the complexity of multi-language systems because adding a new language simply requires connecting it to the central hub, rather than building new models for every existing language pair.

This architectural simplification is so computationally efficient that it historically enabled sophisticated machine translation to operate within severe hardware constraints. For example, Somers et al. leveraged an interlingua-based design in BEDE: A Microprocessor-Based Machine Translation System to run translation pipelines on early microprocessors by abstracting away language-specific structures.

Designing a Language of Pure Meaning

To make an interlingua work, the intermediate representation must capture the underlying semantics of an utterance without biasing toward the syntax, grammar, or vocabulary of any specific language. Researchers have explored several distinct strategies to construct this neutral layer:

Practical Advantages of Decoupled Translation

The benefits of isolating the source and target languages extend far beyond saving computational memory. Decoupling introduces unique functional capabilities:

Source-Language Paraphrasing

Because the analysis phase distills an utterance down to its pure semantic essence, the system can take that interlingual representation and synthesize it back into the original source language. In A Thai Speech Translation System for Medical Dialogs, Alexander et al. demonstrated how this enables source-language paraphrasing for accuracy verification. A doctor speaking English can see a paraphrase of what the system understood before it generates the final Thai translation, ensuring safety in critical environments.

Multi-Engine Hybridization

An interlingual representation provides a stable baseline that can easily cooperate with other translation methodologies. As shown by Zuo et al. in Multi-engine based Chinese-to-English translation system, an interlingua-based engine can be combined with template-based and statistical engines to form a hybrid system, leveraging the structural precision of the interlingua alongside the fluid coverage of data-driven models.

Portability and Edge Deployment

By separating language-specific parsing from the core translation logic, the individual components remain lightweight. This modularity made it possible to deploy early mobile and wearable translation systems, such as the LingWear tourist assistant developed by Fuegen et al. (LingWear: A Mobile Tourist Information System).

Generalizing Beyond Text

The intuition behind interlinguas is not restricted to text-to-text translation. It serves as a universal blueprint for mapping any diverse set of inputs to any diverse set of outputs.

Speech and Sign Language

In global speech translation initiatives surveyed by Waibel et al. (Worldwide ongoing activities on multilingual speech to speech translation), mapping acoustic speech directly to a semantic interlingua insulates the downstream generation phase from the ungrammaticalities and noise inherent in spoken language.

This paradigm even bridges entirely different physical modalities. In Machine Translation between Spoken Languages and Signed Languages Represented in SignWriting, Jiang et al. utilized SignWriting—a visual notation system—as an intermediate representation to bridge spoken words and signed gestures, making sign language translation amenable to standard NLP pipelines.

Database Queries and Code

The exact same architectural logic applies to computer languages. In the PHLIQA question-answering system, Madama (A Control Structure For A Question-Answering System) used a chain of intermediate languages to decouple natural language English queries from the final database schema, ensuring that semantic understanding was not hardcoded to a specific database layout.

Decades later, this concept remains vital in modern polyglot analytics. In PolySem: Efficient Polyglot Analytics on Semantic Data, Emani et al. developed a middleware system that translates diverse front-end query modalities (like SQL and Python) into a single, language-agnostic intermediate representation (Substrait) before compiling them into optimized target execution queries.

SQL Query    ───┐
                ├──► [ Substrait Intermediate Representation ] ──► Optimized DAX Query
Python Query ───┘

The Fundamental Trade-off

While the interlingua approach offers elegant scaling properties, it introduces a significant engineering trade-off: the problem of representation depth.

As Krauwer et al. observed in Transfer in a Multilingual MT System, the depth of an intermediate representation directly dictates its language independence. A shallow representation is easy to parse but retains language-specific biases, making transfer to distant languages difficult. Conversely, a truly universal, deep semantic representation eliminates language bias but is incredibly complex to design and maintain—a challenge also noted by Abeille et al. in Using Lexicalized Tags for Machine Translation.

The history of interlingual research is a continuous effort to balance this trade-off: finding intermediate spaces that are abstract enough to decouple languages, yet practical enough to parse and generate efficiently.

Key Foundational Work

Quelques Applications De La Logique a La Semantique Des Langues Naturelles (1969) — contributed 2 concept elements to this direction.
International Conference on Computational Linguistics COLING 1969: Preprint No. 55 (1969) — contributed 2 concept elements to this direction.
A Control Structure For A Question-Answering System (1975) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_interlingua_based_translation_a4b6f48b,
  title = {interlingua-based translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interlingua-based-translation/rdr_a4b6f48b8015432981419899cd1b633f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_interlingua_based_translation_a4b6f48b_v1,
  title = {interlingua-based translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interlingua-based-translation/rdr_a4b6f48b8015432981419899cd1b633f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake