Machine Translation SPI with multi-level context¶
Status¶
Proposed
Date¶
2025-12-27
Context¶
OmegaT’s existing Machine Translation (MT) Service Provider Interface (SPI) is designed for simple, segment-by-segment translation without persistent context. However, modern foundation models significantly benefit from “multi-level prompting,” where translations are informed by project-wide instructions (e.g., style guides, domain knowledge) in addition to the source segment itself.
To support LLM MT and quality checking with explanatory feedback via external plugins, the OmegaT core SPI needs to be enhanced to handle project-level context. This allows OmegaT to remain decoupled from specific implementations while processing with project-level guidance.
Decision¶
We will extend the OmegaT core SPI to support multi-level context. This includes:
Defining a mechanism to provide project-level context to MT and Issue providers.
Updating the
IMachineTranslationandIIssueProviderinterfaces to accept this context.Ensuring the implementation of specific LLM logic remains in optional, external plugins.
Proposed Changes¶
1. New Interface: IContextSupplier¶
A new interface org.omegat.gui.exttrans.IContextSupplier will be introduced. This interface allows providers to query project-level information.
2. Enhanced IMachineTranslation SPI¶
The org.omegat.gui.exttrans.IMachineTranslation interface will be updated with:
setContextSupplier(IContextSupplier supplier): Allows OmegaT core to pass a context supplier to the MT engine.Methods to handle more granular “suggestions” and metadata.
3. Enhanced IIssueProvider SPI¶
The org.omegat.gui.issues.IIssueProvider interface will be updated to support context-aware issue detection.
This enables connector plugins to perform quality checks that consider project-level instructions and provide detailed rationale for identified issues.
4. Leverage Existing IMTGlossarySupplier¶
We will continue to leverage IMTGlossarySupplier to provide terminology context, which can be combined with project-level context to produce more contextually consistent output.
Consequences¶
Positive¶
✅ Clean Core: OmegaT core remains agnostic of LLM-specific logic, APIs, and dependencies. ✅ Extensibility: Third-party developers can create specialized LLM plugins (local or remote) using a standardized SPI. ✅ Independent Release Cycle: Plugins can be updated frequently to match evolving LLM capabilities without requiring core OmegaT releases. ✅ Contextual Quality: Enables higher-quality translations and better issue detection by providing the necessary context to plugins.
Negative¶
SPI Complexity: Increases the complexity of the MT and Issue Provider interfaces.
Coordination: Requires careful design to ensure that different plugins (e.g., an MT plugin and an Issue Provider plugin) can share configuration and context effectively if they are part of the same suite.
Optional Note: Reference Implementation (MT connector Plugin)¶
(This section describes a potential design for an external plugin leveraging the new SPI; it is not part of the core OmegaT proposal.)
Some users have expressed interest in using modern models for machine translation, specifically local MT models (to ensure data privacy and avoid recurring costs). Frameworks provide OpenAI-compatible APIs that can be leveraged on a user’s local machine.
The proposed SPI changes enable an external connector plugin with the following architecture:
1. MT Connectors and OpenAI Compatibility¶
The plugin can provide a generic connector for Chat Completions-compatible APIs and dedicated connectors for other services.
2. Project-Level Configuration (omegat/mt.xml)¶
The plugin can store its own project-specific settings (model selection, project prompt) in a separate file to ensure backward compatibility and keep the main project file clean.
3. Prompt Construction¶
The plugin constructs prompts by merging:
System Prompt: Base instructions for the MT task.
Project Prompt: Domain and style instructions from
IContextSupplier.Glossary Context: Terminology from
IMTGlossarySupplier.Segment Prompt: The actual source segment to translate.
4. Technical Architecture¶
Here is a high-level overview of the MT connector support architecture:
┌──────────────────────────────────────┐ ┌──────────────────────────────────────┐
│ OmegaT Core │ │ Optional Plugin (JAR) │
│ │ │ │
│ ┌────────────────────────────────┐ │ │ ┌────────────────────────────────┐ │
│ │ MT & Issue Management │ │ │ │ UI: API Configuration │ │
│ └─────────────┬──────────────────┘ │ │ └─────────────┬──────────────────┘ │
│ │ │ │ │ │
│ │ (1) Calls │ │ ▼ │
│ ▼ │ │ ┌────────────────────────────────┐ │
│ ┌────────────────────────────────┐ │ │ │ Properties │ │
│ │ Core SPI Layer │ │ │ │ ──▶ (Storage: omegat/mt.xml) │ │
│ │ │ │ │ └─────────────┬──────────────────┘ │
│ │ │ │ │ │ │
│ │ IMachineTranslation ──────────┼──┼──────────┼───▶ MTConnector (MT SPI) │
│ │ IIssueProvider ──────────┼──┼──────────┼───▶ Issue Provider (Issue SPI) │
│ │ │ │ │ │ │
│ │ IContextSupplier ◀─────────┼──┼──────────┼──── Prompt Construction │
│ │ IMTGlossarySupplier ◀─────────┼──┼──────────┼──── (Context-Aware) │
│ │ │ │ │ │ │
│ └─────────────┬──────────────────┘ │ │ ┌─────────────▼──────────────────┐ │
│ │ │ │ │ Shared connection library │ │
│ │ (2) Provides │ │ │ (Protocol Handling) │ │
│ ▼ │ │ └─────────────┬──────────────────┘ │
│ ┌────────────────────────────────┐ │ └────────────────┼─────────────────────┘
│ │ MT suggestions and checks │ │ │
│ └────────────────────────────────┘ │ ▼
└──────────────────────────────────────┘ ┌──────────────────────────────────────┐
│ API (Local/Remote models) │
└──────────────────────────────────────┘