Agentic AI Translate: An Agentic Translator Prototype for Translation as Communication Design

Technical Summary: Agentic AI Translate

Problem Statement
The paper identifies a fundamental shift in the translation landscape driven by Large Language Models (LLMs). While statistical and neural systems have largely solved the "what" of translation (lexical and grammatical fidelity), achieving high fluency and idiomaticity is now a baseline "Must-Be" quality. The frontier of value has moved to the "how": register, audience design, voice, stance, and cultural framing. Current machine translation paradigms typically treat these communicative dimensions as implicit or apply them via post-editing. The paper argues that the vocabulary of Translation Studies (TS)—concepts like skopos, register, and equivalence—has become an instruction code for generative AI, yet existing systems lack an architecture that operationalizes these theoretical frameworks as explicit, structured inputs rather than unstructured text conversion.

Methodology
The authors present Agentic AI Translate, a research prototype that operationalizes the thesis that translation is "communication design" rather than text conversion. The system replaces the standard text-in/text-out paradigm with a structured, four-stage agentic cycle preceded by an interactive specification phase.

1. Interactive Specification Phase: Before generation, the user engages in a model-assisted dialogue to compose a structured translation brief. This brief is grounded in TS concepts (skopos, audience, register, genre) and covers ten canonical sections (e.g., "Things to preserve," "Things to localise"). The system enforces a "lock" mechanism, ensuring no translation occurs without an explicit, user-endorsed specification.
2. The Four-Stage Agentic Cycle:
   • Identification: An LLM call extracts situational analysis (skopos, audience, register, etc.) from the source text into a structured JSON format.
   • Prompting: A deterministic Python script assembles the final prompt, integrating the user's locked specification, the identification JSON, reference materials, and document-level memory.
   • Generation: A single LLM call (temperature 0.3) produces the draft translation.
   • Verification: An LLM-as-judge evaluates the output against the same locked specification using the GEMBA-MQM protocol. It identifies error spans, assigns MQM categories (Accuracy, Fluency, Terminology, etc.) and severities, and computes a deterministic score. If the score falls below a threshold, the error list is fed back as actionable instructions for a refinement loop (bounded at two iterations).
3. Document-Level Coherence: To handle long inputs, the system employs a "DelTA-lite" memory system. This maintains a running proper-noun ledger, a bilingual summary of the document's progression, and immediate context from the previous chunk, injecting these into subsequent generation prompts to ensure terminological and tonal consistency.
4. Reference Materials: The system accepts four categories of reference inputs (Glossaries, Paired Examples, Parallel Texts, Style Guides), which are injected into the specification proposal, generation prompt, and verification stages.

Key Contributions
The paper's primary contribution is architectural and conceptual, not empirical.

• Executable Embodiment of Theory: The system provides a working codebase that translates the metalanguage of Translation Studies (specifically the frameworks of Reiss, Nord, House, and Vermeer) into machine-readable instructions. It demonstrates that theoretical concepts can serve as direct operational constraints for LLMs.
• The "Spec-Driven" Architecture: The prototype enforces a workflow where the translator acts as a "designer" (defining conditions) and a "verifier" (judging functional fit), rather than a manual drafter. The separation of the situational analysis (Stage 1) from generation allows for inspectable, legible artifacts, making the translation process transparent.
• Open-Source Implementation: The system is released as an open-source Python prototype (approx. 1,200 lines) using the Anthropic SDK and Streamlit, requiring no fine-tuning, vector databases, or GPU resources. It relies entirely on prompt engineering and flow control.

Results
The paper explicitly states that no empirical validation has been conducted.

• The authors do not present comparative MQM scores, user studies, or quantitative data demonstrating that this architecture produces better translations than unstructured prompting or standard MT.
• The "results" are limited to informal observations on multi-paragraph test inputs, where the DelTA-lite memory successfully maintained consistency for named entities (e.g., Natsume Soseki vs. Kushami-sensei) across chunks.
• The paper acknowledges that the current verification stage, which uses the same model family for both generation and judging, is a limitation due to potential self-preference bias.

Significance and Claims
The paper claims significance as a conceptual proof-of-concept. It argues that the "automation paradox" (where generating text is easy, but ensuring it fulfills a specific purpose is hard) can be resolved by treating the vocabulary of Translation Studies as the instruction set for AI.

• Philosophical Shift: The system operationalizes the view that the translator's role in the GenAI era is the design of conditions (voice, register, audience) and the verification of communicative purpose, rather than sentence-by-sentence drafting.
• Transparency: By decomposing the process into distinct, logged stages (Identification, Prompting, Generation, Verification), the system makes the "black box" of LLM translation legible and inspectable, which is presented as a pedagogical and research necessity.
• Future Direction: The authors position this prototype as the foundation for future research, specifically the need for controlled comparative studies to validate whether explicit specification leads to measurable gains in "attractive quality" (style, tone, cultural fit) and the development of structured JSON schemas to replace free-form markdown specifications.