Abstract
Pivot translation, the process of translating between two languages via a common intermediate language (usually English), frequently suffers from "semantic drift." This phenomenon leads to inconsistent meanings across the translation chain. This paper introduces a multiagent coordination layer that leverages automatically generated multilingual "triples" from existing bilingual dictionaries to enforce semantic consistency. Evaluation shows that this approach improves translation quality by approximately 40% for noun-heavy sentences without the need for extensive new parallel corpora.
Introduction
Direct translation services for every possible language pair are often unavailable or prohibitively expensive. Consequently, the industry relies on pivot translation, using a bridge language. However, this cascading approach introduces three critical errors:
- Inconsistency: A single source word is translated differently across various sentences.
- Asymmetry: Translating from Language A to B and back to A does not return the original word.
- Intransitivity: Meaning drifts across the chain (e.g., Japanese ketten [defect] -> English fault -> German Schuld [guilt]).
Traditional solutions require massive parallel corpora, which are scarce for rare language pairs. This research proposes a "meta-level" control to fix these "black box" services from the outside.
Methodology
The proposed Context-Based Coordination approach consists of two primary components: dictionary-based triple generation and a multiagent architecture.
Generating Multilingual Equivalent Terms (Triples)
To maintain semantic integrity, the system identifies true equivalents across Language A, B, and C. By applying graph theory to existing bilingual dictionaries, the system identifies two types of "triangles":
- Loop Triangles: A circular mapping (A→B, B→C, C→A) indicating a high likelihood of a shared specific sense.
- Transition Triangles: A transitive mapping (A→B, A→C, B→C) that confirms the relationship between terms.
Multiagent Architecture
The system wraps existing translation engines in a coordination layer, avoiding the need to modify the underlying services.
- Coordinator Agent: Manages the "context," defined as a set of allowed equivalent terms for the current text.
- Translation Agents: These agents interface with standard translation services. If a service returns a word outside the "allowed" list (signaling drift), the agent replaces it with the pre-validated equivalent from the generated triples.
Evaluation and Results
The researchers evaluated the system using four major bilingual dictionaries and a massive web corpus of 470 million Japanese sentences. Subjective human evaluation was conducted on a test set of 100 sample sentences.
Key findings include:
- High Coverage: The system generated 21,914 trilingual triples, covering 58% of all nouns in the web corpus.
- Quality Improvement: Translation quality improved for 41% of sentences.
- Score Increase: On a 5-point scale, the average quality score increased by 0.47 points.
- Drift Prevention: The system successfully blocked semantic drift, such as the "fault to responsibility" error, by identifying that the terms did not belong to the same semantic triangle.
Discussion and Conclusion
This approach demonstrates that context can be effectively modeled as multilingual constraints. By forcing translation engines to adhere to specific n-tuples of words, semantic wandering is prevented. This "black-box friendly" method allows for improvements to proprietary engines like Google Translate without requiring source code access.
Current limitations include a focus on nouns and a reliance on the quality of available bilingual dictionaries. Future work aims to extend this coordination to verbs and adjectives and to handle more than three languages simultaneously.
Summary updated