Introduction: The Challenge of Structural Divergence in EBMT
Example-Based Machine Translation (EBMT) relies on a knowledge base of previously translated sentence pairs to generate new translations. However, traditional alignment methods that attempt to link full syntactic trees for both languages often fail when encountering "free translations." In languages with significant structural differences, such as Portuguese and Chinese, translators frequently interpret phrases rather than translating them literally.
Research indicates that approximately 63.4% of examples in standard Portuguese grammar corpora exhibit this type of structural divergence. Rigid 1:1 word alignment systems cannot effectively map these pairs, creating a need for a more flexible annotation schema that can handle non-literal correspondences.
The Translation Corresponding Tree (TCT) Framework
To address these limitations, this paper introduces the Translation Corresponding Tree (TCT). Unlike traditional bi-tree alignment, TCT utilizes a single-tree representation based on the source language's syntax. Each node in the TCT is enriched with a "triple sequence interval" that maps the source structure directly to the target string:
- SNODE(n): The specific word index in the source sentence.
- STREE(n): The substring interval of the source sentence covered by the subtree.
- STC(n): The corresponding substring interval in the target sentence.
This approach avoids the "double-error" problem inherent in parsing both languages and allows the system to remain robust even when the target language lacks a high-quality parser.
Handling Non-Literal Translations and Inversions
The TCT generation process involves source parsing, target lexical analysis (word segmentation and POS tagging), and a bottom-up alignment phase using bilingual dictionaries and grammatical heuristics.
A key feature of TCT is its ability to represent structural inversions—cases where the conceptual order flips between languages. When such a divergence is detected, the node is marked with an inversion attribute, visually represented by a horizontal line within the tree structure. This allows the system to capture crossing dependencies and provide a clear reordering roadmap for the translation engine.
Key Insights and Results
The TCT schema successfully represents complex linguistic phenomena, including:
- Many-to-one/One-to-many mappings: Linking single words to multi-word phrases.
- Discontinuous translations: Mapping a single source concept to split parts of a target sentence.
- Non-projective mapping: Managing significant word order changes.
By embracing "free" translation as a standard feature of human communication rather than noise, TCT provides a more realistic and flexible foundation for EBMT systems. While the current workflow involves semi-automated generation followed by manual verification, it significantly improves the representational accuracy of bilingual knowledge bases for structurally diverse language pairs.
Conclusion
The Translation Corresponding Tree offers a streamlined, interval-based mapping solution for machine translation. By focusing on a single-tree architecture and flexible string intervals, it effectively bridges the gap between the rigid requirements of computational linguistics and the fluid nature of human translation.
Summary updated