Introduction
Professional translation involves more than simple word-for-word substitution. While technical terminology often maps directly between languages, general language is characterized by "indirect lexical transfers"—shifts in part-of-speech or metaphor that ensure a text sounds natural to native speakers. For instance, a literal translation of the French il faillit échouer ("he faltered to fail") is awkward; a human translator might instead suggest "he was on the brink of failure" or "failure loomed."
Finding these idiomatic equivalents is a significant bottleneck for human translators. Standard Statistical Machine Translation (SMT) relies on parallel corpora, which are often limited in scope and struggle with radical transformations. This paper presents a "translator’s amenuensis," a digital assistant designed to mine creative, indirect translations from large-scale comparable corpora.
Methodology: The Indirect Transfer Pipeline
The system utilizes a multi-step pipeline to identify non-literal equivalents by leveraging distributional similarity rather than direct translation rules:
- Expansion and Mapping: The process begins with "contextual descriptors" provided by the user. The system identifies words with similar distributions in the source language (SL) using cosine similarity of collocation vectors. These terms are then mapped to the target language (TL) using high-quality bilingual dictionaries.
- Target Refinement: On the TL side, the system generates further similarity classes to capture idiomatic variations. For multi-word queries, a consistency check is performed against a massive TL database to ensure suggestions actually co-occur in real-world usage (e.g., as bi-grams).
- Ranking and Semantic Filtering: Candidates are ranked by distributional proximity to the original query. To mitigate noise, a multilingual semantic tagger filters out suggestions that do not share the same semantic field, ensuring that evaluative polarity (e.g., positive vs. negative) is preserved.
Experimental Results
The system was evaluated using 200 million words of English news and 70 million words of Russian news as comparable corpora. Performance was measured against standard dictionaries (ORD) and an SMT baseline (GIZA++).
As shown in the results, the proposed system achieved a recall of 21.9% for 2-word descriptors, significantly outperforming GIZA++ (13.9%) and the Oxford Russian Dictionary (6.7%). Notably, the system demonstrated high robustness; while the SMT baseline's performance plummeted when moving from news to interview genres (dropping to 3.4%), the proposed system's performance remained stable at 19.5%.
The Impact of Semantic Filtering
A critical component of the system is the semantic filter, which refines the "useful noise" generated by distributional similarity. While similarity clusters often contain related but incorrect terms, the filter helps elevate the most relevant candidates to the top of the list.
The inclusion of the semantic filter significantly improved ranking quality. For 1-word descriptors, the average rank of a correct solution improved from 21.6 to 11.3. For 2-word descriptors, the rank improved to 26.7, ensuring that high-quality, human-style solutions typically appear on the first screen of results for the translator.
Conclusion
By shifting the focus from parallel to comparable corpora, this research demonstrates that idiomatic variation can be modeled as a "family of transformations" rather than rigid word-to-word rules. While the system currently relies on high-quality bilingual dictionaries as a bridge, it provides a powerful decision-support tool that offers creative suggestions often overlooked by standard machine translation models.
Summary updated