Introduction
Machine translation (MT) often stumbles on lexical ambiguity—the phenomenon where a single word in a source language possesses multiple distinct meanings in a target language. This paper proposes a statistical approach that bypasses the need for complex hand-coded knowledge bases or manually tagged source corpora. Instead, it leverages the natural statistical patterns of the target language to disambiguate source words.
The Challenge of Lexical Ambiguity
Lexical ambiguity is a major hurdle in Natural Language Processing. For instance, the Hebrew verb lahtom can translate to "sign," "seal," "finish," or "close." Traditional solutions like syntactic rules or source-language statistics often fail because different meanings may share identical grammar or require unavailable tagged data. The authors focus on target word selection, determining the correct translation based on the specific context within the target language.
Methodology: Leveraging the Target Language
The core innovation lies in shifting the statistical lens from the source to the target language. The authors argue that while a word may be ambiguous in the source, its correct translation is often constrained by its collocations in the target language.
The system operates in three stages:
- Relational Representation: Identifying syntactic relations (e.g., verb-object) in the source sentence.
- Mapping to Target Alternatives: Generating all potential translations for the related words.
- Statistical Selection: Consulting a massive, un-tagged target corpus (English) to find which word combination is most frequent in those syntactic roles.
To ensure robustness, the system uses an odds ratio and the "delta method" to calculate confidence intervals, ensuring that the chosen translation is statistically superior to its competitors.
Evaluation and Results
The model was tested using a 150-million-word English corpus (from sources like The Washington Post and Associated Press) and evaluated on Hebrew and German texts.
- Hebrew-to-English: The model applied to 70% of ambiguous cases with a precision of 92%.
- German-to-English: The model achieved 59% applicability and 75% precision.
The results demonstrate that using relational statistics significantly outperforms a "naive" baseline of simply choosing the most frequent word in the target language.
Discussion and Limitations
While successful, the approach faces challenges. Data sparseness remains a primary issue; if a specific word pair never appears in the corpus, the model cannot make an informed choice. Additionally, because the model focuses solely on the target language, it may occasionally miss nuances intended in the source language or fail when "world knowledge" (e.g., distinguishing "immigration" vs. "emigration" based on direction) is required. Future work suggests using semantic similarity measures to mitigate data gaps.
Summary updated