TL;DR
This paper introduces an empirical method to automatically identify and translate technical terms for machine translation (MT) systems. By analyzing untagged text corpora, it extracts frequently used word sequences and assigns them syntactic categories, ultimately building a domain-specific MT dictionary that significantly improves translation quality.
What Problem Does This Paper Address?
Machine translation (MT) systems often struggle with technical terminology. The meaning and usage of technical terms can differ significantly from their literal interpretations, leading to mistranslations. While corpus-based approaches exist for identifying these terms, many require pre-tagged corpora or focus narrowly on specific linguistic structures like noun phrases. This paper aims to bridge this gap by proposing a method that can identify a wider range of technical terms from readily available untagged text and integrate them into an MT system.
Methods
The proposed method consists of three main phases:
- Retrieving Local Patterns: This phase uses n-gram statistics and an entropy-based measure to identify recurrent word sequences (collocations) that are likely to be technical terms. The entropy calculation helps distinguish meaningful phrases from random word occurrences.
- Identifying Syntactic Categories: Since the retrieved patterns are just word sequences, this step assigns them syntactic categories. This is achieved by using simplified part-of-speech (POS) tags for individual words and then matching these tag sequences against predefined POS templates.
- Making Translation Patterns: Finally, the identified technical terms and their syntactic categories are converted into bilingual translation patterns. The paper notes that for Japanese, this often involves using the English term directly (in katakana) followed by a functional verb like "suru" (to do).
The system then uses these generated translation patterns to build a domain-specific dictionary for a pattern-based MT system.
Datasets
The paper uses a computer manual as its primary dataset for training and evaluation. This corpus contains 167,023 words across 22,041 sentences. No new benchmarks are introduced; the focus is on applying the method to an existing corpus.
Key Insights and Ideas
- Untagged Corpora are Sufficient: The method's ability to work with untagged text is a significant advantage, making it more practical and resource-efficient than methods requiring pre-tagged data.
- Entropy as a Signal for Technical Terms: Using entropy to measure word distribution proves effective in filtering out random word sequences and identifying meaningful technical phrases.
- Simplified POS Tagging Works: The authors found that a simplified set of POS tags, combined with POS templates, is sufficient for identifying syntactic categories of technical terms, even when words are used in non-standard ways.
- Domain-Specific Dictionaries are Crucial: The experimental results strongly suggest that incorporating domain-specific technical terminology into MT systems leads to substantial improvements in translation quality.
Key Results
The experiment demonstrated significant improvements when using the generated domain dictionary:
- Translation Improvement: Translations using the domain dictionary improved by 57.1% in parsing and word selection compared to translations without it.
- Pattern Coverage: The top 10 POS patterns accounted for 49.4% of all identified POS patterns, indicating that the retrieved strings tend to concentrate in specific syntactic structures.
- Syntactic Category Identification: The method successfully identified syntactic categories for a large portion of the retrieved strings, with nouns (N) and noun phrases (NP) being the most common categories, which is reasonable for technical terms.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- POS Template Dependency: The POS templates are noted to be dependent on the features of n-gram strings.
- Potential for More Precise Parsing: While POS templates are effective, a dedicated parser might offer more precise syntactic category identification. However, the authors argue that for limited input patterns, POS templates offer better portability.
- Improving Translation Pattern Generation: The paper suggests that there is room for improvement in how translation patterns are generated. Future work could involve introducing techniques for finding translational equivalents from bilingual corpora.
- Categorizing Uncategorized POS Patterns: The authors plan to investigate the availability of POS patterns that are not categorized into any specific syntactic category.
Summary updated