TL;DR
This paper introduces a novel two-stage approach for acquiring bilingual terminology from comparable corpora, which are collections of texts on similar topics but not direct translations. This approach significantly improves Cross-Language Information Retrieval (CLIR) by not only finding accurate term translations but also by effectively translating entire queries as phrases.
What Problem Does This Paper Address?
Building comprehensive bilingual dictionaries is a labor-intensive and costly process. For many language pairs, especially those with fewer resources, obtaining high-quality bilingual terminology is challenging. This lack of robust bilingual lexicons directly impacts the effectiveness of Cross-Language Information Retrieval (CLIR), where users search for information in one language using queries in another. Existing CLIR systems often struggle with query translation ambiguity, leading to poor retrieval results. This paper aims to address this by developing a method to automatically extract bilingual terminology from readily available "comparable corpora" and then use this terminology to improve phrase-level query translation for better CLIR performance.
Methods
The core of the proposed method is a two-stage translation model for bilingual terminology acquisition from comparable corpora:
- Bi-directional Term-by-Term Translation: The model first learns translation probabilities from source to target language (S→T) and then from target to source language (T→S) using similarity vectors derived from context vectors of terms.
- Merging and Disambiguation: These two models are merged to create a more robust, bi-directional translation model. This process leverages common term pairs that appear in both directions and combines their similarity scores.
Following terminology acquisition, two key techniques are applied to enhance CLIR:
- Linguistics-based Pruning: This step filters translation candidates based on linguistic knowledge, specifically Part-of-Speech (POS) tags. It ensures that translated terms have similar grammatical roles (e.g., noun to noun, verb to verb), reducing noise.
- Phrasal Translation (Re-scoring): Instead of translating queries word-by-word, the system treats queries as phrases. It generates multiple translation candidates for each term and then re-scores these phrase-level translations using various techniques. These re-scoring methods include:
- WWW-based: Checking the frequency of translated phrase combinations on the World Wide Web.
- Test Collection-based: Using existing CLIR test collections and retrieval systems.
- Comparable Corpora-based: Analyzing co-occurrence statistics within the comparable corpora.
- Interactive Mode: Allowing user feedback to refine translations.
Datasets
- Comparable Corpora:
- News articles from Mainichi Newspapers (1998-1999) for Japanese.
- Mainichi Daily News (1998-1999) for English.
- CLIR Evaluation:
- NTCIR-2 test collection (specifically for Japanese-English language pairs).
Key Insights and Ideas
- Comparable Corpora are Rich Sources: The paper demonstrates that even non-parallel, comparable corpora can be effectively used to build bilingual lexicons, overcoming the scarcity of parallel data.
- Bi-directionality Matters: Learning translation in both directions (S→T and T→S) and merging the results leads to a more accurate and comprehensive bilingual terminology.
- Linguistics Adds Precision: Incorporating linguistic information (like POS tags) significantly prunes noisy translation candidates, improving the quality of the extracted terminology.
- Phrasal Translation is Crucial for CLIR: Moving beyond single-term translation to phrase-level translation of queries is essential for mitigating ambiguity and achieving better CLIR performance.
- WWW as a Linguistic Resource: The World Wide Web can serve as a valuable, large-scale resource for re-scoring translation candidates, leveraging its vast amount of text to assess phrase coherence.
Key Results
The proposed two-stage model, combined with linguistics-based pruning and phrasal translation techniques, showed significant improvements in CLIR:
- Improved Average Precision: The two-stage comparable corpora-based model achieved a +27.1% improvement in average precision compared to a simple one-stage model.
- Further Gains with Pruning: Adding linguistics-based pruning further boosted average precision by +41.7% and +11.5% compared to the simple and two-stage models, respectively.
- Phrasal Translation Boosts Performance: Applying re-scoring techniques for phrasal translation yielded substantial gains: 10.35% (WWW-based), 8.27% (NTCIR-based), and 3.08% (comparable corpora-based) improvements over a hybrid approach without phrasal translation.
- Overall Effectiveness: The integrated approach demonstrated positive effects on query translation/expansion and document retrieval for the Japanese-English language pair.
Limitations and Future Work
- Limitations:
- The paper focuses on the Japanese-English language pair. While the approach is presented as generalizable, its performance on other language pairs is not explicitly evaluated.
- The effectiveness of the interactive mode for phrasal translation is discussed conceptually but not quantitatively evaluated in the experiments.
- Future Work:
- Integration of other linguistics-based techniques.
- Combination with transliteration for special phonetic alphabets (like Japanese katakana).
- Further exploration of interactive modes to optimize phrasal translation and user interaction.
Summary updated