Abstract
Identifying term variations in Japanese web search queries is a significant challenge due to the language's complex writing system. This paper presents a statistical model built on the MART algorithm to identify a broad range of variants, including spelling mistakes, script alternates, transliterations, and abbreviations. By mining user session and click-through logs, the system generates candidates and learns semantic similarity features. The proposed method achieves 70% precision with over 60% recall, representing a 38% error reduction over baseline methods.
The Challenge of Japanese Query Variation
In web search, query expansion helps users find relevant results by including related terms. While this often involves simple typos or stemming in English, Japanese presents a unique hurdle. The Japanese writing system employs four distinct scripts: Hiragana, Katakana, Kanji, and the Roman alphabet. A single concept can be represented in numerous valid ways across these scripts.
Traditional edit distance metrics often fail because variations may look entirely different on the surface while being phonetically identical. Furthermore, Kanji characters can have multiple pronunciations depending on context. The scope of variations handled by this research is broad, ranging from simple spelling mistakes to complex semantic relations.
Methodology: Mining User Behavior
The researchers framed term variation identification as a binary classification task. Their approach relies on two primary pillars: candidate generation from logs and discriminative modeling.
Smart Candidate Generation
Rather than relying on exhaustive dictionary lookups, the system mines user logs to find potential variations:
- Session Logs: Identifying "re-writes" where a user issues a second query within a short timeframe that leads to a click.
- Click-Through Logs: Using a Random Walk algorithm on a bipartite graph of queries and URLs. If different queries frequently lead to the same URL, they are likely variations.
Discriminative Modeling with MART
The core of the system is the Multiple Additive Regression Trees (MART) algorithm. The model utilizes 39 features, including:
- Phonetic Similarity: Levenshtein distance applied to Romanized (phonetic) versions of queries.
- Semantic Similarity: Likelihood ratios derived from query logs.
- Character Type Features: Indicators for specific script transitions (e.g., Kanji to Katakana).
Key Results and Performance
The MART-based model significantly outperformed linear models like Maximum Entropy, suggesting that the interactions between linguistic features and script types are non-linear and complex.
The system achieved a precision of 70% at a recall level of approximately 60%. A key finding was the importance of semantic features; removing them caused a substantial spike in error rates. This confirms that user behavior data—the "wisdom of the crowd"—is a vital signal for resolving linguistic ambiguity in search.
Conclusion
This research demonstrates that combining phonetic string similarity with behavioral data from search logs is highly effective for identifying Japanese term variations. While challenges remain—particularly with extreme abbreviations and Kanji ambiguity—the 38% error reduction marks a significant step forward in improving search intent recognition for multi-script languages. Future work will explore character-dependent edit operations to further refine the model's accuracy.
Summary updated