Improving the Multilingual User Experience of Wikipedia Using Cross-Language Name Search

Raghavendra Udupa, Mitesh M. KhapraNAACL 2010

TL;DR

This paper tackles the challenge of non-English Wikipedia users struggling to find relevant English articles due to language barriers, especially when searching for names. They introduce GEOM-SEARCH, a novel cross-language name search system that uses a geometric representation of names to enable users to search in their native language and find English Wikipedia articles. This approach significantly improves the user experience for ESL/EFL users.

What Problem Does This Paper Address?

Wikipedia, while a massive repository of information, suffers from a significant linguistic asymmetry: the vast majority of content is in English, with a much smaller presence in other languages. This creates a problem for users whose native language is underrepresented. These users often need to find information in English Wikipedia but struggle to formulate effective search queries, particularly when dealing with names. Standard English spell-checkers are insufficient because they lack coverage of names and are not designed to handle the spelling variations introduced by a user's native language pronunciation and orthography. This paper aims to bridge this gap by allowing users to search for English Wikipedia articles using names expressed in their native languages.

Methods

The core of the GEOM-SEARCH system lies in transforming the cross-language name search problem into a geometric search problem.

  • Language/Script Independent Geometric Representation: The authors propose mapping names from different languages and scripts into a common, low-dimensional Euclidean feature space. This is achieved by first creating language-specific feature vectors (e.g., using character bigrams) and then using Canonical Correlation Analysis (CCA) to learn linear transformations (matrices A and B) that project these vectors into a shared space. This common space allows for direct comparison of names regardless of their original language or script.
  • Geometric Search: Once names are represented as points in this common feature space, searching for similar names becomes a nearest neighbor search problem. The system leverages the Approximate Nearest Neighbors (ANN) algorithm to efficiently find the closest English Wikipedia names to a given query name.
  • Multi-word Name Similarity: To handle multi-word names, the system first finds the nearest English neighbors for each word in the query. It then constructs a weighted bipartite graph where edges represent the cross-language similarity between individual words. The overall similarity between two multi-word names is then determined by the maximum weighted bipartite matching in this graph, penalizing disproportionate lengths.

Datasets

  • Training Data: For CCA training, approximately 15,000 name pairs were used for each native language. For the transliteration system (TRANS-SEARCH baseline), around 15,000 single-word name pairs were used.
  • Test Data: For evaluation, test sets of 1000 multi-word name queries were created for each of the six native languages (Russian, Hebrew, Hindi, Kannada, Tamil, and Bangla). The "gold standard" for evaluation consisted of the most relevant Wikipedia article for each query and all its redirect titles.

Key Insights and Ideas

  • Geometric Representation as a Unifier: The key insight is that by mapping names into a shared geometric space, the problem of cross-language similarity becomes a tractable geometric search problem, bypassing the need for explicit transliteration or complex linguistic rules.
  • CCA for Cross-Lingual Alignment: Canonical Correlation Analysis proves to be a powerful tool for learning the transformations needed to align different language representations into a common semantic space, effectively capturing the underlying semantic similarity of names.
  • ANN for Scalability: The use of Approximate Nearest Neighbors is crucial for handling the massive scale of Wikipedia's name database, enabling fast retrieval of relevant results.
  • Multi-word Similarity via Graph Matching: The approach to combine single-word similarities into a multi-word similarity score using weighted bipartite matching is an elegant way to account for the structure and word-level correspondences in names.

Key Results

  • Significant Improvement: GEOM-SEARCH consistently outperformed the transliteration-based baseline (TRANS-SEARCH) across all six tested languages.
  • High Recall: On average, the relevant English Wikipedia page was found within the top 2 results produced by GEOM-SEARCH for all native languages.
  • Efficiency: While not explicitly detailed for GEOM-SEARCH's average time, the paper highlights that ANN allows for searching millions of titles in a fraction of a second. The TRANS-SEARCH baseline, which relies on edit distance after transliteration, was significantly slower (e.g., 2.39 seconds for Hindi vs. 0.51 seconds for GEOM-SEARCH).
  • Qualitative Validation: Figures 1 and 2 visually demonstrate that names that are transliterations of each other are mapped to nearby points in the common feature space, validating the geometric approach.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the system is effective, it focuses specifically on name search.
  • The performance of Machine Transliteration systems, even state-of-the-art ones, is noted as not always producing correct transliterations, which is a weakness of the baseline but not the core GEOM-SEARCH approach.
  • The paper mentions that for some languages (e.g., Chinese, Japanese, Korean), character unigrams might be a better choice for features, and syllabification might be beneficial, suggesting potential language-specific optimizations.
  • Future Work:
    • The authors suggest that integrating a Machine Translation system with GEOM-SEARCH could further enhance the user experience by translating the retrieved Wikipedia pages back into the user's native language.
  • While not explicitly stated as future work, the discussion on feature selection for different language families hints at potential avenues for further research into language-specific feature engineering.

Summary updated

BibTeX
@inproceedings{udupa2010improving,
  author = {Raghavendra Udupa and Mitesh M. Khapra},
  title = {{Improving the Multilingual User Experience of Wikipedia Using Cross-Language Name Search}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2010},
}
Figure