Automatic processing of multilingual medical terminology: applications to thesaurus enrichment and cross-language information retrieval

Hervé Déjean, Éric Gaussier, Jean-Michel Renders, Fatiha SadatArtif. Intell. Medicine 2005 · DOI · OpenReview

TL;DR

This paper introduces novel methods for automatically extracting bilingual medical lexicons from both parallel and comparable corpora. These lexicons significantly improve the enrichment of medical thesauri like MeSH and enhance the performance of cross-language information retrieval (CLIR) in the medical domain, outperforming existing general and specialized bilingual resources.

What Problem Does This Paper Address?

The paper tackles the challenge of bridging language barriers in the specialized domain of medical information. Specifically, it addresses the need for accurate and comprehensive bilingual lexicons to facilitate two key tasks:

  1. Thesaurus Enrichment: Medical terminologies, such as MeSH, are crucial for organizing and accessing medical knowledge. However, these thesauri are often monolingual or have limited multilingual support. Manually enriching them with terms from other languages is time-consuming and expensive.
  2. Cross-Language Information Retrieval (CLIR): Users often need to search for medical information in a language different from the one they are querying in. Existing CLIR systems rely on bilingual resources, but these are often incomplete, not domain-specific, or lack sufficient coverage for specialized terminology.

The core problem is the scarcity of high-quality, domain-specific bilingual lexicons that can effectively link medical terms across languages, hindering both knowledge organization and information access.

Methods

The authors propose and evaluate several methods for bilingual lexicon extraction, categorized by the type of corpus used:

  • Parallel Corpora:
    • Single Word Lexicons: They employ a statistical word-alignment model based on iterative proportional fitting (IPFP) to estimate word co-occurrence frequencies in aligned sentences. This results in probabilistic translation lexicons.
  • Term Lexicons: For multi-word terms, they first extract candidate terms using morpho-syntactic patterns and then align them. A key innovation is a term extension strategy that allows for incorporating prepositional phrases (like "of") to better capture complex term translations, using an association score based on single-word alignments.
  • Comparable Corpora:
    • Standard Approach: This method builds context vectors for source and target words, translates target context vectors using a general bilingual dictionary, and then computes similarity using the cosine measure.
  • Thesaurus-Based Approach: This novel method leverages a multilingual thesaurus (MeSH-DMD) as a pivot. It models the probability of a translation pair using concept classes from the thesaurus, considering both "complete-n search" (using top n concept classes) and "subtree search" (using hierarchical thesaurus information).
  • Combined Approach: They combine the standard and thesaurus-based methods using a mixture model (EM algorithm) to leverage the strengths of each.

Datasets

  • MuchMore Corpus: A bilingual collection of approximately 9,300 medical abstracts from Springer Verlag, totaling about 1.2 million words. This corpus is used in two ways:
  • As a comparable corpus (all 9,300 abstracts).
  • As a parallel corpus (a subset of 5,500 abstracts with better sentence alignment).
  • General Bilingual Resources:
    • ELRA German/English dictionary (approx. 50,000 entries).
  • MeSH (Medical Subject Headings) and its German version DMD, extracted from UMLS.
  • Reference Lexicons: Manually extracted reference lexicons for evaluating lexicon extraction (1000 single words and 150 multi-word terms from the parallel corpus, and 150 candidate terms for term alignment evaluation). A held-out reference lexicon of 1,200 pairs was used for estimating mixture weights in the comparable corpus approach.

Key Insights and Ideas

  • Comparable Corpora are Valuable: The paper demonstrates that even without direct sentence-level alignment, comparable corpora can yield valuable bilingual lexicons, especially when combined with thesaurus information.
  • Thesaurus Structure is Key: The "subtree search" method highlights the importance of exploiting the hierarchical structure of thesauri for more accurate and efficient concept selection in lexicon extraction.
  • Term Extension Improves Accuracy: Allowing for the extension of candidate terms (e.g., including prepositions) significantly improves the alignment of complex medical terms.
  • Combination is Powerful: Combining different lexicon extraction methods (standard, thesaurus-based) leads to substantial improvements in lexicon accuracy and coverage.
  • Domain-Specific Lexicons Boost CLIR: Extracted domain-specific bilingual lexicons provide a significant performance boost for CLIR tasks compared to general resources.

Key Results

  • Thesaurus Enrichment: The extracted lexicons are accurate enough for semi-automatic enrichment of thesauri like UMLS.
  • CLIR Performance (Parallel Corpus): Using lexicons extracted from the parallel corpus for CLIR improved average precision from 47% to 54% compared to monolingual retrieval.
  • CLIR Performance (Comparable Corpus): The bilingual lexicon approach using comparable corpora improved average precision from 22% to 40% compared to the baseline.
  • Lexicon Evaluation (Terms): For term alignment, retaining the top 10 candidates yielded an F1-score of 80%, with precision being higher than recall. In nearly 90% of cases, a correct translation was found within the top 10 propositions.
  • Lexicon Evaluation (Comparable Corpus Models): Combined models achieved an F1-score of up to 84%, significantly outperforming individual models. The subtree search was more efficient (fewer classes, faster) while achieving comparable performance to complete searches.
  • Lexicon Coverage: The comparable corpus approach achieved a coverage ratio of about 92% for query terms.

Limitations and Future Work

  • Limitations:
    • The comparable corpus used was "ideal" and biased, meaning results might differ with less favorable comparable corpora.
  • The methods for handling coordinated nouns (e.g., "Academies and Institutes") are not fully covered due to extraction difficulties.
  • The current work focused on bilingual lexicons of single words for comparable corpora; extending to terms is a future direction.
  • Future Work:
    • Conducting experiments with less favorable comparable corpora.
  • Extending the approach to bilingual lexicon extraction of terms from comparable corpora.
  • Investigating the problem of searching collections where only parts are parallel.
  • Further exploring the combination of different approaches for improved accuracy and coverage.

Summary updated

BibTeX
@article{dejean2005automatic,
  author = {Hervé Déjean and Éric Gaussier and Jean-Michel Renders and Fatiha Sadat},
  title = {{Automatic processing of multilingual medical terminology: applications to thesaurus enrichment and cross-language information retrieval}},
  journal = {Artif. Intell. Medicine},
  year = {2005},
  doi = {10.1016/j.artmed.2004.07.015},
}
Figure