Mapping Vocabularies Using Latent Semantics

Abstract

Abstract.Individual users of medical language manifest great variation in the expression of concepts and have difficulty in selecting appropriate terminology when confronted with systems that rely on standardized language, such as MeSH, SNOMED, or ICD, and the special terms sets of systems such as HELP, INTERNIST-I/QMR, and DXplain. Indeed, the need to map natural language into appropriate special terms-as well as the need to map one system's specialized terminology into another's-is one of the problems being addressed by the National Library of Medicine's UMLS System, with its associated information sources maps. The problem is extremely difficult, in part, because such mappings depend on semantic equivalences among terms, not merely the superficial matching of words or phrases.As a general and robust solution to the problem of mapping across vocabularies, we implemented a version of "latent semantic indexing", taking terms from different vocabularies as the 'documents' to be retrieved by natural-language expressions of concepts, taken as 'queries'. In one of several experiments testing our approach, for example, we selected approximately 225 terms each from the INTERNIST-I/QMR, PTXT, and META-1 vocabularies corresponding to clinical findings under the physical exam. We constructed a source matrix of associations between the findings and all the 'words' they contained, supplemented with word-level synonyms and related terms. The resulting source rectangular matrix was approximately 650x3000. Under singular value decomposition, this was reduced to a compressed space of at most 650 dimensions. The performance of the reduced space as a "latent semantic" map of the source domain was evaluated by processing phrases, intended to be interpreted as clinical findings, as term-vectors, projected into the reduced space. In calculating the projections, we considered only the 150 most significant dimensions. Mappings to concepts (i.e., standardized terms) were determined by taking the cosine-distance measure of the vector to all the term-points in the reduced space. As an example of the results, the phrase "stomach discomfort worse after eating" scored as follows ( The power of the approach is that it does not depend on explicit, declarative semantic representations or on word-for-word correspondences among terms; and that multiple vocabularies can be represented side-by-side.Communications should be directed to:

Latent semantic analysis

“…Evans, Handerson, Monarch, Pereiro, Delon, and Hersh (1998) explored the use of LSA and CL‐LSA to associate terms with specialized indexing concepts. Experiments associating English terms with Spanish medical concepts appear to be promising, but no comparisons to monolingual approaches are reported.…”

“…The CL‐LSA method has been applied to many languages, including English‐French (Dumais et al, 1998; Landauer & Littman, 1990), English‐Spanish (Carbonell et al, 1997; Evans et al, 1998; Oard & Dorr, 1998), English‐Greek (Berry & Young, 1995), Portuguese‐English (Orengo & Huyck, 2002) and English‐Japanese (Jiang & Littman, 2001; Landauer, Littman, & Stornetta, 1992; Mori, Kokubu, & Tanaka, 2001). It has also been applied to language triples such as English‐French‐Spanish and English‐French‐German where three‐way document‐aligned corpora were available (Littman, Jiang, & Keim, 1998; Rehder et al, 1997).…”

Chinese-Japanese cross language information retrieval

“…For example, retrieving relevant news articles published in both languages from the Internet. It is understood that several mathematical techniques, like Han character clustering and dimensionality reduction techniques (Evans et al, 1998) can augment and automate the process of finding associations among the Han characters within and across the CJK languages. The vector space model is also flexible for the adjustment of weighting scheme.…”

“…We are aware that such manual modification is not feasible for a large heterogeneous document collection. Dimensionality reduction techniques, fike LSI (Evans at al., 1998;Rehder et al, 1998) or Han character clustering are the potential solutions to automatically discover associations among Hart characters. etc., and they are similar to those used in Chinese.…”

Japanese–English cross-language information retrieval integrating query and document translation methods

“…Thesaural semantic classes [9,19] and Latent Semantic Indexing [3,5,13] are methods that have been employed with this objective.…”