Cognate Mapping - A Heuristic Strategy for the Semi-Supervised Acquisition of a Spanish Lexicon from a Portuguese Seed Lexicon

Cognate Mapping — A Heuristic Strategy for the Semi-Supervised Acquisition of a Spanish Lexicon from a Portuguese Seed Lexicon Stefan Schulza,b Korn´el Mark´ob,c Eduardo Sbrissiaa Percy Nohamaa Udo Hahnc aMaster Program in Health Technology, Paran´a Catholic University, Curitiba, Brazil bDepartment of Medical Informatics, Freiburg University Hospital, Germany cComputational Linguistics Research Group, Jena University, Germany http://www.coling.uni-freiburg.de/ Abstract We deal with the automated acquisition of a Spanish medical subword lexicon from an al- ready existing Portuguese seed lexicon. Using two non-parallel monolingual corpora we de- termined Spanish lexeme candidates from Por- tuguese seed lexicon entries by heuristic cog- nate mapping. We validated the emergent lex- ical translation hypotheses by determining the similarity of fixed-window context vectors on the basis of Portuguese and Spanish text cor- pora. 1 Introduction Medical language presents a unique combination of challenges for language engineering, with a fo- cus on applications such as information retrieval, text mining and information extraction. Document collections – on the Web or in clinical databases – are usually very large and dynamic. In addi- tion, medical document collections are truly multi- lingual. Furthermore, the user population which ac- cess medical documents are really diverse, ranging from physicians and nurses to laypersons, who use different jargons and sublanguages. Therefore, the simplicity of the content representation of the docu- ments, as well as automatically performed intra- and interlingual lexical mappings or transformations of equivalent expressions, become crucial issues for an adequate machine support. We respond to these challenges in terms of the MORPHOSAURUS system (an acronym for MORPHeme TheSAURUS). It is centered around a new type of lexicon, in which the entries are sub- words, i.e., semantically minimal, morpheme-style units (Schulz and Hahn, 2000). Intralingual as well as interlingual synonymy is then expressed by the assignment of subwords to concept-like equivalence classes. As subword equivalence classes abstract away from subtle particularities within and between languages, and reference to them is achieved via a language-independent code system, they form an interlingua characterized by semantic identifiers. Compared to relationally richer, e.g., WORDNET based, interlinguas as applied for cross-language in- formation retrieval (CLIR) (Gonzalo et al., 1999; Ruiz et al., 1999), we use a rather limited set of semantic relations and pursue a more restrictive ap- proach to synonymy. In particular, we restrict our- selves to the specific sublanguage used in the con- text of the medical domain. Our claim that this interlingual approach is useful for the purpose of cross-lingual text retrieval and categorization has al- ready been experimentally supported (Schulz et al., 2002; Mark´o et al., 2003). The quality of cross-lingual indexing fundamen- tally depends on the underlying lexicon and the- saurus. Its manual construction and maintenance is costly and error-prone. Therefore, machine- supported lexical acquisition techniques increas- ingly deserve attention. Whereas in the medical do- main parallel corpora are only available for a lim- ited number of language pairs, unrelated (i.e., non- parallel, non-aligned) corpora might provide suffi- cient evidence for cognate identification, at least in languages which are closely related. In this paper, we …

Content truncated to 500 words