Generalized Template Matching for Semi-structured Text

Conventional template matching for named entity recognition on book-length text strings is generalized by allowing search phrases to capture distant tokens. Combined with word-type tagging and format variants (alternative name/date formats), a few initial templates (class—search-phrase—extract-phrase triples) can label most of the significant tokens. The program then uses its book-length statistics of tag-label associations to suggest candidate text for further template construction. The method serves as a preprocessor for error-free extraction of semantic relations from text obeying explicit semi-structure constraints. On three sample books of genealogical records, an F-measure of over 0.99 was achieved with less than 3 hours’ user time on each book.