Institute of Information Science, Academia Sinica

A substantial proportion language consists of patterns that escape the detection and representation of both symbolic and statistical computational techniques. This territory of language is populated by what Paul Kay has referred to as “partially productive lexico-grammatical patterns” (1997). These are not suitable for capture as n-grams or for storage as multiword wholes because they are not frozen chunks. And since they are of such limited productivity, neither are they readily captured by general combinatorial rules of grammar. This massive portion of linguistic knowledge has thus long represented “a pain in the neck” for computational linguists (Sag et al 2002). In this talk, I present an approach we have developed to capture this ‘in between’ territory with a new sort of machine-tractable language model and to make it accessible and navigable for end users (Wible and Tsao 2010; Tsao and Wible 2013). I also demonstrate our implementation of this approach called StringNet (nav.stringnet.org).