Tracing and Reducing Lexical Ambiguity in Automatically Inferred Grammars

Abstract

While the automated creation of machine-readable grammars is a valuable resource for linguists who wish to work with these grammars for linguistic hypothesis testing, the complexity of developing a system capable of creating such grammars presents a number of obstacles. One obstacle faced by the system this work belongs to is the excessive amount of ambiguity that the output grammars license. Because the system takes input from a diverse collection of resources, the glossing practices between datasets can vary, making it necessary to employ generalized approaches to determining the syntactic function of certain glosses. Such generalized approaches expand the kinds of data that can be used, but at the risk of introducing spurious ambiguity. This study investigates linguistically informed ways to reduce the ambiguity of the inferred lexicons and morphological systems in these grammars. By decreasing the presence of non-inflecting lexical rules and imposing stricter requirements on which roots may be entered into the lexicon, the lexical and morphological ambiguity of the system was reduced without an overwhelming loss of coverage.