Bayesian, inflection class system, lexemes


In this paper we use a Bayesian, agent-based model to explore the emergence of allomorph distributions in inflection class (IC) systems. It has long been understood that irregular inflection occurs mainly among high token frequency lexemes because high frequency leads to word-specific learning, allowing certain lexemes to resist analogical pressure. Over time, these lexemes become ‘stranded’ in low type frequency classes as less frequent lexemes shift class membership. Crucially, these classes partly overlap but do not collapse with high type frequency classes, and as a result detract from speakers’ ability to predict a word’s inflection class membership. Stump and Finkel (2013) extract an empirical generalization that they call Marginal Detraction (MD): low type frequency classes contribute more to the complexity of an IC system than high type frequency classes do. This is represented as negative slopes in Figure 1 based on data from Sims and Parker (2016). Here, IC system complexity (operationalized as conditional entropy) is defined as the average uncertainty associated with one inflected form of a lexeme, given knowledge of one other form of the same lexeme.

Original Publication Citation

Parker, Jeff, Robert Reynolds and Andrea D. Sims. “A Bayesian investigation of factors shaping the network structure of inflection class systems”. Poster at the 1st Meeting of the Society for Computation in Linguistics, Salt Lake City, January 4-7, 2018.

Document Type

Conference Paper

Publication Date



Society for Computation in Linguistics







University Standing at Time of Publication

Assistant Professor

Included in

Linguistics Commons