Using Information Content to Predict Phone Deletion

Different languages present us with rather different likelihoods of deleting consonants. The very common deletion processes that target /t/ and /d/ in various English dialects have been discussed in detail in linguistic research, but English does not exhibit other well-known phonological processes such as /s/ deletion, a process that is very common in a number of Romance languages. A curious linguist might therefore want to know what leads to the observed in-language typology of deletion processes in any language. This paper traces some of the reasons for the in-language typology of deletion processes in English to phone informativity, a concept first introduced in Cohen Priva & Jurafsky (2008). Phone informativity tries to approximate the usefulness of recognizing a phone for what it is: how useful it is for language users to understand that some segment in a sequence of segments they hear is a /t/, for instance, rather than some other phone. This approximation is done by assessing how helpful each phone usually is in word recognition, given a corpus of spoken language data. To evaluate whether phone informativity plays a significant role in determining the in-language typology of deletion processes in English, the paper begins by examining the deletion rates of English stops. The deletion patterns illuminate several shortcomings of theoretical frameworks such as markedness and underspecification, as well as data-driven approaches such as uniform information density in explaining the observed data. The magnitude of influence that phone informativity has on phone deletion is evaluated using corpus-based experiments that control for phonological and phonetic biases, as well as different information theoretic explanations. Finally, a discussion of functional and non-functional explanations in phonology places informativity as a bridge between the two.