A Linguistic Investigation into Unsupervised DOP

Unsupervised Data-Oriented Parsing models (U-DOP) represent a class of structure bootstrapping models that have achieved some of the best unsupervised parsing results in the literature. While U-DOP was originally proposed as an engineering approach to language learning (Bod 2005, 2006a), it turns out that the model has a number of properties that may also be of linguistic and cognitive interest. In this paper we will focus on the original U-DOP model proposed in Bod (2005) which computes the most probable tree from among the shortest derivations of sentences. We will show that this U-DOP model can learn both rule-based and exemplar-based aspects of language, ranging from agreement and movement phenomena to discontiguous contructions, provided that productive units of arbitrary size are allowed. We argue that our results suggest a rapprochement between nativism and empiricism.