Automating construction work: Data-Oriented Parsing and constructivist accounts of language acquisition
Our world is filled with a vast array of objects and their relations and properties. Human infants face the magnificent task of processing experiences with the outside world in such a way that they can later on respond in an adequate manner when similar, but non-identical experiences present themselves. We can call this processing “learning” and an important question studied throughout the cognitive sciences is how humans do it. One domain for which this question is especially important, is that of linguistic systems of communication, as the complexity and open-endedness found therein has led many to believe that some architectural aspects of the cognitive representations of the phenomenon are not learn- able from positive linguistic input alone. This assumption has led to the conclusion that these representations are innately present in the language learner and that there are cognitive mechanisms innately tuned or geared towards acquiring a language (such as a “principles and parameters” approach, cf. Wunderlich 2007). With the linguistic nativist conviction comes the assumption that the representations used are of a fairly abstract nature - after all, the learner would have to be able to acquire any of the thousands of languages being used around the world. Nativists, especially within the Minimalist framework (Chomsky 1993), further support this assumption by pointing to the economy of representation as a driving factor for having a system that is as compact as possible. Importantly, the innate knowledge is part of a mental module pertaining only to language. That is, the representations the learner starts with are domain- specific.
Another school of thought, the empiricist one, states the child does not come equipped with inborn, domain-specific knowledge concerning the architecture or properties of a communication system to be acquired. The acquisition of the complexities of linguistic structure are explained (as far as they are not theory-internal concerns that depend on one’s preconception of the cognitive representation, cf. Tomasello 2003: 7) from experience, through domain-general structure-finding mechanisms such as categorization, schematization and social understanding. Importantly, these mechanisms and representational biases have to exist in the learner’s mind prior to the acquisition of a language system.
Hence, usage-based theorists cannot be argued to believe in a blank-slate learner. The crucial difference from a linguistic nativist position is that the mechanisms and biases are not specific to language, but are shared with other cognitive domains because they either are functions of how the brain in general works (e.g., working memory, entrenchment processes, abstraction) or are part of known evolved cognitive modules (e.g., the figure-ground distinction from the visual system, notions of object permanence). With the nativist position being the dominant one for the last decades, researchers of the empiricist bent face the task of showing that there are flaws in the empirical observations or subsequent inferential processes leading to linguistic nativist conclusions. At the same time, it is crucial that empiricist theorists develop a substitutive, positive, account of language acquisition through experience and domain-general skills. Important work showing flaws in nativist reasoning and providing a novel account has been done. Construction Grammar, in many of its flavors (Langacker 1989; Goldberg 1995; Croft and Cruse 2004), as well as non-constructivist work in language acquisition (Peters 1983) shows how the nativists’ assumed divisions between the core and periphery of the grammar, meaning and the grammar, and linguistic competence and performance cannot be maintained, and at the same time presents an empiricist account of how the architecture and content of linguistic representation emerges as an interaction between a multitude of factors. The work of Tomasello and colleagues (Tomasello 2003) has shown how understanding other people’s (communicative) intentions is crucial for and supportive of acquiring a language, demonstrating how a thitherto overlooked aspect of human cognition solves some of the nativist arguments against acquiring a grammar from experience, as well as presenting a coherent explanation of linguistic development.
In this paper, we would like to add something to the developing usage- based constructivist narrative. This contribution is in part a methodological enrichment and in part an account of the possible domain-general cognitive mechanisms behind the acquisition of the grammatical structures. We believe that computational modeling is an important means for providing us with important insights in the theoretical perspective. First of all, it forces us to translate our fuzzy and imprecise natural-language definitions into extremely precise computational ones. Although this often means a loss in accuracy of description (we will have to give up on the description of some aspects of natural language for our model to be understandable), it provides a gain in the testability of certain claims. Using a well-defined model, then, we can assess claims pertaining to the architecture and content of the representations, the processing mechanisms and the timescales on which these operate.
The computational model we present in this paper is Data-Oriented Parsing or DOP (Scha 1990; Bod 1998, Bod, Scha & Sima’an 2003), and its instantiations
Unsupervised Data-Oriented Parsing (U-DOP) and Meaningful Unsupervised Data-Oriented Parsing (pDOP). The Data-Oriented family of models addresses the question how processing complex, structured exemplars, such as linguistic experiences, may lead to a cognitive system by means of which a language user can assign structure to novel exemplars. As such, it is not a theory about the content of representations, but rather a discovery procedure (for learners and linguists alike) for cognitively useful structured representations. In the following sections, we explain the basic ideas behind the models in greater detail, link it to constructivist assumptions and show how the diverse models can be applied to questions about the acquisition of grammar.