My research interests center around the study of performance constraints on the learning and processing of language. In particular, I am interested in how the integration of multiple partially informative cues may inform the language acquisition process, and how such probabilistic information also constrains adult performance. As a general research strategy, I seek to explain the phenomena in question through the integration of computational modeling, psychological experimentation, and theoretical as well as methodological considerations, letting each line of inquiry inform the others. In the future, I would also like to complement these approaches via collaborations with researchers using cognitive neuroscience techniques, such as ERP, PET or fMRI.
Much of my work has been devoted connectionist models of language (Chater & Christiansen, to appear; Christiansen & Chater, 1992, 1993, 1994), and I am currently editing a special issue of the journal Cognitive Science on ``Models of Human Language Processing: Progress and Prospects'' (with Nick Chater and Mark Seidenberg) bringing together papers from an outstanding group of language researchers. One of my main research projects concerns the learning and processing of recursive syntactic structure. My previous work (Christiansen, 1992; Christiansen & Chater, 1994, in submission) has argued that recurrent neural networks are viable as models of recursive processing because they appear to experience the same processing difficulties as people do. This work is continued in Christiansen (in preparation a), investigating the constraints on various kinds of recursive sentence structures in a connectionist model. Empirical predictions from this model have been successfully tested via a self-paced word-by-word grammaticality judgment task (Christiansen & MacDonald, in preparation). These predictions concern the processing of multiple instances of possessive genitives, multiple PP modifications of NPs, comparisons between the processing of multiple subject relative clauses and the processing of multiple sentential complements, and of grammatical and ungrammatical versions of doubly center-embedded sentences. In addition, MacDonald & Christiansen (in preparation) provides a constraint-based approach to individual variations in sentence comprehension, including a connectionist simulation of the individual differences in the processing of sentences involving subject and object relative clauses found in King & Just (1991). In the near future, I intend to conduct more studies of the constraints on various kinds of center-embedding structures using the same experimental task as above, and to model these and other results (e.g., Gibson & Thomas, 1996) concerning various combinations of S-complements and relative clauses.
A second main research project is the modeling of early speech segmentation, an area in which the encoding of various kinds of sequential information also plays an important role as in the case of syntax above. Earlier work has suggested that recurrent networks constitute a good mechanism for the integration of multiple probabilistic cues in the service of discovering word boundaries in the speech stream (Allen & Christiansen, 1996; Christiansen, in press; Christiansen, Allen & Seidenberg, in press). This work is currently being extended (Christiansen & Allen, in preparation) by simulations using input in which reduction following coarticulation and noise caused by multiple realizations of individual segments creates a highly varied input stream, more closely resembling actual speech than the citation forms used in previous computational models. Future work will test empirical predictions derived from the simulation work and develop a detailed model fitting data from a number of developmental experiments concerning early speech segmentation (e.g., Morgan & Saffran, 1995; Newsome & Jusczyk, 1995).
Another project on my research agenda has been the development of a neural network simulator that allows for the modeling of both comprehension and production within the same recurrent network (Christiansen & Chater, in submission). This simulator permits further computational studies within the two projects above, investigating ways in which production and comprehension may interact and influence each other both during acquisition and adult processing.
The above work has fostered an alternative view of language acquisition, offering a re-appraisal of the poverty of the stimulus arguments by focusing on innate, but not necessarily linguistic, constraints on learning and processing (Seidenberg, Allen, Christiansen & MacDonald, in preparation). This perspective also suggests that the evolution of language may have been constrained by learning and processing mechanisms existing prior to the emergence of language (Christiansen, in preparation b). A learning based explanation of head direction uniformity derived from this theory of language evolution was provided via connectionist simulations (Christiansen & Devlin, in press).
Last modified: April 1st, 1997.
Please email suggestions/errors to mhc27@cornell.edu |