The book by Friederici ( 2017) provides an excellent overview of a major direction in the theory of the language organ. See Section 6 for discussions of potential future work that connects the two areas. Our work can be seen as largely orthogonal to the related work described above, as we attempt to bridge granular neuronal mechanics with the study of complex cognitive processes such as syntax. Rather …neural network models form a description at Marr’s algorithmic level” (Frank et al., 2019). In a survey paper of cognitive models of language that use ANNs, Frank summarizes that “In spite of the superficial similarities between artificial and biological neural networks …these cognitive models are not usually claimed to simulate processing at the level of biological neurons. In this framework, it is nontrivial to implement a single elementary step of syntactic processing, such as recording the fact that a particular input word is the sentence’s subject, whereas in previous work such actions are built into the framework’s primitives. ![]() In all previous work, parsers are written in a high-level programming language, whereas we focus on whether a simple parser can be implemented by millions of individual neurons and synapses through the simulation of a realistic mathematical model of neuronal processing. ![]() The present paper differs from these works in key ways. ( 2013), or the parser of Lewis and Vasishth ( 2005), which is constructed in ACT-R, a high-level meta- model of cognitive processing. Much work attempts to achieve the properties in (b) while maintaining (a), that is, neural and psycholinguistic predictiveness, for example, the PLTAG parser of Demberg et al. Exemplars of this line of work are Jurafsky’s use of probabilistic parsing to predict reading difficulty (Jurafsky, 1996), the surprisal- based models of Hale, Levy, Demberg, and Keller (Hale, 2001 Levy, 2008 Demberg and Keller, 2008a), and the strictly incremental predictive models of Demberg and Keller (Demberg and Keller, 2008b, 2009). ![]() See Keller ( 2010) for a summary of the psycholinguistic desiderata of (b), as well as a discussion of evaluation standards for (a). Such work focuses chiefly on (a) understanding whether high-level parsing methods can be used to predict psycholinguistic data (such as reading-time or eye-tracking data) and neural data (e.g., fMRI and ECoG data from linguistic experiments) and (b) developing parsing methods that have specific, hallmark, experimentally established cognitive properties of human syntactic processing (most importantly incrementality of parsing, limited memory constraints, and connectedness of the syntactic structures maintained by the parser). There is a rich line of work in computational psycholinguistics on cognitively plausible models of human language processing.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |