Towards a Neuro-Cognitive Model of Human Sentence Processing

Towards a Neuro-Cognitive Model of Human Sentence Processing Kei Yoshimoto and Shigeru Sato Graduate School of International Cultural Studies, Tohoku University Kawauchi 41, Aoba-ku, Sendai 980-8576, Japan kei@linguist.jp satos@mail.tains.tohoku.ac.jp Abstract. A formal sentence processing system is proposed which simulates different event- related potential (ERP) elicitation between sentences with and without unambiguous case marking. The electroencephalographical data are based on German subordinate clauses and Japanese sentences. As a formal framework we adopt Dynamic Syntax (Kempson et al. 2001), which enables incremental update of information by underspecifying tree information. This is augmented with default syntactic and semantic specifications which reflect shallow but efficient human sentence processing. 1 Introduction Recent developments in non-invasive measurement of brain activities have paved the way for directly observing how linguistic information is processed in the human brain. However, no attempts have hitherto been made to model human sentence comprehension based on the data provided by those methods from a unified formal point of view. This paper proposes a parser which behaves in a similar manner to humans in eliciting different event-related potentials (ERPs) depending on whether NPs are unambiguously case marked or not. The crucial role therein is performed by syntactic underspecification in incremental processing, which is made possible by Dynamic Syntax (Kempson et al. 2001) adopted as our formal linguistic theory. 2 ERPs in Sentence Processing ERPs, i.e. diverse kinds of electrophysiological response to stimuli, is known to reflect the higher human cognitive capacities. They are also acknowledged as a useful tool to investigate the process of language understanding in the human brain. The most typical ERPs associated with specific steps of processing are outlined below: (i) An early left-anterior negativity (ELAN) is elicited approximately 150-200 ms after a crit- ical word onset which is involved with a word category error. (ii) A left-anterior negativity (LAN) is observed between 100-500 ms after presenting stimuli with morpho-syntactic errors. (iii) An N400 is a negativity which peaks about 400 ms after the word onset. This is typically ascribed to a failure to integrate a word meaning to the preceding context. (iv) A P600, a late centro-parietal positivity occurring between 600-1000 ms, is associated with the integration of different types of information and completion of sentence processing. Mainly based on these electrophysiological data, Friederici (1999, 2002) proposes a neuro-cognitive model of sentence understanding which is composed of three processing steps. An erroneous input sentence is considered to invoke an excessive processing load at at least one of these steps, giving rise to (an) ERP(s) explained in (i) to (iv) above. Phase 1 (100-300 ms): The initial syntactic structure is built up based on the identification of word category. An error in this phase causes an ELAN. 21 Phase 2 (300-500 ms): Lexical-semantic and morpho-syntactic information is processed in parallel without interaction, leading to the assignment of thematic roles. The first subprocess, if over- loaded with an erroneous input, elicits a LAN, the latter an N400. Phase 3 (500-1000 ms): Lexical-semantic and morpho-syntactic information with other types of information is …

Content truncated to 500 words