Department of Humanities and Social Sciences

Indian Institute of Technology Ropar


HSS Seminar


"Processing Hindi Relative Clauses: Implications For Memory Retrieval & Expectation-Based Theories"




Samar Husain


April 22, 2016 (Friday) at 2:30 PM


Venue: Conference Room 2



Cross-linguistically, relative clauses (RCs) have been employed to test opposing predictions made by expectation and memory retrieval theories (e.g. Grodner and Gibson, 2005; Vasishth and Lewis, 2006; Staub, 2010; Bartek et al., 2011; Levy et al., 2013; Jaeger et al., 2015). Broadly, surprisal (an expectation-based theory) predicts processing facilitation for a highly frequent syntactic structure (Levy et al., 2012); it also predicts facilitation at a phrase as the distance between this phrase and its prior dependents increases (Levy, 2008). Memory-based theories (Gibson, 2000; Lewis and Vasishth, 2005), on the other hand, predict a processing slowdown at a phrase under such conditions. Hindi has three types of RCs: right-extraposed, embedded and correlatives. So far no study has investigated 'online' processing of these RC types and its implication for the above theories. In experiment 1 we investigate processing in correlatives using the self-paced reading (SPR) paradigm. The relative pronoun in the correlative clause is associated with a demonstrative phrase in the matrix clause. When the distance between the relative pronoun and the demonstrative phrase was increased (long condition), reading time at the region following the demonstrative phrase was faster compared to the short condition. This goes against both expectation and memory-based theories. Such a facilitation could point to syntactic processes, e.g. reference resolution, that fall outside the purview of the theories under consideration. Experiments 2-4 investigate embedded and right-extraposed RCs. Previously, Kothari (2010), using an offline judgement study, showed a global preference for embedded RCs over right-extraposed RCs. This preference was shown to be modulated by the length of RC and the length of the matrix VP (VP length). Such a pattern is consistent with memory-based theories and can be understood as dependency length minimization for the head noun and the relative-pronoun/RC-verb. This contrasts with expectation-based theories that predict a preference for right-extraposed RCs due to their higher frequency count in a Hindi treebank (cf. Levy et al., 2012). In two SPR studies we found no support for memory-based theories, instead the results favored expectation-based account. In particular, RC-verb was read faster when RC length was long compared to short. A similar pattern was found for VP length. In addition, participants were surprised to encounter short right-extraposed RC and took longer to process the short condition compared to long. Finally, reading time at the RC-pronoun was slower in the case of embedded RC compared to right-extraposed RC. However, an offline judgement task (replicating Kothari, 2010) found support for memory-based theories; there was a global preference for embedded RCs over right-extraposed RCs, right-extraposed RC was only preferred when RC length was long. And, when the length of matrix VP increased RC extraposition was dispreferred.   


Brief Bio-sketch of the speaker


Samar Husain is an Assistant Professor in the Department of Humanities and Social Sciences, IIT Delhi, India. He finished his PhD in Computational Linguistics from IIIT-Hyderabad, India in 2011. From 2011-2014 he was a post-doctoral researcher at the Vasishth lab, University of Potsdam, Germany. His broad research interests are human sentence processing, natural language parsing, natural language modeling, and dependency grammar. His current research investigates questions concerning the predictive nature of human sentence processing, working-memory constraints and the interaction between the two. In addition to the research on sentence comprehension, he has also investigated reading processes in Hindi and Urdu using the eye-tracking methodology.


  back to seminars….