Parsing errors in Hindi: Investigating limits to verbal prediction in an SOV language

The role of prediction during sentence comprehension is widely acknowledged to be very critical in SOV languages. Robust clause-final verbal prediction and its maintenance have been invoked to explain effects such as anti-locality and lack of structural forgetting. At the same time, there is evidence that these languages avoid increased preverbal phrase complexity due to working-memory constraints. Given the critical role of prediction in processing of SOV languages, in this work, we study verbal predictions in Hindi (an SOV language) to investigate its robustness and fallibility using a series of completion studies. Analyses of verbal completions based on grammaticality (grammatical vs ungrammatical) as well as their syntactic property (in terms of verb class) show, as expected, frequent grammatical completions based on effective use of preverbal nouns and case-markers. However, there were also high instances of ungrammatical completions. In particular, consistent errors were made in conditions with 3 animate nouns with unique/similar case-markers. These errors increased in the face of adjuncts of differing complexity following the preverbal nouns. The grammatical and ungrammatical completions show that native speakers of Hindi posit structures with at most 2 verbal heads and 5 core verbal relations, thus highlighting an upper bound to verbal prediction and its maintenance in such configurations. A rating study confirmed that certain errors found in completion tasks can lead to grammatical illusions. Further, a detailed analysis of the completion errors in such cases revealed that the parser ignores the complete preverbal nominal features of the input and instead selectively reconstructs the input based on their frequency in the language to form illicit parses at the expense of globally consistent parses. Together, the results show that while preverbal cues are effectively employed by the parser to make clause-final structural predictions, the parsing system breaks down when the number of predicted verbs/relations exceeds beyond a certain threshold. In effect, the results suggests that processing in SOV languages is susceptible to center-embeddings similar to that in SVO languages. This highlights the over-arching influence of working-memory constraints during sentence comprehension and thereby on the parser to posit less complex structures.