Local Coherence and Case-Marker Exchange Cause Parsing Errors in Hindi Apurva Apurva@hss.iitd.ac.in Indian Institute of Technology, Delhi Samar Husain Samar@hss.iitd.ac.in Indian Institute of Technology, Delhi June 23, 2020 Abstract Prediction in verb-final languages is considered to be very robust (Konieczny, 2000; Konieczny and D¨ oring, 2003; Vasishth, 2003; Levy and Keller, 2013; Husain et al., 2014). In such languages (e.g., Hindi), case-markers can help in making the upcoming verbal prediction precise, leading to processing facilitation (Levy and Keller, 2013). However, recent investigations in Hindi (Apurva and Husain, 2016, 2018a,b,c,d, 2019) have found that predictions are fallible under some circumstances leading to parsing errors. For example, Apurva and Husain (2016) found that parsing errors due to incorrect predictions increases in sentences with non-canonical word order or where multiple case-markers share the same lexical form. However, the cause for such parsing errors still remains unexplored. This work fills this gap in the literature. In order to do so, we first replicated Apurva and Husain (2016) with a modified design. Results show that two major causes of parsing errors are local coherence and case-marker exchange due to similarity of case-markers. The work highlights the role of certain parsing heuristics and working memory constraints while comprehending rare structure in a head-final language like Hindi. Keywords: Local coherence, Sentence comprehension, Prediction, Head-final languages, Cloze task, Parsing error, Encoding interference 1. Introduction The human language comprehension system is predictive in nature. On the basis of degree of pre- dictive capacity the sentence processing literature divides languages into two categories: • Head initial/medial languages • Head final languages On average, speakers of a head-final language are exposed to more instances of verb-final struc- tures compared to a speaker of a head-initial lan- guage. This exposure makes their parser adapt to the head-final structure (e.g., clause final verb structure) of the language. This adaptability has been argued to lead to better prediction and main- tenance of the clause final verb in head-final lan- guages. This is unlike the speakers of head me- dial/initial language where they get to see the verb quite early in a sentence. Therefore as far as the clause final verb prediction is concerned head-final languages are considered to be better in terms of verb prediction. This assumption is widely used in sentence processing literature to explain some well known effects such as anti-locality (Konieczny, 2000; Konieczny and D¨ oring, 2003; Vasishth and Lewis, 2006; Levy and Keller, 2013), and lack of forgetting effect (Gibson and Thomas, 1999; Va- sishth et al., 2010), etc. However, the assumption of robust prediction in head-final languages was never validated system- atically. Some recent studies (Apurva and Husain, 2016, 2018a,b,c,d, 2019) checked this assump- tion via several Cloze tasks in Hindi. The focus of these studies was to check the robustness of the predicted parse (grammatical vs ungrammat- ical) during comprehension in Hindi (a head-final language). The results of these cloze tasks showed that predictions in Hindi were quite robust. However, they also found a large number of parsing er- rors due to incorrect predictions as well. In par- ticular, Apurva and Husain (2016) showed that the overall percentage of valid parse during the cloze task was close to 50%. The remaining 50% of the parses are either incorrect (∼ 20%) 1