Clausal parsing helps data-driven dependency parsing: Experiments with Hindi

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1279–1287, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Clausal parsing helps data-driven dependency parsing: Experiments with Hindi Samar Husain*, Phani Gadde*, Joakim Nivre† and Rajeev Sangal* * Language Technologies Research Centre, IIIT-Hyderabad, India. † Department of Linguistics and Philology, Uppsala University, Sweden. {samar, phani.gadde}@research.iiit.ac.in, joakim.nivre@lingfil.uu.se, sangal@mail.iiit.ac.in Abstract This paper investigates clausal data-driven de- pendency parsing. We first motivate a clause as the minimal parsing unit by correlating in- ter- and intra-clausal relations with relation type, depth, arc length and non-projectivity. This insight leads to a two-stage formulation of parsing where intra-clausal relations are identified in the 1st stage and inter-clausal rela- tions are identified in the 2nd stage. We com- pare two ways of implementing this idea, one based on hard constraints (similar to the one used in constraint-based parsing) and one based on soft constraints (using a kind of pars- er stacking). Our results show that the ap- proach using hard constraints seems most promising and performs significantly better than single-stage parsing. Our best result gives significant increase in LAS and UAS, respec- tively, over the previous best result using sin- gle-stage parsing. 1 Introduction There has been a recent surge in addressing pars- ing for morphologically rich free word order lan- guages such as Czech, Turkish, Hindi, etc. These languages pose various challenges for the task of parsing mainly because the syntactic cues neces- sary to identify various relations are complex and distributed (Tsarfaty et al., 2010; Ambati et al., 2010; Nivre and McDonald, 2008; Tsarfaty and Sima'an, 2008; Seddah et al., 2009; Gadde et al., 2010; Husain et al., 2009; Eryigit et al., 2008). There has also been a lot of interest in building ensemble systems (Zeman and Zabokrtsky, 2005; Sagae and Lavie, 2006) and parser stacking (Ni- vre and McDonald, 2008; Martins et al., 2009) to improve the overall parsing accuracy by combin- ing the strengths of multiple parsers. In this paper, we formulate clausal parsing as a two-stage setup where intra-clausal relations are identified in the 1st stage and inter-clausal rela- tions are identified in the 2nd stage. We attempt to find out whether this two-stage parsing ap- proach that has earlier been successful in con- straint-based systems for parsing Hindi (Bharati et al., 2009) can also benefit data-driven parsing approaches (Nivre et al., 2006), and compare two ways of implementing this idea, one based on hard constraints (similar to the one used in con- straint-based parsing), and one based on soft constraints (using a kind of parser stacking; (Ni- vre and McDonald, 2008). We show that one of the ways in which clausal parsing helps is by better learning of features that leads to improved label accuracy for Hindi. In particular we show that ambiguous case markers (or suffixes) that appear with many relations can be disambiguated successfully. We also show that the setup reduc- es many of the traditional MaltParser (Nivre et al., 2006) errors (McDonald and Nivre, …

Content truncated to 500 words