Coreference Annotation Scheme and Relation Types for Hindi

Coreference Annotation Scheme and Relation Types for Hindi Vandan Mujadia, Palash Gupta, Dipti Misra Sharma FC Kohli Center on Intelligent Systems (KCIS), IIIT-H, India vmujadia@gmail.com, palash.gupta03@gmail.com, diptims@gmail.com Abstract This paper describes a coreference annotation scheme, coreference annotation specific issues and their solutions through our proposed annotation scheme for Hindi. We introduce different co-reference relation types between continuous mentions of the same coreference chain such as ‘Part-of’, ‘Function-value pair’ etc. We used Jaccard similarity based Krippendorff‘s’ alpha to demonstrate consistency in annotation scheme, annotation and corpora. To ease the coreference annotation process, we built a semi-automatic Coreference Annotation Tool (CAT). We also provide statistics of coreference annotation on Hindi Dependency Treebank (HDTB). Keywords: Hindi, Coreference, Annotation Scheme 1. Introduction There has been considerable research for coreference annotation on various languages (English(Hovy et al., 2006), French(Mitkov et al., 2000), Spanish(Recasens et al., 2007), Czech(Nedoluzhko et al., 2013), etc), on diverse domains like newspaper texts, bio-medical journals, etc. Coreference annotation is a time consuming, challenging and an expensive task. To overcome these challenges, various coreference an- notation tools have been developed like CorefDraw (Harabagiu et al., 2001), GATE (Cunningham et al., 2002), PALinkA (Orasan and Sb, 2000), MMAX2 (Müller and Strube, 2001) and BART (Versley et al., 2008). All these provide text based visualization for annotation. Also there has been considerable work on coreference type relations, like, (Recasens et al., 2010) presented a typology of Near-Identity Relations and motivated the need for a middle ground category be- tween identity and non-identity in the coreference task. For Indian languages, co-referentially annotated corpora are few in numbers and mostly for Hindi. As stated in (Dakwale et al., 2012), most schemes which are meant to be for English are not applicable for Indian languages due to their free word order. (Dak- wale et al., 2012) presents an anaphora annotation scheme and applied it on Hindi Dependency Treebank for limited set of pronominal categories, particularly for concrete anaphors. (Dakwale et al., 2012) used key-value pair attributes on anaphor chunk-head to represent its referent(s). Compared to English coref- erence annotation (and its representation), very little work has been done on Indian languages. This paper tries to fill that gap by describing our coreference annotation scheme for Indian Languages with Hindi examples. We also discuss issues related to coreference annotation specific to Hindi such as distribution of markable span, level of annotation, concept of chain, type relation between mentions(referential entities), full and partial membership of mentions in a chain, multi-chain membership of mentions in chains and their representation in SSF(Bharati et al., 2007). We point out several issues in the earlier annotation schemes and describe a consistent solution to those issues. Co-referentially annotated corpora were, traditionally, not annotated to capture the degree of relations between the referential entities. Our scheme also includes the degree of relation between continuous referential entities of the same chain into different relation types. This includes rela- tions such as ‘Part-of’, ‘Instance-of’, ‘Function-value pair’ etc. These relation types would help various ap- plications like question-answering, …

Content truncated to 500 words