MASALA: Modelling and Analysing the Semantics of Adpositions in Linguistic Annotation of Hindi

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5696–5704 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5696 MASALA: Modelling and Analysing the Semantics of Adpositions in Linguistic Annotation of Hindi Aryaman Arora, Nitin Venkateswaran, Nathan Schneider Georgetown University Washington, D.C., USA {aa2190, nv214, nathan.schneider}@georgetown.edu Abstract We present a completed, publicly available corpus of annotated semantic relations of adpositions and case markers in Hindi. We used the multilingual SNACS annotation scheme, which has been applied to a variety of typologically diverse languages. Building on past work examining linguistic problems in SNACS annotation, we use language models to attempt automatic labelling of SNACS supersenses in Hindi and achieve results competitive with past work on English. We look towards upstream applications in semantic role labelling and extension to related languages such as Gujarati. Keywords: SNACS, semantic parsing, Hindi 1. Introduction Case markers express semantic roles, describing the re- lationship between the arguments they apply to and the action of a verb. Adpositions (prepositions, postposi- tions, and circumpositions) further express a range of semantic relations, including space, time, possession, properties, and comparison. Languages have different strategies for encoding these kinds of semantic relations. Hindi–Urdu1 uses a case-marking system along with a large postposition inventory (Kachru, 2006; Koul, 2008). Idiosyncratic bundling of case and adpositional relations poses prob- lems in many natural language processing tasks for Hindi, such as machine translation (Ratnam et al. 2018, Jha 2017, Ramanathan et al. 2009, Rao et al. 1998) and semantic role labelling (Pal and Sharma 2019, Gupta 2019). Many models for these tasks rely on human- annotated corpora for training data, such as the one cre- ated for the Hindi–Urdu PropBank (Bhatt et al., 2009), and in Kumar et al. (2019). The study of adposition and case semantics in corpora is also useful from a compar- ative/typological linguistic perspective, in comparing and categorizing the encoding of such relations across languages. To that end, we release a completed Hindi corpus annotated for adposition and case semantic labels using the SNACS formalism (Schneider et al., 2018a, 2020). We approach the problem of automatic tagging of these labels using a variety of language models and explore what these models learn. Drawing on parallel SNACS corpora in English, German, Mandarin, and Korean, we compare strategies for encoding semantic roles across languages. 1Hindi and Urdu are two registers written in two different scripts of a single language (usually called ‘Hindi–Urdu’ or ‘Hindustani’) with a largely identical grammar. While our corpus is in Hindi in the Devanagari script, the linguistic portions of our work (e.g. annotation guidelines) are applicable to Urdu as well. 2. Background Hindi is a language of India, of the Indo-Aryan branch of the Indo-European family, and one of the best- resourced South Asian languages for research in nat- ural language processing and computational linguistics (Joshi et al., 2020). Hindi has a small number of core case markers as well as a large class of adpositions for signalling semantic relations. …

Content truncated to 500 words