📄 Towards an Annotated Corpus of Discourse Relations in Hindi

Summary

This workshop paper takes first steps toward a Hindi counterpart of the Penn Discourse Treebank (PDTB). PDTB guidelines were designed for English, and Hindi has free word order (default SOV) and rich, optional case marking, so it is unclear how well they transfer. No ready-made inventory of Hindi connectives existed. A list built from intuition and from translating PDTB connectives turned out to be incomplete, so the authors examined real text: 9 short stories totalling about 8,000 words. From this sample they catalogued explicit connectives by syntactic type and noted where Hindi diverges from the PDTB. They then compared each connective type with its treatment in the existing sentence-level Hindi dependency annotation. Some relations, such as single subordinating conjunctions, are already fully captured there. Others are only partly covered, including paired conjunctions, non-finite subordinators and semantically vague markers. A lexically anchored discourse layer would fill these gaps. The work lays the groundwork for annotation guidelines and a future Hindi Discourse Relation Bank, and adds to cross-linguistic research on connectives.

Key points

• Data and approach: A pilot study of 9 short stories (about 8,000 words), using the PDTB's lexically grounded method and Arg1/Arg2 naming. It covers explicit connectives only; implicit connectives, senses and attribution are left for future work.
• Broader coverage than the PDTB: Five connective groups are identified:
  • subordinating conjunctions, often paired across clauses (e.g., "because…therefore");
  • sentential relative pronouns that link eventualities;
  • non-finite subordinators and postpositions;
  • coordinating conjunctions;
  • discourse adverbials.
• Discovery during annotation: Some functions only appeared in real data. For example, the complementizer ki can also act as a temporal subordinator.
• Frequencies and constraints: "and" (117 occurrences) and "but" (51) are among the most frequent connectives. The position of conjunctions in their clause is limited; for instance, "because" cannot occur mid-clause. No adverbial had a non-adjacent Arg1.
• Link to the dependency treebank: The sentence-level dependency labels include rh (cause), k7t (time) and vmod (verbal modifier). Subordinators are not represented as words in these trees, and paired conjunctions are not linked as a single connective, so the discourse layer would add this missing information.