A Graph-Based Framework for Structured Prediction Tasks in Sanskrit

We propose a framework using energy-based models for multiple structured prediction tasks in Sanskrit. Ours is an arc-factored model, similar to the graph-based parsing approaches, and we consider of word segmentation, morphological parsing, dependency syntactic linearization, prosodification, “prosody-level” task introduce this work. search-based framework, which expects graph as input, where relevant linguistic information encoded nodes, edges are then used indicate association between these nodes. Typically, state-of-the-art morphosyntactic morphologically rich languages still rely on hand-crafted features their performance. But here, automate learning feature function. The function so learned, along with search space construct, encode consider. This enables us substantially reduce training data requirements low 10%, compared neural models. Our experiments Czech Sanskrit show language-agnostic nature train highly competitive both languages. Moreover, our incorporate language-specific constraints prune filter candidates during inference. obtain significant improvements by incorporating into model. In all discuss Sanskrit, either achieve results or ours only data-driven solution those tasks.