Poetry to Prose Conversion in Sanskrit as a Linearisation Task: A Case for Low-Resource Languages

Abstract

The word ordering in a Sanskrit verse is often not aligned with its corresponding prose order. Conversion of the verse to its corresponding prose helps in better comprehension of the construction. Owing to the resource constraints, we formulate this task as a word ordering (linearisation) task. In doing so, we completely ignore the word arrangement at the verse side. kāvya guru, the approach we propose, essentially consists of a pipeline of two pretraining steps followed by a seq2seq model. The first pretraining step learns task specific token embeddings from pretrained embeddings. In the next step, we generate multiple hypotheses for possible word arrangements of the input (Wang et al., 2018). We then use them as inputs to a neural seq2seq model for the final prediction. We empirically show that the hypotheses generated by our pretraining step result in predictions that consistently outperform predictions based on the original order in the verse. Overall, kāvya guru outperforms current state of the art models in linearisation for the poetry to prose conversion task in Sanskrit.

A Graph-Based Framework for Structured Prediction Tasks in Sanskrit

“…For dependency parsing, we use UAS and LAS. Similarly for linearization tasks, we follow Krishna et al (2019) and report the performance of the systems using BLEU (Papineni et al 2002), Kendall's Tau (τ) score (Lapata 2003), and perfect match score, namely, the percentage of sentences with exact match to the input. For WS, MP, and the joint task of WS and MP, we use macro-averaged Precision, Recall, and F-Score (Krishna et al 2018).…”

“…a fromSrīnivāsa Aiyaṅkār (1910). 26 Krishna et al (2019), the current state of the art in syntactic linearization in Sanskrit, use a test set of 3,017 textlines from this data set. We use the same test data for all these three tasks.…”

“…These baselines are provided with additional data for their training. For syntactic linearization, we use 95,000 prose order sentences crawled from Wikipedia by Krishna et al (2019). Similarly for prosodification, we use 100,000 verses from DCS and Vedabase.…”

“…kāvya guru. This is a seq2seq model (Krishna et al 2019) with gated CNNs (Gehring et al 2017), using a sequence-level loss (Edunov et al 2018). The model uses a weighted combination of expected risk minimization and the token-level log likelihood with label smoothing as loss.…”

“…All our models use as low as 10% of the training data as required by the current neural SoTA models in Sanskrit for various tasks. We used around 9.26% (10,000) and 1.5% (8,200) of training data, as against 108,000 and 0.5 million training sentences for the SoTA neural models in syntactic linearization (Krishna et al 2019) and word segmentation (Hellwig and Nehrdich 2018), respectively.…”