Biraj Silwal

Biraj Silwal

Contextual Embeddings have yielded state-of-the-art results in various natural language processing tasks. However, these embeddings are constrained by models requiring large amounts of data and huge computing power. This is an issue for low-resource languages like Nepali as the amount of data available over the internet is not always sufficient for the models. This work has taken an incomplete BERT model with six attention heads pretrained on Nepali language and finetuned it on previously unseen data. The obtained results from intrinsic and extrinsic evaluations have been compared to the results drawn from the original model baseline and a complete BERT model pretrained on Nepali language as the oracle. The results demonstrate that even though the oracle is better on average, finetuning the small embeddings drastically improves results compared to the original baseline.

Biraj Silwal

The distributed representations currently used are dense and uninterpretable, leading to interpretations that themselves are relative, overcomplete, and hard to interpret. We propose a method that transforms these word vectors into reduced syntactic representations. The resulting representations are compact and interpretable allowing better visualization and comparison of the word vectors and we successively demonstrate that the drawn interpretations are in line with human judgment. The syntactic representations are then used to create hierarchical word vectors using an incremental learning approach similar to the hierarchical aspect of human learning. As these representations are drawn from pre-trained vectors, the generation process and learning approach are computationally efficient. Most importantly, we find out that syntactic representations provide a plausible interpretation of the vectors and subsequent hierarchical vectors outperform the original vectors in benchmark tests.