Blog
-
A glimpse into Open Information Extraction with Large Language Models
This poster was presented at the Annual LCT Meeting, University of Malta, June 2023 and is based on the work by Wang et al. Large Language Models(LLM’s) store linguistic and relational information. The knowledge gained by LLM’s can be exploited and extracted to perform partially unsupervised open information extraction. Extracting n-ary representations from text in an unsupervised manner helps in many tasks, such as building knowledge bases and knowledge graph construction.
-
On Cross Lingual Learning and Dravidian Languages
Despite their richness and diversity, low-resource languages have not received as much attention from NLP researchers as high-resource languages like English and Spanish. However, recent progress in transfer learning, unsupervised learning, and data augmentation techniques show promise for improving NLP systems for low-resource languages. Leveraging the latent symmetry learned by multilingual language models through joint training, this report explores how cross-lingual learning can benefit the understanding of Dravidian languages, specifically, Telugu, Malayalam and Tamil. The blogpost covers tasks related to question answering, transliteration, code-switching, and hate speech detection. This non-exhaustive survey aims to facilitate further research in these important and socially beneficial tasks.
-
On Dual Encoders
This blogpost attempts to provide a glimpse at the potential and different applications of dual encoders while keeping in mind their pitfalls too and is inspired by Luan et al. (2021) and their use of dual encoders as first stage retrievers.
-
On Pruning Large Language Model's
This blogpost describe the compressing of BERT, in the context of the Lottery Ticket Hypothesis. Through emperical evidence obtained by fine tuning on several tasks, it is found that 30-40% of the parameters in the BERT model can be discarded.
-
A look at Diglossia
The term Diglossia, derived from the Greek word, diglōssos(two tongues) was brought into popular usage in the English language by Charles Ferguson who wrote an influential paper of the same name (Ferguson, 1959).
-
About Interpretable Machine Learning
Interpretable machine learning, as specified by Carvalho, Diogo V et al (2019) in “Machine learning interpretability: A survey on methods and metrics” enables a user to verify, interpret and understand the reasoning of a system.
-
Barebones Search Engine Implementation
This blog documents the process of implementing a search engine that takes a query term as input and ranks relevant documents in the corpus.