We worked on a benchmark competition organized at SemEval-2024 focusing on detecting grammatically sound but semantically incorrect outputs generated by NLP models. Our team explored various techniques, including LLM-based approaches with prompt refinements, fine-tuning over NLI-based data, and ensemble models. Our ensemble based pipeline leveraging multiple signals and insights from LLM-based prompting and supervised learning, achieved competitive results, placing us in the top 10 among 50 teams across two different setups. 

Working on some cool areas of adopting LLM based modelling for industrial use-cases. Diving into core issues of model fine tuning on enterprise data, performing evaluation & comparative analysis of different LLMs and generic transformers based open source models. We have been working with LLama-2, OpenChat, Mistral and other open sourced LLMs to benchmark and compare with GPT-X based models to develop effective and scalable systems to pioneer for industry use-cases and solutions.
Staye in touch to explore & learn on this emerging thread.

We conducted a comparative analysis of transformer-based models for modeling tabular data, specifically on an industry-scale dataset. Previous studies showed promise on smaller datasets but fell short on larger industry-scale datasets due to challenges such as handling high-dimensional data and efficient preprocessing of categorical and numerical features. The study provided crucial insights into optimal data preprocessing, compared pre-training and direct supervised learning methods, discussed strategies for managing features, and highlighted trade-offs between computational resources and performance. Focusing on temporal financial data modeling, the research aimed to facilitate the systematic development and deployment of transformer-based models in real-world scenarios, emphasizing scalability.

This project focuses on understanding natural language in the context of financial domain and in general. Some underlying aspects we have been investigating are: i) detecting temporal expressions and performing its normalization for the unstructured data, ii) paraphrase understanding and generation for a business unit specific vocabulary and jargon, and iii) structured sentiment analysis, focusing on fine grained analysis of input text and extraction of important cues and sentiment expressions. The main goal is to address the nuances introduced for the financial data modeling and processing, which generic available models fails to capture.

An overview can be checked at following page: SOURCE

This project focuses on building end-to-end automated system for querying databases. Goal is to support performing reporting and business analytics while querying databases as performing general free order text search commonly used while interacting with search engines e.g. Google. We are building next generation reporting and analytics tools to support analysts for business reporting and also provide novel and convenient ways to interact with such NLIDB systems.

An overview can be checked at following page: SOURCE

With the rise in Deep Learning based models and pre-trained large embeddings for NLP modeling, there is a great need to explain how the NLP models make predictions and make these explanations human interpretable. This project focuses on making NLP models predictions more interpretable and explainable for the unstructured data, for a given task, e.g. Sentiment Analysis prediction.

I have worked extensively on search and retrieval models and their evaluation during my PhD and PostDoctoral period. Below is a list of search and related topics, I have worked in past. 

• Adhoc Search

• Enterprise Search

• Semantic Search

• Cross Lingual Search

• Multi-Modal Cross Lingual Search

• Conversational Search

I have been fortunate to also work with some cool collaborators, on topics of Sentiment Analysis, Semantic Similarity and other NLP related problems in past.

**More details can be checked in the corresponding papers.