Patterns and problems : analyzing document layouts and improving mathematical reasoning of llms

Abstract

Documents play a pivotal role in conveying information, serving as integral carriers of knowledge across various domains. Their importance lies in their ability to encapsulate ideas, facts, and insights, thereby facilitating communication and record-keeping. Document analysis, as a field, involves the systematic examination of documents to extract meaningful insights, patterns, or information. While current document analysis tools have made strides, they face challenges such as limited accuracy, scalability issues, and struggles in handling diverse document formats. This highlights the pressing need for more robust and advanced document analysis tools. This research work attempts to address some of the broad domain difficulties associated with documents and their analysis. This research explains the solutions in the domain of domain adaptation based document layout detection, Detecting and Recognizing Tables within document images. Secondly, the Large Language Models (LLMs) proved to achieve state-of-the-art results on extensive, and complex NLP related tasks. But sometimes LLMs fails to solve basic mathematical reasoning tasks. Focusing on this, I’ve worked on proposing a extensive mathematical dataset for training LLMs to enhance their mathematical reasoning capabilities and proposed a efficient approach for solving physics problems using Reinforcement Learning with Human & AI feedback.