Semantic Application Based on the Bhagavad Gita: A Deep Learning Approach

Recent evolutions in Natural Language Processing involve techniques to convert text into meaningful high-dimensional numerical vector representations that can be used for various applications such as sentiment analysis. However, these vector representations can be computationally heavy and challenging to interpret. This paper proposes using autoencoders with cosine similarity as their loss function to generate encoded vectors that have intentionally discarded some semantic information. After the test, several autoencoder models like this were created to generate advice and question-answering trained on the entirety of the Hindu holy book of Bhagavad Gita by comparing it with the same algorithm working on the original vectors and concluding subjectively that the quality of the output has improved, while significantly reducing the time taken to generate an output. The findings suggest that intentional data loss through autoencoders is a promising technique for dimensionality reduction on word embeddings. The neural network was trained on the entire dataset using cosine similarity as the loss function. The work could explore the potential of this approach in other areas of natural language processing and develop more objective measures for evaluating its effectiveness.