Abstract:
"Semantic searches allow users to retrieve relevant information from raw data. Advancement in embeddings such as transformer models have made the semantic retrieval more efficient. However, the use of transformers requires more resources than a lexical search. And retrieval systems with lexical search tend to be inaccurate. Therefore, creating a retrieval pipeline for a raw document with a decent speed and accuracy is difficult.
Sentence transformers are the advancement in transformers that are more suited for
semantic search. Combination of several transformer architectures will further improve the
accuracy of retrieval. Integrating sentence encoders along with word-based architectures will Improve the speed of retrieval architectures.
This research presents ensembled semantic retrieval technique to search a raw PDF with
low latency. A novel combination of word embeddings with sentence encoders to retrieve
information is introduced by this research."
