Problem Statement:

Accessing and retrieving information from scanned documents, such as PDFs, can be challenging due to the lack of searchable text data. Traditional search methods rely on manual indexing or keyword-based searches, which are often inefficient and error-prone. This presents a significant barrier to organizations seeking to extract insights and knowledge from their document repositories.

Input:

The input to our OCR-based NLU semantic search engine consists of scanned documents in PDF format. These documents may contain valuable information, but the lack of searchable text data makes it difficult to extract insights and retrieve relevant information using traditional search methods.

Output:

The output of our solution is a set of scanned documents retrieved from the AWS Elasticsearch database, based on semantic similarity to the user query. By leveraging natural language understanding (NLU) techniques and deep learning models, we enable users to access and retrieve relevant information from scanned documents with ease.

Challenges Faced:

One of the main challenges in this project was accurately extracting text data from scanned documents using OCR (Optical Character Recognition) technology. Additionally, processing and analyzing large volumes of text data to generate embedding vectors while maintaining real-time performance posed a significant challenge. Furthermore, ensuring accurate semantic similarity comparison and efficient retrieval of documents from the Elasticsearch database presented additional complexities.

Proposed Solution:

Our solution leverages AWS Textract, a machine learning service for OCR, to extract text data from scanned documents. The extracted text data undergoes a series of text pre-processing stages using NLP techniques to enhance quality and relevance. A transformer-based deep learning model is then used to generate embedding vectors from the text data, which are stored in an AWS Elasticsearch database with proper indexing. A Streamlit-based UI provides an intuitive interface for users to input text queries, which are converted into embedding vectors and compared with vectors in the Elasticsearch database to retrieve semantically similar scanned documents.

Summary:

The completion of our OCR-based NLU semantic search engine represents a significant advancement in document retrieval and knowledge discovery. By combining OCR technology with deep learning models and NLP techniques, we enable organizations to unlock insights and extract valuable information from scanned documents with ease. Our solution not only improves efficiency and productivity but also promotes collaboration and innovation by providing users with seamless access to relevant information. With its intuitive interface and powerful search capabilities, our semantic search engine empowers organizations to harness the full potential of their document repositories and drive informed decision-making.