Implement knowledge mining and information extraction solutions

The optimal strategy is to retrain the model using only the 150 clean digital catalog documents as the primary training set, then evaluate if adding the watermarked documents in smaller batches improves field-level accuracy metrics.

This approach is correct for several reasons:

1. Immediate Quality Improvement: Starting with clean, high-quality documents (the 150 digital catalogs) will establish a strong baseline model that can quickly achieve higher accuracy. The current 71% accuracy is likely being dragged down by the problematic documents with handwritten annotations and watermarks.

2. Time Efficiency: With only 10 days to deployment, this strategy allows for rapid iteration. Training on 150 documents is significantly faster than 450, enabling multiple training cycles and refinements within the tight timeline.

3. Incremental Enhancement: After establishing a solid baseline with clean documents, the team can systematically test whether adding watermarked documents (which are partially readable) improves or degrades performance. This data-driven approach prevents contaminating the model with low-quality training data.

4. Preserves Labeling Investment: The team's 40 hours of labeling work is not wasted - they can selectively use the labeled data based on quality assessment.

5. Production Viability: A model trained on consistent, clean data is more likely to reach the 88% accuracy threshold quickly and maintain stable performance in production.

Why the other approaches are less suitable:

Preprocessing all 450 documents through an OCR enhancement pipeline would consume significant time (likely several days for setup, processing, and validation), making the 10-day deadline very difficult to meet. Additionally, OCR enhancement has limitations with handwritten text and heavily watermarked regions, and may not sufficiently improve quality to justify the time investment.

Segmenting into three separate models and using compose functionality adds architectural complexity that requires additional development, testing, and maintenance effort. This approach would likely exceed the 10-day timeline when considering the need to train three models, implement routing logic, and thoroughly test the composite solution. It's also operationally more complex for the procurement team.

Applying data augmentation techniques like rotation and brightness adjustment is designed for scenarios with insufficient training data or to improve model robustness to variations. However, these techniques don't address the fundamental issue: poor source data quality from handwritten annotations and watermarks. Data augmentation cannot create readable text where it's obscured or illegible, and may actually introduce additional noise that further degrades accuracy. Given the tight timeline and existing accuracy issues, this approach would likely not yield the necessary improvements to reach the 88% threshold.

The correct answer is that five labeled documents of the same form structure are required to train a custom template model in Azure AI Document Intelligence.

Azure AI Document Intelligence's custom template model (formerly known as custom form) requires a minimum of five labeled training documents that share the same layout and structure. This is the official requirement documented by Microsoft for training template models.

Template models work by learning the fixed positions of fields in documents with consistent layouts. The five-document minimum ensures the model has enough examples to identify patterns in field locations and extract data accurately from similar documents.

Why the other options are incorrect:

The option suggesting ten labeled documents sets an unnecessarily high minimum requirement. While more training data generally improves model performance, Azure AI Document Intelligence only requires five documents as a minimum threshold.

The option mentioning five documents with different layout variations describes the requirements for a neural model (custom neural), not a template model. Template models specifically require documents with the same structure, as they rely on consistent field positioning.

The option suggesting three labeled documents falls short of the minimum requirement. While three documents might seem sufficient, Microsoft has established five as the minimum number needed to adequately train a template model and achieve reliable extraction results.

It's important to note that while five is the minimum, providing more labeled examples (up to the maximum of 500 pages or files) typically results in better model accuracy and performance.

The correct answer is Azure AI Document Intelligence.

Azure AI Document Intelligence (formerly known as Form Recognizer) is specifically designed to extract structured data from domain-specific documents using labeled datasets. This service provides:

• Custom model training capabilities where you can train models on your own labeled documents
• Support for both template-based approaches (for structured forms with consistent layouts) and neural network approaches (for varying document layouts)
• The ability to extract key-value pairs, tables, and other structured information from documents
• Pre-built models for common document types and the flexibility to create custom models for domain-specific documents

Why the other options are not the best answers:

The second option involving Azure Machine Learning with AutoML is a general machine learning platform. While it can be used for classification and feature extraction, it is not specifically optimized for document intelligence tasks and would require significantly more custom development work to achieve document extraction capabilities. It lacks the pre-built document processing features that are essential for efficient document understanding.

The third option mentioning Azure Cognitive Search is primarily a search and indexing service. While it can enrich documents through custom skills, it is not designed as a training platform for custom document extraction models. Its primary purpose is information retrieval rather than model training for structured data extraction.

The fourth option combines Form Recognizer Studio with Azure OpenAI Service. While Form Recognizer Studio is indeed part of Azure AI Document Intelligence, the combination with OpenAI Service is unnecessary for the core capability described in the question. Azure AI Document Intelligence alone provides all the required functionality for training custom models with labeled datasets using both template-based and neural approaches. Adding Azure OpenAI would be relevant for other scenarios like document summarization or advanced language understanding, but not for the specific training capabilities mentioned.