The game-changing "Attention Is All You Need" paper marked the beginning of the Transformer model movement, which introduced a new way to handle sequential data. This architecture is the foundation for many of today’s advancements in AI, particularly in natural language processing (NLP).
Unlike older models, which used Recurrent Neural Networks (RNNs) or Convolutional Neural Networks (CNNs), Transformer models rely exclusively on attention mechanisms. These mechanisms allow the model to focus on the most important parts of data, whether it’s translating text, classifying documents, or understanding complex language patterns.
These advances have brought significant improvements in areas like text classification, language modeling, and even image recognition, making them invaluable for applications like mortgage origination.
AI Evolution in Mortgage
Over the past year, generative AI and large language models (LLMs) have become increasingly accessible, leading to an explosion of new AI tools. However, not all AI models are suited for every task, and some require significant resources to deploy effectively.
For example, while Transformer-based models can achieve impressive results in understanding language, they also require vast amounts of data to function correctly. This can be a barrier for many credit unions that may lack the resources to collect and manage the data needed to train these models effectively. However, for credit unions with access to large datasets and a commitment to innovation, these technologies can solve complex challenges.
Transforming Document Processing with AI
One of the most significant challenges in mortgage origination is efficiently processing documents. Credit unions are increasingly turning to AI to automate the extraction and classification of mortgage documents, which traditionally require manual data entry and time-consuming review. Dark Matter Technologies has been a leader in this space with its Artificial Intelligence Virtual Assistant® (AIVA®) solution.
AIVA® Docs has been particularly effective in automating document indexing and classification.
In 2024 lenders used the AIVA suite of AI virtual assistants, which can be integrated into the Empower ecosystem or work with other LOS platforms, to extract 36 million data labels with 99% service level agreement accuracy. That’s an unprecedented volume and accuracy level, and we’re very proud of that achievement.
Improving Performance for Diverse Document Types
While highly structured documents like tax forms are easier for AI to handle, others – like pay stubs or employment verification – pose more of a challenge due to their variability. These documents require specialized training, often leveraging extensive data sets to ensure that the AI system can accurately process them across a wide range of scenarios. AIVA, however, has made notable progress in this area, with some clients achieving F1 accuracy rates above 95% for variable document types.
Harnessing the Power of Attention Modeling
The real power of AI lies in its ability to understand context. Attention modeling—one of the key advancements in Transformer models—helps AI systems understand the relationships between different pieces of information in a document, even when that information is spread out or inconsistently formatted. This is crucial for accurately processing mortgage documents, which often require understanding both the language and the spatial layout of text on a page.
By incorporating attention mechanisms, AIVA can better navigate complex documents, ensuring that the right data is extracted and classified with the highest degree of accuracy. This ability to blend language understanding with spatial awareness allows for intelligent document processing that is significantly faster and more reliable than human intervention.
Adapting to Complex Scenarios with AI Flexibility
At the heart of successful AI-driven document processing is flexibility. No single model can solve every problem on its own. Complex challenges, such as those found in mortgage document processing, require a combination of AI tools and models. AIVA is designed to adapt to different scenarios, using various models to identify the best solution for each document type.
This adaptability is what sets AIVA apart. It doesn’t just rely on one approach but leverages a range of technologies, including open-source solutions, proprietary models, and industry-specific tools. This "AI marketplace" approach ensures that AIVA always uses the best possible tool for each document, ensuring accuracy and efficiency in document processing.
Why Credit Unions Should Embrace AI
The future of mortgage document processing lies in AI-driven automation. For credit unions, embracing these technologies can streamline operations, reduce costs and improve member service. By leveraging AI tools like AIVA, credit unions can process mortgage documents at speeds and accuracy levels that far exceed what is possible with manual methods.
AI early adopters will be well-positioned to stay competitive. The ability to process documents quickly and accurately can significantly enhance the member experience and improve operational efficiency—key factors for success in the competitive mortgage lending market.
Brian Swanson is the Director of AI/ML at Dark Matter Technologies, and Boaz Reisman is an AI specialist.