Automated Information Extraction using Ai Based Loan Processing Automation

Matt Payne
August 8, 2022

With the world becoming more digital, companies are looking for ways to cut down on bottlenecks and speed up some of their slowest processes. For banks and commercial loan lawyers one of the slowest processes is the extraction of key information and information exchange during the loan process. Any loan reviewer or processor will tell you it takes a ton of time and communication. Different things like running credit checks or signing the paperwork can cause a simple loan to take forever, and understanding key information is required at all stages of the process. For example, ICE Mortgage Technology (a leading cloud-based provider for lenders) reported that, on average, it took about 50 days to close a loan. (Source).

As automation continues to grow in industries requiring large amounts of documents and document review the timeline from start to finish is getting shorter and shorter. Loan review and processing automation has evolved tremendously in the past 2 years and is at the forefront of the ROI document processing automation. 

What is Automated Loan Processing?

Automated loan processing is a streamlining technique to decrease the amount of paperwork, communication, time, and data re-entry during the loan process by utilizing software, machine learning (ML), and other effective cloud-based solutions. Combining all the different steps of the loan process into a custom processing pipeline allows for a more systematic approach, leading to faster decisions for clients looking to secure a loan.

One of the biggest bottlenecks during the loan process is information extraction. This bottleneck has two sides to it, the first being communication. We are all humans and do what we can to serve our clients and pass information effectively. But sometimes it can be a huge time commitment that requires rereviewing documents or moving through notes. Utilizing cloud-based software that is always available, information can be extracted without clients and lenders having to get on the phone. Not only does this save the lenders time and energy, but it also ensures that the information gathered is accurate and well documented.

The second side is clearing up miscommunications. Since both parties constantly exchange calls and emails, data is easily lost or overwritten with new and possibly inaccurate data. Entirely relying on an automated loan process will cut down on the human erect, the time spent re-entering data, and data disputes when things are inaccurate. While an automated loan process is something all lenders are heading towards, pulling it off is still very difficult. 

What Are The Steps To Secure a Loan?

The loan process can be incredibly long and requires a lot of documents to be filled in or reviewed by the parties involved. The loan process can be broken down into three key parts, application, underwriting, and loan closing. 

Loan Application

During the first phase, you’ll work closely with a loan officer to gather the information needed to prequalify your loan. 

There will be small talk, predominantly questions about getting to know each other, but most of the time will be spent answering questions like the following.

  • Are you looking to refinance a pre-existing debt, and if so, why aren’t you able to pay the balance?
  • Is this a personal or business loan?
  • Tell us how you plan to utilize this loan, what ROI you anticipate from this capital, and what amount you’ll be able to pay back?

Once all of these questions are answered, and your documents are finally turned in, the bank will review and send you a proposal. After reviewing this proposal, you can move on to the underwriting process.

Underwriting Process

what is the job of the mortgage underwriter?

Next, in the underwriting phase, you will finally start working on your assigned loan. Realize up until this point, only documentation and information have been shared. This documentation still has not been verified till this step, when the underwriter will do a deep dive into the different financials of your company.

Let’s hope that in this long and extensive manual process.

  • The bank hasn’t made any errors inputting information
  • The applicant hasn’t made any errors inputting information
  • There hasn’t been any miscommunication between both parties

Once everything is verified, the underwriter will work up a credit memo. This credit memo will be presented to the credit officer of the financial institution, who will either approve or reject the loan. 

Loan Closing

Finally, the credit officer has accepted your loan, and the initial closing has been kicked off. You’ll be signing the final paperwork and reviewing the last documents during this final stage. These documents include security agreements, Deeds of Trust, and the Note.  Again, mountains of data are parsed through, checked, and verified for correctness.

What if We Could Eliminate Most of This Process?

Through modern ai based models we can parse the essential information out of documents to allow both parties a more effortless flow of data and information. While we know that all documents must be submitted, what if there was a way to find and extract the key information and parse it into a format that makes it easier to pass or reference in the different loan application steps?

For example, let’s say we’re reviewing a loan applicant's balance sheet. If we saw a line that the business currently has an outstanding loan that is overdue, we would probably instantly decline this candidate. However what if this was one of the last documents we reviewed for this applicant at the bank. The financial instituions would have invested hours into this application, while this candidate never had any chance of receiving a loan (once the line item on the balance sheet was found).

example balance sheet

Now, many banks deal with these wasted hours thousands of times per year.

With recent advances in machine learning and artificial intelligence, we can leverage intelligent systems that can take in all of this information at once, and leverage it to make quick, impactful decisions that can speed up the loan process for your customers.

How does it work?

What now makes Automated Loan Processing Possible?

Natural Language processing is one of the fastest-growing subsets of machine learning. With the release of models like GPT-3 in June 2020 and BERT in November 2018, we’ve entered a timeline where models can understand text semantics. This earlier work wasn’t always easy to use or quick to deploy.

Why Can We Now Better Understand Semantic Of Text

Word2Vec single word similarities

In 2013, a natural language processing technique was made popular called Word2Vec. The high level idea is that given some text, we can map each word to a vector. This may not seem novel, but with a word represented as a vector you can now bring in other techniques from mathematics. Some of these techniques include distance and direction that allow us to better understand word meaning and relationships. When we can represent words as vectors, figuring out the semantic meaning of our text becomes a reality when we can define words as vectors. When we take these vectors and map them into some n-dimensional space, things like euclidean distance and cosine similarity give us a decent idea of how closely related these vectors (words) are.

Pitfalls of Word2Vec

While (at the time) borderline revolutionary, there are pitfalls in the Word2Vec process that correlate to the strengths of NLP used in modern document processing. 

Single Word Embedding

word2vec example vectors

Word2Vec will make one vector representation for each word, and while there may be many different usages, these will all be classified into one vector. To understand this a little more deeply, read the excerpt below.

“We went to the fair to play some games, and Billy didn’t play fair.

In our example, we’d get a weak representative vector for the word fair as it would be the blend of the location and the way to play games.

Word Ordering

When trained Word2Vec will create word embeddings for you to utilize on your text. However Word2Vec (like the name suggests) does not utilize the complete sentence only the exact word to get the embedding.

While this may not seem like a problem, if we trained our model on a dataset that was using the word “fair” to represent the location, and while testing our dataset, the word fair came up about cheating - our model would struggle to understand the semantic meaning of our words.

Out of Vocabulary

The out-of-vocabulary problem is the last and probably the worst pitfall of Word2Vec. 

When a model is trained, our model will only know those words that existed in the training set. If while trying to find the semantic meaning of some text we run into a word that didn’t exist in our training - we will not have a vector for it. When this vector is missing, model accuracy falls off a cliff. This is pretty easy to understand; think of a sentence and remove one of the words from it. Whatever that word could change the whole meaning of the sentence.


BERT nlp architecture
BERT Architecture

In 2018, a different way of looking at text became popular with a new architecture and framework for thinking. Instead of focusing on the words what if we focused on the sentence? BERT (Bidirectional Encoder Representations from Transformers) is a state-of-the-art natural language processing algorithm. The general idea of BERT leverages many of the same ideas of Word2Vec, but instead of focusing only on that specific word, what if we focus on that word and the words around it?

BERT also allows for text to be read in any direction. This is a huge plus as before sequential reading was only allowed. This may not seem a big deal but the example below explains why it’s so impactful. During training, your model sees the sentence. 

“It was nice for Billy.”

We know from Word2Vec that our model will create a word embedding from all the sequential occurrences of the word Billy up to this point. Since this is the first time seeing this word, Word2Vec figures out that Bily is probably a noun and creates the embedding. However, since BERT can read the whole sentence.

“It was nice for Billy to retrieve the stick.

BERT can deduce that not only is Billy a noun, but it’s also actually an animal.

And if there was more to the story.

“It was nice for Billy to retrieve the stick; he’s such an awesome dog.”

BERT would be able to deduce that Billy is a male dog. While seeming simple now it’s easy to see how impactful algorithms like BERT allow us to derive deep meaning from the text we are provided.

These advances in natural language processing have opened the door for huge movements in automated document processing and information extraction. We’ve leveraged these learned NLP relationships with text extraction computer vision models to build state of the art loan automation pipelines.

How our machine learning solution works

width.ai document processing pipeline
Modules overview

With recent advances from our machine learning engineers here at Width.ai, we’ve developed a complete system that leverages state-of-the-art information extraction knowledge and text understanding models to automate various parts of the loan process.

Realize that every single document will be pushed through this state-of-the-art machine learning system, creating accurate and repeatable results for you and your clients.

Document Input

example W-2 document
An example W-2 we will extract information from

Loan documents such as tax returns, bank statements, W2s, or AR & AP can be uploaded in formats such as PDF or images. Our document processing pipelines are designed to support a wide range of variance in document formats or background noise for a given document type. This means you don’t have to use specific document templates or write specific rules to support variations in client documents. 

We’ve been able to use this same document processing pipeline for over 50 formats of legal document cover sheets (read)

Text Extraction From Loan Processing Documents 

Text fields recognized and drawn around in the W-2
Text fields recognized and drawn around in the W-2

We’ve built a custom OCR model that extracts machine & handwritten text from these specific loan documents with state of the art accuracy and the ability to understand variance in text parameters. On top of the raw text extraction is a module used to clean up and refine the text for our domain specific use case. This helps to greatly reduce errors in the NLP models downstream in common loan documents. 

Information Extraction & Understanding From Loan Documents

information extraction from w-2 document in the form of pairs
Extracted key value pairs from the W2 

A key part of information extraction being able to reduce manual effort is being able to correlate extracted information with fields you need. If we’re going to extract dollar amounts, names, and other entities we want to know what they correlate to in the loan process. 

We’ve built a domain specific entity recognition and matching model with the ability to not only extract key text from the documents but create correlations such as:

1. Correlate text to other text in the documents. Correlate dollar amounts to people or actions, lawyers to people, or othe information to loan seekers.

2. Correlate text from the document to form fields. Take your key information from the documents and autofill forms with the correct information. This allows you to truly automate the process of going through the loan application process.

3. Correlate text from documents to database fields. Automatically structure extracted text from the document into your database for other processes.

The model uses a mix of fine-tuned models that are domain specific to loan automation with our baseline document processing architecture underneath. We use a mix of state of the art natural language processing models to achieve the process and combination of text extraction and language understanding and fine-tune these ideas with loan documents. 

What You’ll See

Our loan automation pipelines can be deployed as a raw rest API or with a UI attached. The pathway we take mostly depends on how you plan on using the pipeline and how you plan on passing data. 

When using the loan automation pipeline as a piece of an larger automation pipeline it’s most common to deploy as a rest API and add a connection. We help you define a JSON input and structure an output that fits with the rest of your pipeline. We’ve also integrated our pipelines with tools such as ZenDesk and PagerDuty that allows you to automatically send notifications to humans. 

If you’re looking to have a human pass documents into the pipeline or don’t have upstream automation than a UI to upload documents works well. A simple one touch upload and run allows you to kick off the pipeline when you want to.

What do our accuracies look like?

Our baseline document processing pipeline has reached 94% accuracy on a standard document processing and information extraction dataset with over 11,000 examples. The dataset has over 20 classes labels and is a standard in the deep learning world. Not only does this dataset evaluate accuracy for text extraction tasks but post-OCR tasks such as named entity recognition. 

Fine-tuning our loan automation pipeline on your specific use case boosts the accuracy even further. Fine-tuning allows our models to understand your specific task and language cues that affect the target variables. Once the models have seen real examples of the loan documents you want to process and relationships between key fields and text that relationship is learned. 

Why your business needs an automated loan processing solution

Extracting text from legal documents
Extracting text from legal documents

By utilizing loan automation pipelines you will spend less time reviewing loan documents, finding key information, and completing the loan process. 

Repeatable Solution 

The repeatable nature of automation software allows you to standardize steps in the loan process. The fields you receive when processing a document are the same every time and don’t change unless you want them to. 

What Value it Can Bring

The benefits of loan automated loan processing are seen quickly when considering the costs of manual loan processing. Cutting down the time required for manual resources in the legal industry is always a great way to lower costs. With the current hourly rate in this industry the ROI you can get by automating any amount of manual processing is going to be large. 

On top of the reduction of required manual labor by expensive professionals software also allows you to reduce the time required. Many document processing tasks take as little as 2.7 seconds to complete and allow you to move through loan review processes quicker.

Let Width.ai Help You

Want to learn more about how we can deploy a custom loan automation pipeline with over 94% accuracy for your business? Let’s talk about how we can automate your processes -> Width.ai

width.ai logo