August 20, 2018

Back to school special: AppZen AI 101

Ryan Floersch

The AppZen artificial intelligence platform uses a combination of deep learning, machine learning, and semantic modeling to comprise our artificial intelligence platform.

Deep Learning & AppZen

Deep learning is what allows machines to solve complex problems based on different sets of training data. In theory, a machine with a deep learning system can watch a TV cooking show and learn how to cook or watch a mechanic and learn how to fix a car. When it comes to AppZen, we use deep learning for:

1.   Understanding complex receipts and documents

AppZen applies deep learning models to different types of documents (like hotel receipts or meal itineraries) to extract different components from each unique type. When it comes to receipts these are things like understanding line items, dates, merchant names, and dollar amounts while also understanding where these things typically appear on each type of receipt.

2.   Understanding natural language

AppZen uses deep learning models in 40+ languages to understand the context of the text within a receipt, which is often just a few words on each line item. For example, the system has learned that a “baggage” has strong associations to flying and that “pint” has strong association to alcohol. In more complex cases our machines also understand the nuances between words that might have different connotation if used independently or together. For example, the word “ginger” by itself will be associated to food, “ale” will be associated to alcohol, but “ginger ale” will have no association to either.

3.   Learning from user feedback

User feedback is used to update models so that the system improves over time. For example, over time, the AppZen AI has learned when the word spa might indicate that a business is actually a spa (Diane’s Day Spa Services) and when it is not a spa (Westin Hotel and Spa). In the latter case, user feedback has taught the machine that when the word “spa” appears as part of a hotel name that this merchant is likely not a spa.

Machine Learning & AppZen

Machine learning is typically used to answer binary questions based on training data. For example, if you have the temperature, pressure, precipitation, humidity, and any other weather related measurements for a given day over the past few years, you can build a machine learning model that uses that information to predict whether it will rain or not. At AppZen, machine learning is used for the following:

1.   Receipt classification

Classifies and understands whether a receipt is a hotel, meal -- or any other expense type. With such high volumes of different receipt types pushed into the AppZen platform over time, our machines are able to make quick decisions on what type of receipt they are looking at based on the variable inputs it has learned from in the past.

2.   Attribute extraction

Similar to classification, AppZen machines have learned over time from previous data sets how to extract certain values based on a receipts classification. All receipts, for example, have certain things in common like a merchant name, total amount, tax amount, purchase date, etc that AppZen looks for.

Semantic Modeling & AppZen

While machine learning guides AppZen’s AI to the first step of information gathering, AppZen uses a domain model of expenses that models the relationships between concepts and their attributes. For example, a hotel bill is a type of expense. Once the AppZen machines have recognized that a receipt is from a hotel, it will understand to look for things such as “number of nights”, “room rates,” “mini-bar,” “start date,” “end date,” etc -- since those types of terms are associated with a standard hotel bill. The AppZen AI has learned over time from these types of data sets and is able to predict what should be outlined on a standard hotel bill when they appear.

Over the next few weeks we will be exploring each of these fields more in more depth. Be sure to follow AppZen for the latest information on how we are on the forefront of AI technology.