Many people wonder how RPA differs from Conversational AI and NLP. The short answer is that Conversational AI and NLP use a combination of rules and probability, whereas RPA uses structured inputs and defined logic.
RPA uses structured inputs and defined logic
RPA (Robotic Process Automation) is a generic term to describe any type of process automation where a piece of software mimics the repetitive actions of a human user. It’s most commonly used to mimic the clicks of a human user to perform actions such as data entry on a user interface, complete forms, trigger actions on a screen, upload files to an ERP system and more. RPA is preprogrammed, meaning there is no inherent ability in these systems to learn from their actions or environment.
Expense processing is a great candidate for RPA. Though mostly handled by one department, each expense report is processed with multiple steps. It might start with retrieving an attachment from an email or information from an app, categorizing the expenses, routing the invoice to the right person for approval, then uploading to a finance system for payout. All of these aspects can be handled by robotic process automation software.
Conversational AI and NLP uses a combination of rules and probability
Artificial Intelligence is primarily software that has the ability to learn from the data it processes. Machine Learning is a key component of modern AI systems that give AI entities the ability to extract and learn patterns from data. Based on the type, the Machine Learning algorithm could learn from experience in controlled environments where rules are well defined (e.g. playing chess), learn from explicitly defined labels that are usually set by a human (e.g. NLP, fraud detection) or simply identify and delineate patterns (e.g. finding similarity in data).
NLP, or Natural Language Processing, is a sub-field of Machine Learning that deals with performing language related tasks such as translation, comprehension, summarization or understanding pieces of text. A wide field in itself, the most commonly deployed language models perform understanding tasks and are used primarily in chatbots.
NLU (Natural Language Understanding), takes a piece of text (typically a sentence) and maps it to an “intent”, which is a summary of what that sentence or phrase means. For example, “I need to pay my bill” may be summarized to “pay bill” intent.
An NLU model is explicitly trained by humans, who decide what intents should be learned by the model. This is known as “supervised learning”. Intents are trained by showing the model sample “utterances”, which are sample phrases for an intent. “I need to pay by bill”, “bill payment”, or “how do I pay my bill” may be examples of utterances to train the intent “pay bill”. The model then summarizes these examples into a machine representation that renders meaning to these sentences (known as “vector space”).
When a new phrase that is similar in meaning comes along (e.g. “How much is my invoice? I want to pay it”), the model finds the closest match in vector space and then assigns a probability of how well the input phrase matches the intent. For example, the above phrase may match “pay bill” intent with a probability of 0.92, or, 92% match. If this value is above a set threshold, the chatbot will trigger the action or response associated with the “pay bill” intent. If multiple matches are found, then the one with the highest probability is chosen. If the match is below the threshold, then the chatbot may decline to answer. All the responses and actions for each intent are preassigned by human conversation designers.
The combination of RPA and natural language processing (NLP) allows an enterprise to expand the variety and complexity of the transactional journeys they expose their customers to through conversational channels.