ANA: The Automated Nursing Agent

An AI companion for counteracting elder loneliness

Overview

Researchers at Amii, led by Osmar Zaïane (Fellow, Amii; Professor, University of Alberta), are developing ANA (Automated Nursing Agent), a conversational software agent (also known as a chatbot) designed to converse with elderly individuals who are living at home. Companionship is important for the mental and physical wellbeing of an elderly individual. Seeking companionship can be challenging for seniors and often not even possible. By building an intelligent automated system, our researchers seek to develop an emotionally-sensitive tool that provides information and verbal assistance to seniors who wouldn’t otherwise have companionship. ANA relies on techniques for natural language processing, emotion mining, knowledge graphs and information extraction in order to help with social needs and assist with simple home healthcare needs.

About the Work

ANA, which is currently in limited-prototype form, brings together numerous research lines into a single application. Working as both a personal assistant and a digital companion, ANA will build a knowledge base of personalized facts and memories (such as important people, places, activities and prescriptions), carry on engaging conversations that express and respond to emotions, and also answer impersonal questions from sources on the Internet. This will give ANA the ability to not only fulfill social needs but also assist with simple home healthcare needs such as prescription reminders.

ANA extracts information from text obtained from a speech-to-text converter; from this, it builds a personalized knowledge base that allows it to answer personal questions. ANA can also answer impersonal questions from sources on the Internet. When ANA receives an inquiry, it first searches the personalized knowledge base for individualized responses; if the answer cannot be located, it turns next to the Google Knowledge Graph as a general knowledge base. The system can even use templates and the personalized knowledge base to initiate conversations or to continue conversations initiated by the individual.

The system also leverages emotion mining techniques to express and respond to emotion during a conversation. In testing, the team’s model has been able to reliably express responses that matched requested emotions in most cases—though researchers note that some emotions, like surprise and love, are easier to express than others.

The main scientific objective of ANA is to build an automatic response generation agent, using classifiers that learn to generate adequate responses - i.e. sequences of words, in the context of a dialogue between a human and an agent. The classifiers used include Recurrent Neural Networks and other techniques that allow the injection of emotions, guide the topics of discussion, and more. In addition, researchers seek to build other classifiers to detect emotions, signs of dementia, and frailty from transcribed utterances. A further scientific objective is the extraction of personalized knowledge (entities and relationships) from the text and finding appropriate ways to represent this knowledge so that it can be used by the neural network generating responses including this knowledge. In the long term, the research team is seeking to build a system with logic reasoning that works in tandem with a neural network. The reasoning would be done on the collected knowledge base and provide the appropriate information to the neural network for the generation of responses.

The tool is designed to be deployed on common consumer tablets, and leverages the features of these devices for a variety of tasks. For example, it contains a database of jokes and uses reinforcement learning to choose which jokes to tell based on how well previous jokes were received. Other activities provided by ANA include reading, playing games, and providing means to tell their life story. By encouraging the elderly to share life experiences, it improves the connection between the agent and the user. Researchers are continuing to work on improving their limited-prototype so they can begin testing how the system works in real-life settings.

Additional Links:

• Automated Nursing Agent: A Software Agent for At-Home Elderly Care
• Augmenting Neural Response Generation with Context-Aware Topical Attention
• ANA at SemEval-2019 Task 3: Contextual Emotion detection in Conversations through hierarchical LSTMs and BERT
• ANA at SemEval-2020 Task 4: mUlti-task learNIng for cOmmonsense reasoNing (UNION)
• "Avoiding elder loneliness using artificial intelligence" from the University of Alberta's Faculty of Science