Digital Me

Digital Me

Established: April 1, 2016

Overview

Digital Me: Toward Digitalizing Everybody in the World

startjpeg

Introduction: Artificial Intelligence (AI) applications such as chat bots, software assistants etc. are attracting increasing attention from both academic and industry. Most existing work aim to assist people in information acquisition and task completion scenarios using knowledge from public data sources such as Web, enterprise documents, social media etc. Differently, digital Me (DM) aims to digitalize knowledge of each person for building her personal “Avatar”, through which way the DM agent of each individual can participate in digital work and life activities on behalf of this person for improving her productivity. DM agent could be considered as extended and never lose memory of each person. It can handle repeat work of each person in digital format. It can participate in social communications on behalf of each person. The knowledge learning process of DM agent is in an AI + HI mode, which means it actively learns from people to improve its knowledge and capability with no ending. With a life-long learning from each person, the knowledge and personal opinion of each person will be digitalized and will never die in a digital world for various application scenarios.

 

Uniqueness: Compare with many existing AI applications, three key points make DM unique:

  1. User vs contributor. Most existing work consider people as users of their service. DM considers each person as player and contributor of the AI game since everyone can easily build their own agent for their work, life and social activities. This approach has its uniqueness to make the AI popular among all users.
  2. Certain answers vs personal answers. Most existing work provide information service and task completion service with certain answers. However, many questions such as medical problems have no obvious certain answer since different doctors may have different opinions on the same symptom. DM allow users to browse the information with personal opinion from different people.
  3. Documented knowledge vs undocumented knowledge. Most existing work mining knowledge from public data sources which are documented for extraction purpose. DM not only learns knowledge from documents but learns from people by active learning and from environment by reinforcement learning, which you cannot see in any existing documents if no agent to learn and digitalize the knowledge.

 

Scenario Examples: In this 1-pager document, we give three digital work example scenarios from three different perspectives.  Firstly, the DM agent of a user could be considered as her extended memory and she can ask her agent about the knowledge she has known but cannot remember. For instance, Mary can ask her agent, i.e. D-Mary, where is the slides she made for review of next Tuesday since she cannot remember the slides title. As the assistant of a department, there are so many similar requests and questions repeatedly appear, D-Mary can automatically answer frequently asked questions on behalf of her using her knowledge after authorization. In the DM agent society, D-Mary can contact agents of other users for opinion collection and summarization on some target topics such as lunch options with no additional communication cost. As a summary, DM can improve productivity of each person in digital work scenarios from three perspectives, which are:

  1. Assist common tasks for time saving
  2. Reduce efforts on repeat work
  3. Reduce cost on communication

The DM for digital life scenarios will be introduced in separate document.

 

Technology Briefing: Technically, the core technologies of DM are knowledge mining, active learning etc. The technical challenges could be summarized in the list below:

  • Knowledge Mining: translate your data into knowledge
    • Common sense knowledge learning
    • Knowledge extraction for document understanding
  • Actively learning: learning your personal knowledge
    • Active learning UI and active conversation model
    • Active knowledge learning
  • Reinforcement learning: learning knowledge together with people from environment
  • Natural language understanding for human computer interaction
    • NER and EL
    • Intent classification
  • Engineering
    • Pipeline, tools, cloud service
    • Multi-agent communication interface

Contacts


Zhongyuan Wang

Dawei Zhang

Jun Yan

Wei-Ying Ma

Group

  Data Mining and Enterprise Intelligence Group, MSRA