ExperTwin is a Knowledge Advantage Machine (KAM) that is able to collect data from your areas of interest and present it in-time, in-context and in place to the worker workspace. This research paper describes how workers can be benefited from having a personal net of crawlers (as Google does) collecting and organizing updated data relevant to their areas of interest and delivering these to their workspace.
Swan(sea) Song – personal research during my six years at Swansea ... and bey...
ExperTwin: An Alter Ego in Cyberspace for Knowledge Workers
1. ExperTwin: An Alter Ego in
Cyberspace for
Knowledge Workers
C. Toxtli, C. Flores-Saviaga, M. Maurier, A. Ribot, T. Bankole, A. Entrekin, M.
Cantley, S. Singh, S. Reddy, R. Reddy
2. Problem statement
Knowledge workers (i.e. news writers, researchers) are benefited from having
the right information (i.e. in context), in time (i.e. auto suggestions) and in
place (i.e. in their workspace).
Querying and filtering multi-domain knowledge bases (i.e. Google) are time
consuming tasks. The collected information is usually moved to the workspace
and the friction of switching contexts cause interruptions and add up to
reduced productivity and increased stress (Czerwinski 2000, Iqbal 2007, Mark
2008)
3. Example
Imagine that you are writing an article about the relation of the United States
government to the North Korea government.
Maybe you need to know:
● What are the last actions from North Korea (query focused in North Korea)
● What is the United States government expecting (query focused in U.S.)
● How previous agreements had evolved (query ordered by time)
Then you collect, organize and cite the found information.
4. Solution - ExperTwin
In order to empower knowledge workers to be able to get opportune in-context
information in their workspace, we present ExperTwin, a Knowledge Advantage
Machine (KAM) capable to manage personal semantic networks.
5. Goal
The purpose of this research is to envision how a knowledge worker workspace
can be enhanced by applying Knowledge Advantage Machine frameworks
such as Vijjana (Makineni 2015).
6. Terminology
Knowledge Advantage (KA): Just as Mechanical Advantage played a key role in
the industrial era, the concept of Knowledge Advantage could be applied to deal
with the information explosion problem, and it is defined as the ratio of time it
takes to accomplish a knowledge based task to amount of time it takes to
search for the relevant knowledge.
Knowledge Advantage Machine (KAM): Any machine (or an app) that increases
the KA may be thought of as a KAM.
Knowledge Unit (KU): referred in this paper as JANs. Knowledge Object that
contains all the metadata of each content.
8. Knowledge Discovery
ExperTwin indexes the knowledge
from web sources, local sources,
web feeds and email.
ExperTwin crawlers constantly
updates the Knowledge Base from
these sources.
10. Learning Agent - Natural Language Processing
Purpose: Keyword extraction will, with a degree of accuracy, tell what the
purpose of many articles are. From aiding in determining relevance to user
preferences.
Keyword
Extraction
1. Text to obtain
keywords from
2. Number of keywords
wanted
3. Title of text if
obtainable
Dictionary of
Keywords with
weights.
Perform NLP with NLTK
and RAKE_NLTK libraries
11. Learning Agent - Machine Learning
According to the user preference of a content over different contexts, the
classifier give an extra weight to each content.
Preprocessing
1. Run through the database
2. Generate keywords for
every JAN in database
3. Define user defined
keywords
4. Label article as class 1/class
2 based on the results of
step 3
5. Collect master document
Tensorflo
w
12. Learning Agent - Machine Learning
1. CPU based tensorflow®
2. Learn vocabulary and
term document matrix
with scikit learn
3. relU + sigmoid activation
functions wt 50% dropout
4. Train with 70% of data
5. 87% test accuracy
Training
https://goo.gl/aRXEbp
Tensorflo
w
13. Learning Agent - Machine Learning
1. Load saved neural
network architecture
2. Query the database for
unclassified JANs
3. Retrieve content &
transform to document
term matrix
4. Make predictions
5. Update database
Testing/Processing
https://goo.gl/9q5azK
1. CPU based tensorflow®
2. Learn vocabulary and
term document matrix
with scikit learn
3. relU + sigmoid activation
functions wt 50% dropout
4. Train with 70% of data
5. 87% test accuracy
TrainingPreprocessing
1. Run through the database
2. Generate keywords for
every JAN in database
3. Define user defined
keywords
4. Label article as class 1/class
2 based on the results of
step 3
5. Collect master document
14. Learning Agent - GraphDB
The semantic network is stored in a
graph database by linking the
keywords to the JANs and assigning
different weights.
● Each twin has a meta-knowledge base
● Stores its biases and reasoning for relating data
● Self-representing (see image)
● Allows us to rank articles by relevance in real time
● Searchable
18. Visualization - Work area
● Need login (through Google Sign-In with a gmail address)
● Many users can use the interface at the same time
● Users need to set up interest keywords (add/delete)
● Keywords associated with user listed
● Users can pick keywords in dropdown or search to start
browsing
19. Visualization - Work area
● Context choice: Research / Professional / Study / Social / Others
● Will help in the choice/ranking of the articles
● Drag and Drop: to add files or folder to the database
● Help:
○ To send articles (url) to database through an email inbox@aiwvu.ml
○ To download Chrome Extension to add articles to database
20. Visualization - Content suggestions
From user search, get ten best ranked articles
● Thumbnail (if any)
● Title of article
● Date of publication
● Article clickable for a preview
21. Visualization - Content suggestions
Each article listed can be open in preview:
● Title
● Date of publication
● Source
● Full content
● User rating
23. 2D & 3D
visualizations
A search -> list of articles
4 types of 3D representations
available:
● Table
● Sphere
● Helix
● Grid
24. Graph visualizations
Articles and their relationship available in Graph 3D representation
Populated by a user search
Each article = node
Link = keyword shared by nodes
26. ● This work only focuses in how a Knowledge Advantage Machine
frameworks can be applied to implement an enhanced workspace for
knowledge workers.
● Evaluations should be performed to determine how much this tool can help
information workers to improve their work by being assisted by ExperTwin.
Limitations
27. Conclusions
We propose ExperTwin a Knowledge Advantage Machine that enhances the
knowledge worker workspace by adding in-context information retrieval
capabilities and information analysis visualizations to improve knowledge based
tasks.