KI
Projekt
4/07
Connecting People with Questions
to People with Answers
Mehmet H. Göker, Cynthia Thompson, Simo Arajärvi, Kevin Hua
The Connection Machine helps PricewaterhouseCoopers LLP (PwC) partners and staff to solve problems by connecting
people to people. It allows information seekers to enter their question in free text, finds knowledgeable colleagues,
forwards the question to them, obtains the answer and sends it back to the seeker. In the course of this interaction, the
application unobtrusively learns and updates user profiles and thereby increases its routing accuracy. The Connection
Machine combines aspects of expertise locators, adaptive case-based recommender systems and question answering
applications.
1 The Power of Connected People
Information, knowledge and experience are key success factors and the most important competitive advantage for any
business. However, most of this core corporate asset is in
the heads of the employees and cannot be easily accessed,
shared or distributed. Capturing and protecting it in documents (electronic or otherwise) is not only cumbersome, but
the documents become rapidly outdated and the maintenance effort required to keep document collections up-todate is formidable.
Furthermore, in the complex business scenarios of today’s
world, problem solving requires an increasingly large amount
of specialized knowledge. It is nearly impossible for one individual to be an expert in every aspect of a company’s business and deliver comprehensive solutions. Problem solving
requires co-operation and the sharing of ideas and information. The size of a corporation and the collective knowledge of
its employees are only valuable if these employees can share
their information and cooperate. We believe that the best way
to provide the most up-to-date and accurate information to
those who seek it is by putting them directly in touch with
the experts and implemented this idea in the PwC Connection Machine.
tion and helps PwC partners and staff to get answers to their
questions and to solve problems together. The Connection
Machine matches incoming questions to the expertise profiles of users, routes questions to the experts whose profiles
are of the highest similarity until one of them answers, collects the answer and relays it back to the seeker. Experts can
also refer the question to a more knowledgeable colleague
from their personal network. Users can interact with the Connection Machine by email or through its web interface.
By providing answers to questions rather than just locating people, the Connection Machine acts as a virtual, adaptive
expertise provider. The application merges technologies from
expertise locators (e.g. expertise profiles, routing of questions) with adaptive case-based recommender systems (e.g.
user modeling, selection of experts to route the question to)
and question answering applications. From the seeker‘s perspective, the application encapsulates all experts in the firm
and provides real answers to real questions.
3 Overview of the Connection
Machine
Figure 1 provides a general overview of the interaction
between the information Seeker, potential Providers and the
2 Directory Systems, Expertise
Locators and the Connection
Machine
Most firms allow their employees to search for other colleagues by means of directories or expertise locator systems.
Directory systems are essentially computerized versions of
phone-books and help users to find the contact information
of the person they are looking for using the name or departmental information. Expertise locators extend the data contained in the employee directories with work experience and
areas of specialization. They sometimes provide the ability to
search for experts using free text and typically permit users
to select and contact the suggested experts from within the
application [1, 2]. Both directory systems and expertise locators return contact information of potentially knowledgeable
people as a result of a query. It is left to the seeker to contact
the experts and obtain the answer.
The Connection Machine extends directory systems and
expertise locators beyond the pure search for contact informa-
Figure 1: Overview of the application workflow in the PwC Connection
Machine
Auszug aus: Künstliche Intelligenz, Heft 4/2007, ISSN 0933-1875, BöttcherIT Verlag, Bremen, www.kuenstliche-intelligenz.de/order
23
KI
Projekt
Connection Machine. An interaction with the Connection
Machine starts with an Information Seeker entering a question in free text format, as if he/she were asking a colleague
a question via email. The Seeker is also able to specify the
urgency of the question, the name of a client the question
relates to as well as additional, optional, structured information (e.g. knowledge domain, line of service, industry) to be
used to locate appropriate potential Providers (Figure 2).
The Connection Machine processes the query, finds a set
of matching potential Providers, contacts them via email (Figure 3) and places a visual indicator in the Summary page of
the web interface. In addition to the question, the potential
Providers are informed of the Seeker’s contact information
(e.g. name, line of service) and of the timeframe in which the
question needs to be answered.
After receiving a question, the potential Providers may
choose to respond either via web interface or via email.
Potential Providers may
• offer an answer to the question;
• request additional information from the Seeker;
• refer the question to other potential Providers; or
• decline to answer.
Once one of the potential Providers offers an answer or
requests additional information, he/she becomes the Provider for the interaction. From this point on, the Connection
Machine facilitates communicates between the information
Figure 2: Web interface of the PwC Connection Machine
4/07
Seeker and the Provider. It notifies and removes other potential Providers from the problem solving conversation.
If a Provider chooses to answer the question, the Seeker
is notified via email and a visual indicator in the web application. Upon receiving an answer, a Seeker can choose to accept
it and close the request, ask a clarification question about the
answer, or reject the answer and request a second opinion.
If a (potential) Provider decides that someone else from
his/her personal network is better suited to answer the question, he/she may choose to refer the question. In this way the
(virtual) personal network of the seeker is expanded and the
Connection Machine can learn about users who may have
been missing from its initial set of profiles. If the question is
declined by all contacted potential Providers in the first set,
it is sent to a second set. If these do not respond either, the
question is sent to the Domain Manager of the domain for
further processing.
4 User Modeling and Retrieval of
Potential Providers
To execute the workflow described above, the Connection
Machine needs to be able to determine who is potentially
able to answer the question by matching a seeker‘s query
against information it stores about other users. These user
profiles can be initialized from documents the user authored,
a resume, prior engagement histories or similar. To achieve
consistently high accuracy over a long period of time, the user
profiles are continuously updated with appropriate sections
of the interaction with the Connection Machine (Figure 4).
Figure 4: Profile maintenance and usage in the Connection Machine
A user model in the Connection Machine contains three types
of user profiles, each of which captures one aspect of a user’s
preferences or capabilities:
• The interest profile denotes the topics a user is interested
in. It is updated from the questions a user asks and the
associated clarifications.
• The expertise profile represents the topics that the user is
knowledgeable in. It is updated with the questions a user
could answer, the answers he/she provided, and all associated clarification conversations.
• The referral profiles represent the topics in which the user
is able to refer questions. It is updated with referred questions, any clarification conversation associated with them,
and any referral comments.
Figure 3: Sample email from the PwC Connection Machine
24
We also make a distinction between positive and negative profiles. Negative profiles contain information that the user does
not want to be associated with.
Auszug aus: Künstliche Intelligenz, Heft 4/2007, ISSN 0933-1875, BöttcherIT Verlag, Bremen, www.kuenstliche-intelligenz.de/order
KI
4/07
Projekt
The current version of the Connection Machine does not
utilize the interest and referral profiles for routing purposes.
These profiles are collected for future applications such as
intelligent document routing and the analysis of social networks.
The representation that is used to store the content of the
user models has a significant impact on the capabilities, flexibility, maintainability and learning abilities of the system. For
details on our approach to user modeling please refer to [6].
Once the system starts being used, the interest, expertise
and referral profiles are updated directly from the interactions. The user is also able to manually update and manage
his/her profiles by adding relevant documents or keywords.
The profile changes caused by a user’s interactions with the
Connection Machine are visible in the profile section of the
application as well and can be removed by users if they should
choose the do so.
The technology we used to implement these functions
in the Connection Machine is similar to User Adaptive, CaseBased Recommender Systems [3, 4, 5]. It is worthwhile to note
that the items in the case-base of the Connection Machine
are continuously evolving user models where each model
contains multiple profiles. Additionally, rather than being the
final goal, the retrieval process is an intermediate step and
users, whose expertise profile matched the query, are utilized
in the workflow to route questions to. The resulting interaction between the Seeker and Provider is the desired outcome
for the application. This is obviously not the case for standard
recommender systems where the retrieved items are suggested to the user as potential solutions.
5 Current Status
The Connection Machine has been in pilot with approximately
5000 users from five user groups (domains) within PricewaterhouseCoopers (U.S. firm only) for one year (July 2006 to July
2007). After the successful completion of the pilot phase, the
application has been deployed on PwC‘s knowledge gateway
and is currently accessible to all 30,000 US employees of the
firm.
To initiate a pilot with a user group, we collected all available and relevant directory information, resumes, and documents that represent user‘s experience. Once the profiles had
been populated with this initial data, we sent an announcement to the members of each group and give them access to
the application. Each domain is also assigned a Domain Manager who ensures that the questions flow smoothly.
Between July 2006 and July 2007, the Connection Machine
received 838 questions. 244 of these were subsequently
withdrawn due to various reasons (e.g. test questions, queries intended for search engines, or questions which were
answered through alternative channels). Of the remaining
593 questions, 437 were answered (74%), and 105 were waiting for a response (18%). For 36 questions (6%), the potential
provider requested clarification of the question and 15 seekers (2%) had provided such clarification.
At first look, the number of questions and answers in the
Connection Machine may seem to be low compared to standard web applications or search engines. However, the interactions in the Connection Machine are of an entirely different
quality: they contain concise questions related to a business
problem and highly specific answers to these. Considering
the drastically different nature of the domain and the busi-
ness value of answers of this nature, the numbers cannot be
compared to utilization numbers of search engines or similar
application.
We also noted that, based on their experience with
standard search engines, users are accustomed to express
their problem in keywords and feel awkward to ask their real
question. For our pilot, we had to remind users that they can
phrase their question as if they were asking a colleague and
to provide all information a colleague might need to know
to provide a good answer. With the Connection Machine,
users can enter what they actually want to know rather than
abstracting their question artificially. We consider support for
this ability to be a very important and real gap in knowledge
management technology.
The application was able to suggest a potential provider
to send the question to 99% of the time. This high number is
due to the fact that we always have domain managers in the
loop to help in situations where the Connection Machine was
unable to locate a provider. 91.3% of the time one of the contacted potential providers started to interact with the seeker
to provide an answer and 88% of the answers provided were
deemed sufficient and no second opinion was requested.
The seekers have the option to specify the timeframe in
which they need to have their questions answered. To allow
time for a potential second set of providers and for the domain
manager to act, the actual time each provider has to answer
a question is one third of the time the seeker states. Table 1
shows the percentage of questions for which the interaction
was started within the expected timeframe as well as the percentage of questions answered on time. As one would expect,
the more time the experts have, the more likely it is that they
will answer.
Response Needed in
4 Hrs
1 day
3 days
1 week
Interaction started by
Provider
50.43%
63.04%
76.42%
91.71%
Answered on time
35.04%
40.58%
54.72%
71.50%
Table 1: Response Times in the Connection Machine
6 Future Work and Summary
As next steps, we are planning to utilize the user models of
the Connection Machine for tasks such as targeted content
distribution to interested parties, routing content to experts
for verification, personalization of portals, as well as the creation of communities of interest and expertise. By analyzing
interest, expertise and referral profiles for the entire organization, analyses of gaps in knowledge could be performed and
areas of concentrated expertise or interest highlighted.
In summary, the PricewaterhouseCoopers Connection
Machine allows information seekers to enter their question in free text, finds knowledgeable colleagues, forwards
the question to them, obtains the answer and sends it back
to the seeker. In the course of this interaction, the Connection Machine unobtrusively updates and refines the interest,
expertise and referral profiles of each user. Rather than just
locating people, it extends the concepts of directory systems
and expertise locators and acts as a virtual (adaptive) expertise provider and answers questions.
Auszug aus: Künstliche Intelligenz, Heft 4/2007, ISSN 0933-1875, BöttcherIT Verlag, Bremen, www.kuenstliche-intelligenz.de/order
25
KI
4/07
Projekt
References
1.
Ackerman, M., Pipek V., Wulf V. (Eds), “Sharing Expertise: Beyond
Knowledge Management”, MIT Press, Cambridge, Massachussetts
and London, England, 2003
2. American Productivity & Quality Center (APQC), “Expertise Locator
Systems: Finding the Answers”, Benchmarking Report, APQC Publications, Houston, Texas, 2003 (www.apqc.org)
3. Thompson C., Göker M., Langley P., „A Personalized System for
Conversational Recommendations“, Journal of Artificial Intelligence
Research (JAIR), Volume: 21 (2004), pp 393-428
4. Bridge, D., Göker, M., McGinthy, L., Smyth, B., “Case-based Recommender Systems”, The Knowledge Engineering Review, 2006, Cambridge University Press, to appear.
5. Burke, R., “Hybrid Recommender Systems: Survey and Experiments”.
User Modeling and User-Adapted Interaction. 12(4), pages 331-370.
6. Göker, M., Thompson, C., Hua, K. , Arajärvi, S, „The PwC Connection
Machine, An Adaptive Expertise Provider“, in T.R. Roth-Berghofer et
al. (Eds.): Advances in Case-Based Reasoning, Proceedings of the
8th European Conference on Case-Based Reasoning, ECCBR 2006,
Fethiye, Turkey, September 4-7, 2006. LNAI 4106, pp. 549 – 563,
Springer-Verlag Berlin Heidelberg 2006
Contact
Mehmet H. Göker, Cynthia Thompson,
Simo Arajärvi, Kevin Hua
PricewaterhouseCoopers LLP
Center for Advanced Research
Ten Almaden Blvd, Suite 1600
San Jose, CA 95113
Email: {mehmet.goker, cynthia.thompson, simo.arajarvi,
kevin.k.hua}@us.pwc.com
Simo Arajärvi is a senior Java/JEE developer with
12 years of experience and various certificates from
Sun Microsystems and Oracle Corporation. Originally from Helsinki, Finland, Simo studied economics prior to pursuing software. He graduated from
Florida Southern College with a BS in CIS, and holds
an MBA from the University of Florida. Recently,
Simo has concentrated on Service Oriented Architecture and other strategic initiatives within PricewaterhouseCoopers.
Mehmet H. Göker leads the Connection Machine
Project at the PricewaterhouseCoopers Center for
Advanced Research in San Jose, California. His main
research focus is on using Case-Based Reasoning
and User-Adaptive technologies to develop intelligent, human centric knowledge management
systems that further the state of the art in applied
information technology. Mehmet received his Ph.D.
in Mechanical Engineering from the Department of
Machine Elements and Engineering Design at the
Darmstadt Institute of Technology in Darmstadt,
Germany. His M.Sc. degree in Aerospace Engineering is from the University of Michigan, and his M.Sc.
degree in Computer Engineering from the Bogaziçi
University in Istanbul, Turkey.
Kefeng Hua joined the Center for Advanced
Research of PricewaterhouseCoopers as Knowledge Mgmt. Systems Developer/Researcher to
work on the Connection Machine in 2004. Kefeng
received his Ph.D. in Computer Science (Artificial
Intelligence) in 1994 from Swiss Federal Institute
of Technology at Lausanne, a MS in Computer Science (Artificial Intelligence) from Beijing Institute
of Technology and a BS in Computer Science from
Hohai University in China.
Cynthia Thompson joined PricewaterhouseCoopers Center for Advanced Research in 2004. She is
the project leader for Magic Lens, which expedites
the review and comparison of even paper-based
documents and previously worked on the Connection Machine Project as a Knowledge Representation researcher and developer. Cindi received her
Ph.D. and M.A. in Computer Sciences from the University of Texas at Austin, and her B.S. in Computer
Science from North Carolina State University.
26
Auszug aus: Künstliche Intelligenz, Heft 4/2007, ISSN 0933-1875, BöttcherIT Verlag, Bremen, www.kuenstliche-intelligenz.de/order