Web servicesknowledgebasetechnology

The Insider’s Guide to Knowledge Base Technology
Developments at the Cutting Edge of eService

by Doug Warner, RightNow Technologies

 2002 RightNow Technologies, Inc.

Contents
Executive Summary ____________________________________________ 1
Philosophical Approach _________________________________________ 1
Current Artificial Intelligence (AI) _________________________________ 1
Auto-Adjusting Answers_________________________________________ 2
Auto-Generated Answer Relatedness ______________________________ 2
Natural Language Processing_____________________________________ 3
Suggested Solutions ___________________________________________ 4
Emotion Detection _____________________________________________ 5
Information Browsing __________________________________________ 5
Knowledge Maintenance ________________________________________ 7

Future Directions for AI in RightNow eService Center_________________ 7
Information Gap Analysis _______________________________________ 8
Visualization__________________________________________________ 8
Blue Sky_____________________________________________________ 9
About the Author _____________________________________________ 10
About RightNow Technologies ___________________________________ 10


Executive Summary

RightNow eService Center (RNeSC) is RightNow Technologies' flagship
Internet customer service product: robust, easy to use and easy to
administer. An article by PC Week stated, “Companies that do not want to
save money on customer support should not buy RightNow Web.” (Feb. 14,
2000, page 30). In February 2002, InfoWorld gave their highest ranking,
"Deploy Immediately" for RightNow eService Center, saying, “This is how
eService can save your company money: By building a comprehensive
knowledge base of the most common support issues, it can anticipate
customers' questions and answer them automatically, without costly service
and support reps getting involved.” (Feb. 15, 2002). In this paper we’ll
examine the various forms of artificial intelligence (AI) contained in RNeSC,
the implications AI has in this area, and how RightNow Technologies plans to
utilize AI in future products.

Philosophical Approach

Before considering the various artificial intelligence methods implemented
and proposed in RightNow Technologies’ products, we need to step back and
look at our approach from a philosophical perspective. Our goal is to create
the industry’s most usable and easily administered suite of tools for Internet
customer service, also known as eService, which requires the development
and adaptation of AI algorithms to the eService domain. To achieve this goal,
we’re striving to create the ideal, self-maintaining, organic system. An
organic Internet customer service system responds to the interests of the
customer, yet requires no administration to address those interests. This sort
of personalized automation can only be achieved using AI.

Because the organic approach means keeping the system running with little
or no administration, we consider any task requiring administration to be a
problem in need of a solution. Administration should only be necessary when
the underlying eService system is unable to adapt to, or meet, the end user's
needs. Our use of AI relies heavily on organic adaptive methods, while
avoiding those methods that require extensive, manual fine-tuning.

Current Artificial Intelligence

All versions of RNeSC have some level of embedded AI, and each new
release has improved and expanded the use of this groundbreaking
technology. Currently, AI appears in RNeSC in the form of automatically
ranking the most helpful Answers based on the interests of the customers.
Customers are also able to view Answers that are specifically related to those
they already read as well as view the larger scale interrelationships of the
Answers. Natural language processing exists in several areas of the product,
including generating automated responses to questions as well as detecting
and using the emotional content of incoming messages. Moreover, in keeping
with the organic approach, none of these features requires administrative
intervention, requiring embedded AI to ensure appropriate configuration
across all customer uses.
1

Auto-Adjusting Answers

The original approach to auto-adjusting Answers was basic but effective.
Customers were explicitly asked for feedback on all of the Answers they
viewed. Those questions with the highest level of feedback moved to the top
of the Answers list.

Subsequent additions to RNeSC include expanded AI applications in the
ordering of the Answers list. For those customers providing feedback on a
specific Answer, RNeSC now offers a wider range of positive responses (those
that move the Answers upward on the Answers list, as well as allowing
negative responses. Negative responses indicate a particular Answer was not
helpful, which means this solution will, over time, move down the list.

Implicit feedback is a patent pending technology available only in RNeSC.
RightNow uses implicit feedback to collect reactions from customers who do
not provide explicit responses to Answers. Analyzing the Answers each
customer viewed, and the order in which they viewed them, generates the
data used to determine Answer usefulness.

Customer questions change over time and RNeSC’s Answer orderings reflect
that. Every item containing any usage information has a time stamp
indicating that item’s most recent update. If the information becomes “stale”
(not accessed for a given time period), then its relative importance is
downgraded. Less popular information then drops down the list, while
information accessed more often rises to the top. We call this “information
half-life.''

What AI technique describes Answer ranking auto-adjustments?

Several standard AI techniques, including collaborative filtering and recommender
systems, come close to describing our approach, but still fail to grasp the essence
of what we are doing. However, one rather new AI approach seems to more closely
describe our auto-adjustment scheme. Under the domain of artificial life, our
approach to Answers auto-adjustment could be considered “swarm intelligence.”
The only distinction between most applications of swarm intelligence and RNeSC’s
implementation is that the customer's directed seeking behavior is what drives the
system, instead of relying on simple programmed agents.

SmartAssistant™ Answer Relatedness

Related information presented to Web users can take many forms. One
example is a shopping aid, such as Amazon.com, which suggests books
based upon a customer’s stated reading preferences. For eService Center,
related information takes a slightly different route. Here, you want to present
users with Answers related to those previously viewed, hoping if the current
Answer does not solve their question, then the related Answer should.
Thanks to RNeSC’s patent-pending, organic, adaptive AI techniques,
customers find solutions faster by viewing intelligently linked Answers
related to other solutions already viewed.
2

RNeSC assumes each customer session involves a new search for a single
answer. Underlying technology automatically performs click-stream analysis
to track, record and leverage the paths users take through the knowledge
base. Contrary to typical recommender systems, RNeSC's AI analysis is
conducted from an answer-focused approach, rather than customer-focused.
As a result, when building information relatedness, the question asked by a
customer is not as important as the way in which the customer reached that
answer.

How are Answers organically related in RNeSC?

Information relatedness is a two-part process. The first part relies on a type of
swarm or agent approach. In this case, it is assumed each customer session is a
directed search for a single item existing in the knowledge base. Each step taken
by a customer in the quest for this information builds on those steps already
taken. This series of steps generates the second part of the process—an
information map. The result of a large number (swarm) of customers (agents) is a
detailed map of how each item in the knowledge base relates to the others.

To initialize relationships on a new knowledge base as well as identify
relationships, which visitors fail to detect, RNeSC also periodically detects textual
similarity between existing Answers and automatically adds relationships between
highly similar documents.

Natural Language Processing

Natural Language Processing (NLP) has been applied to several areas of
RNeSC. NLP is a process enabling the computer to “read'' a piece of text and
gain information about its content. There are numerous NLP approaches, but
most are distinguished from a basic keyword searching routine by analyzing
the order and grouping of the words in the text. More complex methods also
attempt to identify part-of-speech (POS) or grammatical structures and
perform an ontological analysis of the words and sentences. The result is a
reasonable interpretation of questions asked in the customer's native
language.

NLP exists in RNeSC in numerous areas. It is most recognizable in the
searching interface where users enter complete sentences, sentence
fragments or simple binary keywords for searching. In addition, the "Similar
Phrase" option within RNeSC searching allows an intent-based search style
where users find Answers using different language than is contained in the
search terms.

NLP is also prominent in SmartAssistant "Suggested Solutions," which is a
feature available in multiple areas within RNeSC, but is commonly considered
as a method to generate automatic responses to incoming customer
communication. Similarly, emotion detection and information browsing make
use of NLP to an even greater extent than the basic searching interface.

3

What form of NLP is currently used in RNeSC?

The NLP used by RNeSC is based on a modified keyword-based statistical
method. In some cases, we also use part of speech tagging localized for
numerous languages. The rudiments of ontological analysis exist in limited
places in RNeSC. This statistics-based combination of approaches work
especially well for RNeSC because of the highly organized Answers knowledge
base. When searching, instead of attempting to find a general interpretation for
the question by a statistical analysis, the result set is narrowed down into
smaller and smaller areas within the list of Answers. More than simple word
frequency and keyword matching are used, and steps have been taken to
circumvent the pitfalls of standard statistical NLP methods. The requirements of
detecting emotion and generating information browsing structures are more
complex than most searching requirements, so these techniques incorporate not
only the statistical methods, but also POS tagging and other more aggressive
approaches.

SmartAssistant Suggested Solutions

Aside from searching, the most recognizable application of NLP in RNeSC is in
Suggested Solutions. With Suggested Solutions, RNeSC "reads" a complete
customer query and determines the conceptual intent of the question to
return an Answer from the system. Suggested Solutions is a tool available
for both automatically responding to customer queries as well as aiding
customer service representatives working to address a customer inquiry.

The Suggested Solutions feature works by taking a complete customer query
and breaking it down into component parts. These parts are then analyzed
for conceptual RNeSC. Depending on the use, the resulting Answers might
be returned to the customer via email, the standard RNeSC Web interface or
available to the customer service representative through their normal work
interface.

How does Suggested Solutions detect conceptual intent?

Suggested Solutions is an integration of our search approach and our
information browse approach. Suggested Solutions takes advantage of the
power of RNeSC's NLP searching by providing a much larger amount of
contextual language to allow a more robust match with the available Answers.
These various statistically based NLP techniques are then augmented by
evaluating how closely the conceptual aspects of the search results match with
the information browse classification of the customer question. This step of
focusing search-style results by using our automated techniques that find
conceptual structure in semi-unstructured text makes for a highly accurate
response to the customer’s question.

4

SmartSense™ Emotion Detection

While the broad topic of emotion has been studied in psychology for decades,
very little effort has been spent on attempting to detect emotion in text and
virtually no research exists on how to automate the process with a computer.
With RNeSC, we have created the first integrated system, called SmartSense,
to use emotion detection in text as a useful part of customer service
workflow.

In RNeSC, an emotion indicator is available on all correspondence sent into
and out of the system. Identifying emotion on incoming correspondence
provides two important functionalities. First, companies use this as a method
of triage. With this use, angry customers are routed to specially trained
employees; customer fan mail can be handled with an automated response,
while normal communication enters through normal channels. Second, each
customer service representative begins each interaction with an expectation
of the emotional content of the correspondence and marshal their resources
appropriately.

Identifying emotion on outgoing correspondence is also vitally important to a
customer service organization. Customer service managers use this tool as
an indicator of employee performance and behavior and employees can use
this tool as an incentive to help them achieve the highest quality response.

How is Emotion Detection performed in RNeSC?

Detection of emotion in text in RNeSC is a multi-step process. First, the text is
tagged with a language-independent POS tagger. This POS tagger learns
sentence structures for a language as a set of transition rules. These rules are
then applied to the text to label each word as a noun, verb, etc. Once words
are labeled, they are checked for emotional valence and assigned an emotional
rating. Based on the POS tagging, language specific modifiers such as ‘not’ and
‘very’ are applied to modify the preliminary valence score. The scores are then
summed across all sentences and finally run through a fuzzy-logic process to
determine an overall score for the correspondence. In RNeSC, valence is
determined from a proprietary list of emotionally charged words, abbreviations,
and emoticons. Administrators of the system are free to add new emotion
words, or change the values associated with existing words.

Information Browsing

People are diverse in their learning and interaction styles. Traditional
searching paradigms are biased towards one particular interaction style while
ignoring others. Specifically, there exists a key difference in human memory
for recall vs. recognition tasks. Recall tasks are the more challenging ‘fill in
the blank’ or ‘short answer’ sorts of test questions you may remember from
school. Recognition tasks are more similar to the easier ‘multiple choice’ test
questions. The difference between these two tasks is whether or not you
need to produce information without available examples or whether you
simply recognize relevant information when you see it. Traditional search
5

tasks require the user to generate the exact search term they desire, but
many users might not know the correct language to describe their problem.

Information browsing is a search interface available in RNeSC. This interface
automatically evaluates all public Answers to determine the inherent
structure in the information and provides an appropriate label for each
conceptual group within that structure. Thus, when an end-user does not
know the correct terms to describe their problem, they can drill down
through the automatically generated topics in this alternate search interface
on words they recognize, instead of words they are required to produce.

How are the topics automatically generated in the RNeSC Browse
interfaces?

Information browsing is the most complex technique in RNeSC. The process of
generating topics and summaries requires: usage-biased natural language
feature detection including POS tagging, advanced statistical clustering
techniques, learned classification rules, extractive summarization, and
numerous usability heuristics. Each of these techniques is a novel adaptation or
completely new algorithm and to our knowledge, these techniques have never
before been used together.

The feature detection stage combines the same techniques used for POS tagging
as we use for the emotion detection described earlier as well as using a similar
approach to our auto-adjusting answers to identify those terms most used when
users search the knowledge base.

The clustering stage uses a novel linear-time statistical algorithm to generate a
hierarchical organization of the knowledge. This technique is highly modified to
be adaptive so it works in general on any type of company data. This technique
uses fuzzy logic to iteratively adjust the clustering algorithm parameters to best
adapt to the current data.

Once the hierarchical structure is gleaned from the data, this structure is
learned so new Answers added to the system are correctly placed in the
structure. This step allows for rapid placement of new information as well as
provides a framework of describing the important concepts to include in the
descriptions of the hierarchical clusters. In addition, these learned rules allow
individual answers to appear in multiple locations in the structure as
appropriate.

Summaries are generated for each group in the hierarchy allowing users to
identify the contents of that group quickly. These summaries consist of the
conceptually most important words out of all Answers, as determined through
POS, statistical and heuristic measures.

6

Knowledge Maintenance

Dynamic information systems, such as the RNeSC knowledge base, require
periodic maintenance to maintain optimal performance. This traditionally
requires knowledge base administrators to periodically clean up the
relationships and rankings to remove any outdated information that has crept
into the system. However, this manual approach runs counter to the
philosophy of an organic knowledge base, where manual administration is
avoided whenever possible. Several AI maintenance functions are integral for
RNeSC functionality, including automatically determining settings in the
system for information aging as well as adapting system functionality based
on user load.

One example of how RNeSC adapts to its usage is within the information-
browsing framework. There is a tradeoff between consistency and
adaptability within the information browse hierarchical structure. A fully
adaptive system is able to change information groupings with each added or
deleted piece of information, while the overly consistent structure does not
recognize when completely novel information was added to the system.
RNeSC makes the best of both worlds by recognizing when changes to the
available Answers are similar to the current information structure or if they
introduce new concepts into the structure.

Many other areas within RNeSC are automatically adaptive to the use of the
system, allowing for increased ease-of-use without additional administrative
overhead. In most cases, a simple preference setting sets a site-wide bias
for the rate of adaptability for these algorithms, from conservative to
aggressive.

What techniques are used for knowledge maintenance in RNeSC?

Most knowledge maintenance techniques are some combination of heuristics,
statistics, and fuzzy logic. Each individual case is evaluated on its merits. The
simple cases use simple heuristics, whereas complex scenarios like the adaptive
analysis of the information browsing structure. This technique uses a complex
combination of heuristics, statistics, and fuzzy logic to function as people would
expect without having to consider its operation.

Future Directions for AI in RNeSC

The application of organic AI plays a continuing role in the increasing
capabilities of RNeSC. Existing AI methods and features will be updated, and
new AI methods and features added. Present AI methods include heuristic,
statistical, fuzzy logic, swarm and agent approaches.

Any AI method that adequately solves a required task will be seriously
considered, but only those methods that function with minimal or no human
administrative setup, intervention or steering will be used in RNeSC and
other RightNow products. RightNow plans to update existing features and
include new functionality to address deeper NLP analysis, Answers
7

relatedness, user action prediction, information representation and
visualization.

Various techniques may be used for additional maintenance features,
including any of the standard machine-learning approaches. Maintenance
features may also incorporate various types of neural networks, which
perform well on clustering and pattern recognition. The auto-cleanup and
noise reduction approaches under consideration include not only several
statistical methods, but also genetic algorithms.

Aside from focusing on organic AI approaches, the requirements of each new
task must also dictate the understandability of the results. When humans
must interpret the results, techniques such as fuzzy logic and statistical
analysis usually provide the most transparency. However, on tasks where no
human analysis is required, conceptually opaque techniques such as neural
networks might prove more appropriate.

Information Gap Analysis

The concept of an information gap is important within the domain of
knowledge base management. Traditionally, RNeSC has relied on human
workers to identify repetitive questions that should be created as public
Answers. However, automated approaches are more likely to identify
frequently asked questions distributed across a number of support personnel.
Similarly, this technique applies equally well to identify support workload and
indicate root causes of problems entering the RNeSC system.

Identifying missing information in a knowledge base is an extension to the
information browsing techniques described above. In addition to using these
techniques to identify groups of Answers, the same techniques must be
applied to identify groups of customer questions entering the system.
Finally, these two information groupings must be compared to reveal
discrepancies between the public and private information that indicate an
information gap.

Visualization

A key element of any eService application is ease of use. Unfortunately,
many of these AI-based approaches tend to deliver complex results to the
customer. When the results are not exactly what the customer requested, it
can be confusing. Therefore, any suite of AI tools used to help untrained
users must include visualization tools to aid user interpretation and
understanding of results. RNeSC will include a map of the knowledge base
information space to help beginners find solutions in large knowledge bases.
The first information map likely to become part of the RNeSC product is
based on similar approaches as Kohonen Self-Organizing Maps. Kohonen
Self-Organizing Maps are a variant of neural networks that plot information
in the form of topographic maps.

8

Blue Sky

Ease of use is the goal for all products in the RNeSC suite. With the field of
AI expanding and evolving, most techniques fitting the organic approach
touted by RightNow Technologies will fall within the RightNow product line.
These techniques range from various types of statistical methods for
machine-learning and fuzzy logic, to various neural and Bayesian networks,
genetic algorithms and beyond. Research into AI continues at institutions
around the world and our all-important “ease of use” will be enhanced as
new methods and techniques are developed.

9

About the Author

Doug Warner is one of the original software developers for RightNow
Technologies, a Bozeman-based provider of Web customer service systems.
Doug holds a BS in Computer Science and Psychology from New Mexico
Institute of Mining and Technology, a MS in Psychology from the University of
New Mexico, and has nearly completed his PhD in Psychology and Computer
Science from the University of New Mexico. He has taught a wide variety of
undergraduate and graduate courses at the University of New Mexico and
Montana State University. Doug has authored or co-authored seven academic
papers and has one granted patents and seven pending patents relating on
topics addressed in this paper. Prior to RightNow Technologies, Doug was a
member of a team researching approaches to work with knowledge bases in
virtual reality and parallel computing environments.

About RightNow Technologies

RightNow Technologies, a recipient of UPSIDE Magazine’s 2002 Hot 100
Private Companies Award, is the leading eService solutions expert,
engineering business solutions that deliver rapid time-to-benefit and quick
return on investment. RightNow delivers these benefits to more than 1,000
customers such as: Air New Zealand, Ben & Jerry's, British Airways, Cisco,
Fujitsu, Maxtor, Orbitz, Remington, Sanyo and more than 100 public sector
clients including the Social Security Administration and the State of Florida.

RightNow's multi-channel eService suite, which is Section 508 certified,
supports Web-based self-service, email response management, live chat and
collaboration, reporting and service metrics. RightNow Locator, which directly
links a company’s Web presence with its real-world locations, provides
customers the information they need to purchase products or obtain services
locally.

Founded in 1997, RightNow has offices in Bozeman, Dallas, London, and
Sydney, with an associated office in Tokyo. RightNow's products are available
in 14 languages worldwide. For further information visit
http://www.rightnow.com/.

For further information visit www.rightnow.com or email info@rightnow.com.

RightNow Technologies, Inc. North American Sales –West UK–International Sales
40 Enterprise Blvd phone +1-406-522-4200 phone +44 (0)1753 89 4900
PO Box 9300 toll free 1-877-363-5678 fax +44 (0)1753 89 4901
Bozeman, Montana 59718-9300 fax +1-406-522-4208
phone +1-406-522-4200 Australia – International Sales
toll free 1-877-363-5678 North American Sales –East phone +69 2 9657 1366
fax +1-406-522-4227 phone +1-972-323-5600 fax +69 2 9657 1353
toll free 1-877-277-3898
fax +1-972-247-4055

10

Web servicesknowledgebasetechnology

More Related Content

What's hot (11)

Viewers also liked (7)

Similar to Web servicesknowledgebasetechnology (20)

More from Automotive Digital Marketing Professional Community (20)

Web servicesknowledgebasetechnology