Artificial Intelligence Terms and Definitions

General AI Terms and Definitions

Case Based Reasoning (CBR)

(NOT used by RightNow) This is a question answering technique where prior exemplar cases (literally from prior customer questions) are used to answer new, incoming questions. Comparisons between cases are complex, and usually customized to the topic domain of the questions. Cases are usually complex entities including fields intended to be read by the system, not humans. They are stored as question-action pairs. Strengths: can provide very accurate responses to previously seen inquiries. Weaknesses: responses are literally individual prior responses, thus more personalized and less generalized.

Expert System

(NOT used by RightNow) This is a question answering technique where the knowledge of an expert in the field is exhaustively queried, then represented as a set of probabilistic if-then rules. Strengths: can work with incomplete information and know how to query for more complete information; will answer as well as the expert. Weaknesses: brittle as the domain changes beyond what was initially extracted from the expert.

Decision Tree

(NOT used by RightNow) This is a question answering technique where the prior customer behavior (or sometimes the expected behavior) is modeled as a set of yes/no decisions leading to an answer. Strengths: works well for trouble-shooting Weaknesses: can become brittle if not kept up-to-date, yet is computationally intensive to keep up-to-date.

Natural Language Processing (NLP)

This is used in the context of searching, and refers to the computer's ability to understand human (natural) language. This term is mostly meaningless because it is loosely used to cover many techniques, from those that are quite trivial to those that are so complex they are still cutting-edge research. The techniques include, among others,

identifying (and then ignoring) common words (commonly called stop words),
recognizing multiple-word phrases,
identifying word roots (commonly called stemming) (for example, walks, walked, and walking are all based on the word root walk),
identifying part-of-speech (for example, 'bill' can be a proper noun, a noun, or a verb),
identifying and correcting misspellings,
using synonyms, and
identifying and using general versus specific terms (for example, vehicle is more general than automobile, which is more general than truck, which is more general than pickup, which is more general than Toyota Tacoma).

RightNow uses techniques 1-6 for various purposes in the system.

Clustering

(Used by RightNow) This is a technique to find groups of related information in data. There are many different clustering algorithms and approaches. Strengths: good way to automatically find groups of information. Weaknesses: usually computationally intensive.

Classification

(Used by RightNow) This is a technique to add new items to known groups. This can be used in conjunction with CBR, Decision Trees, Clustering, or many other techniques. There are many different classification techniques and approaches. Strengths: computationally efficient to properly place new items. Weaknesses: the groups must be known before-hand, and new items will always get placed in an existing group (even if it is a poor match).

Neural Networks (also Neural Nets, never Neural Networking)

(Used by RightNow) This is a computational learning technique to find patterns in data. The name comes from the fact that they are loosely modeled on the functional workings of biological neurons. There are numerous different types of neural networks. Strengths: good way to automatically find patterns in data. Weaknesses: computationally intensive, and no way for a human to interpret how the answer was derived.

Self Organizing Map (SOM)

(Used by RightNow) This is a particular type of Neural Network that is used to Cluster information. It can generate maps of an 'information landscape' keeping conceptually similar items in close proximity to each other. Strengths: results are displayable in a way that is easy for humans to interpret. Weaknesses: computationally intensive, and no way for a human to interpret how the layout was derived.

Swarm Intelligence

(Used by RightNow) This is a computational learning technique to find patterns in data. It is loosely based off of the concept of multiple entities competing for resources in a limited resource environment. The entities quickly identify the relevant bits of information (food) and can rapidly change to new information once it becomes more relevant (better food). Strengths: very robust, computationally distributed technique. Weaknesses: computation is distributed (ie, calculations are always approximate) and can take a long time to stabilize.

Bayesian Learning

(Used by RightNow) This is a computational learning technique to find patterns in data. It is explicitly a probability based system that has the probabilities generated from the statistics in the training data. Strengths: Computationally simple, fairly robust. Weaknesses: Susceptible to bad data, and results not as refined as more complicated techniques.

Spidering

(Used by RightNow) This describes the process of scanning the links in a web page to find documents to index.

Index (Search Index)

(Used by RightNow) This is a method for increasing the processing speed for searching. Documents are processed into their 'interesting' components, which are then individually stored in the index. When a search query is entered it is compared against the index, not the individual documents. Strengths: vastly increases searching speed. Weaknesses: there is usually information from the original that is not in the index, hence can't be searched on.

Industry Terms and Definitions

Knowledge Base (KB)

(Used by RightNow) This describes any collection of knowledge items, whether documents, cases (in CBR), or anything else. This term also implies at least one efficient means for extracting relevant knowledge items. In RightNow it refers to our answers (plus external documents) as well as our incidents, and we treat these three areas as if they were three separate knowledge bases. The extraction techniques we use for answers include

the initial set of top ranked answers,
four different search techniques, and
two different browse interfaces.

In external documents we just just one search style. In incidents we just use one search style.

Cross-sell/Up-sell Advisor (CUA) or Up-sell Cross-sell Advisor (UCA)

(Used by RightNow) This describes a product or a process for recommending items for sale or purchase discounts based off of individual customer characteristics (such as prior purchases, and demographics), as well as specific business rules (such as over-stock or under-stock).

RightNow Specific Terms

SmartAssistant

This is a name given to any RightNow technology that automatically provides answers in novel ways. Currently there are two SmartAssistant technologies in our products:

Suggested Solutions: This technique allows for automated responses to customer questions, either by a rule working via email or via the web page, or by request of the CSR while working an incident. This technique uses NLP and Classification (on our Clusters) to find close matches between a detailed customer inquiry and a small number of answers.
Related Answers: This technique automatically generates links between answers. These are visible only on the end-user, and currently appear at the bottom of an answer display page. This technique uses Bayesian learning and Clustering.

Similar Phrases (search)

This is the name we have given to our search approach that uses the more advanced NLP approaches of spelling correction and synonym expansion.