The term artificial intelligence is used rather loosely these days. For an automatic system to possess true artificial intelligence it must learn by itself, and by learning is not meant the blind incorporation of information (e.g. User inputs), but rather the self-critical ability to select which data are relevant and which are not. Herein lies the dispute between the AI systems truly applying some form of learning and the dialogue simulation systems which learn by human intervention. The latter group is often represented in contests such as the Loebner Competition and the Chatterbox Challenge. This article sketches some of author's experiences with dialogue systems.

 

 

The term artificial intelligence is used rather loosely these days. For an automatic system to possess true artificial intelligence it must learn by itself, and by learning is not meant the blind incorporation of information (e.g. User inputs), but rather the self-critical ability to select which data are relevant and which are not. Herein lies the dispute between the AI systems truly applying some form of learning and the dialogue simulation systems which learn by human intervention. The latter group is often represented in contests such as the Loebner Competition and the Chatterbox Challenge. This article sketches some of my experiences with dialogue systems, in particular Elbot. Incidentally, if Elbot asked about the definition of artificial intelligence his response is: Artificial intelligence was conceived as the opposite of authentic intelligence. But since the latter does not actually exist, the developers of AI are in a serious dilemma.

Lingubots

Kiwilogic does not build artificial intelligences, but rather commercial dialogue systems with a well-defined area of expertise, usually company FAQs and Website navigation. This is a realistic application of current technology, as opposed to a system which is able to converse convincingly on every conversational topic with which it might be confronted. The systems Kiwilogic builds are designed to recognize large classes of inputs in all their synonymical variations and associate them with a desired response. The answers are written to conform to a meaning, and not a particular choice of words. This gives the author of a Lingubot complete control over the Lingubots character and range of expression, as the pool of answers is not influenced by any factor of erroneous learning. Lingubots have the ability to note topics of conversation and return to them, and are able to follow-up on typical responses given by a particular Lingubot answer.

The principal of the Lingubot technology is to bring automatic dialogues as close as possible into the realm of human conversation. This allows Lingubots to assume tasks as varied as those handled by a human employee. With the well-rounded technology for automatic dialogues in natural languages behind it, a Lingubot shows enormous flexibility in reliably assuming these roles.

Internally a Lingubot is a collection of recognitions (rules for recognizing an input) which are associated with responses. The recognitions are arranged in the order of complexity and are then checked in this order. Answers assigned based on a complex rule will fit more convincingly than answers of a recognition with low complexity. Several words occurring together have more meaning than a simple keyword, in which context is unknown.

A dialogue system learns by human intervention. The Lingubot Creator software is optimally geared towards this task. All dialogues with a Lingubot are transcribed for analytical reference. A variety of analysis functions allows the Lingubot author to meaningfully expand the knowledge base of the Lingubot in question, by offering different views into the collected conversations. For example, It is easily possible to view the user inputs which triggered an answer, with an eye towards reducing misunderstandings (correcting errors in recognition). The user responses to a particular Lingubot answer may also be viewed, allowing the meaningful expansion of dialogue trees having as many branches as are needed. An additional strategy for expanding the knowledge base of a Lingubot is to examine the inputs for the default or safety-net answers which are given when the Lingubot could not understand anything in the input. Among these user inputs will be found commonly asked questions, which the Lingubot has not yet been prepared to answer.

sum:

0 €

» Library of the Month
» Workshop
» Software Engineering
» Programming Tools
» Interview
» High quality and
secure programming
» Reverse Engineering
» Portable GUI
» Artificial intelligence