In this paper, Richard S. Wallace of the A.L.I.C.E. Artificial Intelligence Foundation gives a technical account of two things he created: A.L.I.C.E., the Artificial Linguistic Internet Computer Entity, and AIML, the Artificial Intelligence Markup Language used to author it. The abstract states that A.L.I.C.E. was “the first AIML-based personality program” and that it “won the Loebner Prize as ‘the most human computer’ at the annual Turing Test contests in 2000 and 2001.” The paper traces the history of the A.L.I.C.E. and AIML free-software effort back to 1995, noting that more than 500 volunteers around the world contributed to the bot’s development.
The core idea Wallace describes is deliberately simple. A bot’s “personality” is a set of AIML files made of stimulus-response modules called categories, each containing a pattern (the stimulus) and a template (the response). The software stores these categories in a tree managed by an object called the Graphmaster, and when a user types something, the Graphmaster searches for a matching pattern, takes context into account, and returns the associated template. A few markup tags, including the recursive srai tag and the context tags that and topic, let an author compose more complex, human-seeming replies out of these basic building blocks.
Wallace is candid that the strategy is one of “deception and pretense,” a lineage he traces through the history of AI back to ELIZA, whose pronoun-swapping trick AIML reimplements. The point is not that the bot understands, but that carefully arranged pattern-response rules can carry a surprising amount of conversation.
For a business reader, A.L.I.C.E. matters as the bridge between 1960s pattern-matching and the modern chatbot industry: it was open source, community-built, and template-driven, and its AIML approach powered countless customer-service and novelty bots for years before neural language models arrived.