ELIZA: Not Just Regex — The Architectural Depth of the First Chatbot
The article debunks the myth of ELIZA's primitiveness, showing that it was based on a universal natural language processing engine separate from behavior scripts. The author emphasizes that the architectural solutions introduced by Joseph Weizenbaum in 1965 anticipated modern approaches to creating dialogue systems.
An article on Habr examines the ELIZA system created by Joseph Weizenbaum in 1965. It debunks the myth that ELIZA was just a set of regular expressions: in reality, it was a full-fledged natural language processing engine where the processing logic was separated from the bot's scripts (skins). The main script was called DOCTOR and simulated a psychotherapist, but ELIZA could load any scripts with keywords and templates. The article details the inner workings: scanning input, collecting keywords with priorities (ranks), decomposition using templates with fragment capture, pronoun replacement (transformation), response reconstruction, using memory (a stack) to store transformations, and fallback scripts. It notes that this architecture—separating a universal mechanism from application logic—anticipated modern approaches to large language models (LLMs). Also discussed is the "Eliza effect": people's tendency to attribute understanding and empathy to programs, despite the machine having no real consciousness. A story is recounted about Weizenbaum's secretary who wanted to be alone with ELIZA. The article concludes with an invitation to test a restored version of ELIZA in a browser and discuss which modern architectural solutions will seem naive in 60 years.
Source: Habr — хаб ИИ —
original