A rule-based system responds by applying rules written in advance. In a chatbot, it might detect a word, phrase, or type of problem and select a prepared reply. This differs from a large language model (LLM), which generates responses probabilistically rather than selecting only from authored messages. section 1.1 section 9.5
Two kinds of chatbot
ELIZA used simple pattern matching, often asking open-ended questions or giving generic replies when it did not recognize a phrase. Woebot could identify a user's challenge and respond with a message written by its team and reviewed by clinicians. Neither example means the system understands a person as a therapist would: people attributed empathy to ELIZA despite its simple mechanism. section 1.1 section 9.5
Authored responses offer a reason to consider rule-based chatbots for sensitive settings: their available messages can be reviewed before use. That does not establish that they are suitable mental health treatments. Their conversations are also less flexible and may feel less engaging than conversations with LLM-based chatbots. Conversely, even when an LLM is trained to respond in particular ways, ensuring that every reply follows mental health guidance is difficult. section 9.5
Rule-based approaches have a broader history in natural language processing. Building them for language tasks required extensive if-then logic and linguistic expertise, and the results could be brittle. By the 1990s, statistical methods were achieving strong machine-translation results by learning patterns from translated documents instead of relying on people to specify grammatical rules. section 1.1
