A message understanding and response system recognizes and answers messages based on the message writer's intent in unconstrained natural language text messages. The system has a set of knowledge bases with linked domain specific words, phrases, and regular expressions relating to the domain of the writer and the domain of the respondent. The writer's domain is represented by special purpose lexicons linked to representations of typical intents. The typical intents are linked to a domain knowledge base of typical and appropriate respondent actions. The system is initialized by manually classifying a training text corpus according to the respondent's policies. A lexical analysis tool with prototypical intents and phrases indicating intents is applied to the training text corpus, which includes the domain specific characteristics of both the writer and the respondent. The output results are an operable knowledge base which is a conjunction of keywords used to communicate between the two domains of the writer and the respondent. During automatic operation, the input text is pre-processed to remove irregularities in a manner similar to how the data in the training text corpus was regularized. Sets of extracted keywords and concepts are matched against the sets of stored, pre-classified keywords and concepts, producing a list of intents. The intents and other extracted features are then mapped to appropriate actions as defined by the system operator. The actions use the common linked domain knowledge terms to formulate a textual reply which is tailored to and answering the intent of writer of the input message.