6.1. Problem Definition and Approaches to the Solution

Human-machine communication can be seen as an exchange of messages coded with methods that enable their transmission through a physical medium [MOR 98]. The process which produces a representation of the messages to be communicated is called coding. Messages are structured into words, themselves represented as sequences of symbols belonging to an alphabet and organized according to a given lexicon. Phrases are built from words using grammar rules and other constraints which constitute knowledge sources (KSs) [JUN 96].

Speech understanding and voice dictation systems perform a decoding process [BAH 83, DUG 95, JEL 76, JEL 80, JEL 97] using knowledge sources to transform the speech signal carrying the message into representations of this message at various levels of abstraction. The decoding process can produce word sequences or conceptual hypotheses.

As opposed to human-human communication, human-machine communication must produce data structures in a deterministic way. In this case, deterministic means that a computational system must generate the same representation each time the same signal is processed. Knowledge sources used by machines for decoding messages are only models of those used by humans for the same or similar purposes. One of these models is the language model [FED 95, KNE 93, KUH 90]. It models constraints based on which words are acceptable in a given language.

One of the most important difficulties in natural language processing by computers originates...