GEMS: A Model Of Sentence Production

The paper describes GEMS, a system for Generating and Expressing the Meaning of Sentences, focussing on the generation task, i.e. how GEMS extracts a set of propositional units from a knowledge store that can be expressed with a well-formed sentence in a target language. GEMS is lexically distributed. After a central processor has selected the first unit(s) from the knowledge store and activated the corresponding lexical entry, the further construction of the sentences meaning is entrusted to the entries in the vocabulary. Examples of how GEMS constructs the meaning of a number of English sentence types are briefly described. I . Constructing the meaning of sentences Most work on natural language generation has been concerned with the production of connected text (Davey, 1979; Goldman, 1975; Mann and Moore, 1981; Meehan, 1977) or with language generation as a goal-directed, planned activity (Appelt, 1980; Mann and Moore, 1981). Less attention has been dedicated to the linguistic details of sentence generation, i.e. to constructing a general device for imposing the appropriate linguistic form to the content that must be expressed (but see Kempen and Hoenkamp, 1982). The aim of this paper is to describe GEMS, a system for Generating and Expressing the Meaning of Sentences. GEMS takes a store of knowledge as input and gives English sentences expressing that knowledge as output. The knowledge contained in the knowledge store is purely conceptual knowledge with no trace of linguistic form. There is no partitioning of knowledge in parts which can be expressed by single sentences or by single lexical items, no grammatical labelling of items as verbs, nouns, or subjects, objects, etc., no other traces of syntactic or lexical form. Hence, a first task of GEMS is to extract from the knowledge store the knowledge which it is appropriate to express in a well-formed sentence, i.e. to generate the meaning of the sentence. Since the meaning thus constructed must be expressed with a specific sequence of words, two further tasks of GEMS are to select the semantic and grammatical morphemes that make up the sentence and to put them in the appropriate sequential order. Producing sentences is a goal-directed activity: what one says depends on one's goals. GEMS however is a model of how to say something, not of what to say. When it arrives at a decision point on what to say, GEMS makes a random choice. Hence, GEMS is not a complete model of the activity of producing sentences but only a model of the linguistic constraints on the communication of knowledge and ideas. GEMS conceives the knowledge necessary to produce sentences as largely distributed in the lexicon. This change from previous more centralized version of GEMS (see Parisi and Giorgi, 1981; 1983) has been suggested to us by Oliviero Stock and Cristiano Castelfranchi and it is related to our view of a lexically distributed sentence comprehension process (see Stock, Castelfranchi, and Parisi, 1983; Parisi, Castelfranchi, and Stock, in preparation). The lexical entries are procedures that activate each other in a given order when a sentence is produced, although the order of activation may not coincide with the external sequential order of the words in the actual sentence. When executed the entries' procedures (a) extract the sentence's meaning from the knowledge store, (b) lexicalize this meaning with the appropriate semantic and grammatical morphemes, and (c) put these morphemes in the correct sequential order. A central processor has the task of searching the knowledge store for knowledge to be expressed and the lexicon for the lexical entries that can express this knowledge. However, the main task of the central processor is to start the construction process and to keep a record of the order of activation of the lexical entries. The overall scheme of GEMS is represented in Fig. 1