Text planning selects and organizes the propositions, events and states that should be described in a discourse, but a second intelligent process is needed to generate semantic and syntactic specifications that can actually be realized in natural language. This process is called sentence planning [KKR91]. The key function of sentence planning is to select and adapt linguistic forms so that they are suitable for the local context in which they are to appear.
One important issues in sentence planning is to determine the content of definite noun phrases that are able to refer uniquely to an object and thereby allow a hearer to identify the object quickly and naturally. Research that addresses this problem includes [App85,DH91,Rei91,RD92]. Names for machine-internal representations of objects must be replaced with key properties of the object. The content needed to refer uniquely to an object varies greatly, according to the salience or attentional availability of the object in context, as well as the salience of distractor objects with similar properties. In the best case, the object is the most salient entity of the appropriate type, and it can be described using a pronoun---with almost no content at all. On the other hand, if the object has not been mentioned in discourse at all, it may be necessary to provide a ``complete description'' of the object, including detailed information about its size, color, location and type. In the range of cases in between these extremes, the system must construct a reduced description, which should contain enough information to uniquely identify the object but should also be brief, so that the discourse remains clear and concise.
Of course, material meant to satisfy other purposes than unique referability may be included in a noun phrase as well -- for example, to convey an object's location to the hearer so they know where to look for it [App85], to convey an object's condition to the hearer so that they either take appropriate precautions (e.g., ``hot'') or do what's appropriate to achieve that condition (e.g., ``well-oiled''), etc. The assumption is that this material will have been selected during text planning: sentence planning decisions involve whether to allocate the information to separate sentences (producing more, but perhaps simpler sentences) or to include in noun phrases (producing fewer sentences, but perhaps involving more reasoning on the part of the hearer).
Similar choices appear to be involved in building other sentential constituents [SD96]. Like descriptions of entities in noun phrases, descriptions of actions in verb phrases, descriptions of events in dependent clauses, and temporal and spatial locations in adverbials vary depending on the context. A clear and concise discourse will include enough information about them so that the hearer can determine which is meant, but not so much that the text is confusing or hard to read. Moreover, generating descriptions serves as a natural paradigm for integrating a variety of choices about which lexical items and which syntactic structures should be used to realize a sentence. It is a useful software engineering move.
Sentence planning must be a separate stage from text planning for two reasons. The first is its dependence on salience, which is determined both by the structure of discourse and the syntactic and semantic structure of immediately preceding sentences. Salience is not available during text planning. The second is the basis for including content, which differs from that of text planning. In text planning, content is included based on complex, abstract operators that organize propositions into arguments. In sentence planning, as noted above, content is included in referring expressions by simple operators that add small units of information, so as to identify objects or achieve in a terse manner goals identified during text planning.
Nevertheless, like text planning, sentence planning requires a rich representation of the world and the hearer. World knowledge must provide an inventory of properties about each object that can be included in descriptions. Hearer knowledge must include how the hearer can use such properties to rule out alternatives to the object being described.