Catalogue of Artificial Intelligence Techniques
Semantic Head-Driven Generation
Aliases: Head-driven Bottom-up Generation
Keywords: generation, natural language generation, non-terminal symbol
Categories: Natural Language
Author(s): Nicolas Nicolov
Semantic Head-Driven Generation is a technique for generating a sentence (or a phrase) from its corresponding semantic representation. The algorithm traverses the derivation tree of the sentence in a semantic-head-first fashion. Starting from a node (root), which contains the input semantics and some syntactic information (like the category of the phrase to be generated), the algorithm selects a grammar rule in which the semantics of the mother node is identical to the semantics of the root. If the selected grammar rule passes the semantics unchanged from the mother node to a daughter node (chain rule), the daughter node becomes the new root and the algorithm is applied recursively (this is the top-down prediction step). If a grammar rule is reached which does not pass the semantics unchanged from the mother node to a daughter node (a non-chain rule) then the grammar rule is expanded and the algorithm is recursively applied to the list of the daughters (the mother node of the non-chain rule is called the pivot). The top-down base case occurs when the non-chain rule has no non-terminal daughters; that is, it introduces lexical material only. When all daughters of a non-chain rule have been generated, the algorithm attempts to `connect' the pivot to the root. This is done in a bottom-up fashion using chain rules. Larger and larger trees are built which contain the tree dominated by the pivot. This proceeds until the top node of the new tree is compatible with the initial root. The Semantic Head-Driven Generation algorithm is an illustrative instance of a sophisticated combination of top-down prediction and bottom-up structure building. It properly handles left-recursive rules and semantically non-monotonic grammars which have been problematic for previous purely top-down and bottom-up generation strategies respectively. The overall processing is directed by the input logical form and the information in the lexicon which leads to a goal-directed generation with good performance results.
No references to display.