VIE-GEN A Generator for German Texts

Overview

VIE-GEN is a generator that produces German text from a semantic representation. It is a component of the German language dialogue system VIE-LANG [2], implemented in INTERLISP. The input to VIE-GEN is part of the episodic layer of the semantic network SEMNET, its outputs are German sentences. The generator is not restricted to single sentences, it contains features for creating coherent structures (e.g. generation of anaphora and gapping). VIE-GEN is designed to suit the idiosyncracies of the German language: it is able to produce various alternative word orderings (in German word order is not as strict as e.g. in English), it considers syntactic differences between main clauses and dependent ones and it is able to correctly produce all inflectional forms being found in German.

In order to be able to handle all these features, VIE-GEN performs its task in two steps which shall be referred to as verbalization phase and realization phase, respectively. The input to the verbalization phase is a part of the episodic layer of the semantic net, which is to be supplied by the dialogue component. This reflects the nearly classical distinction between “what to say” (decided by the dialogue component) and “how to say it” (decided by the generator) (e.g. [11]). The verbalization process is strongly data driven, its main sources of information being discrimination nets (DNs) and the so-called syntactico-semantic lexicon (SSL). By application of the DNs and by evaluating the SSL an intermediate structure (IMS) is created, forming the input to the realization phase.

The IMS is a tree, whose nodes represent either single words (terminals) or groups of words, i.e. constituents (nonterminals) together with their features. The “1exeme”-property of a node contains the canonical form of the word (actually a pointer to a lexicon entry so that morphological data can be accessed in the last step of processing) or, in the non-terminal case, a list of pointers to the dependent nodes. Admissible features are the complement type, an identifying marker that carries information about the individuated source concept (in SEMNET) the node stands for (used as a link for phrase heads and playing an important role in the generation of anaphora and gapping), number, tense, preposition (prepositional phrases are treated like noun phrases, the only difference being in the preposition feature being bound to a non-NIL value) and some others. Details on the IMS are to be found in a separate chapter below.

The task of the realization phase is the production of surface sentences out of the IMS. This task is divided into the following subprocedures:

• The predicates (in the grammatical sense) of the IMS are split into their finite and non-finite parts (in German, the predicate often is a discontinuous constituent)

• All constituents of a sentence are sorted in a standard order. The most interesting parts of this subtask are the partitioning of the IMS into sentences, maintaining their connections, and selection of appropriate types of sentences.

• Transformations are applied to these sentences, reordering the constituents in case of questions and subordinate phrases and the generation of gapping or anaphora.

• Noun phrases are linearized.

• In a last step, morphologic synthesis is performed.