Natural Language Generation Systems

Authors

Publication date 5 June 1993

Language generation systems natural language generation natural language systems text generation. Natural language generation systems

Mathematics Subject Classification ID

Proceedings, conferences, collections, etc. pertaining to computer science (68-06) Natural language processing (68T50)

Review text

The book contains a collection of ten articles dealing with various aspects of natural language generation systems. The first paper (by R. Gabriel) addresses the problem of so-called ``deliberate writing, i.e. writing considered as a creative, carefully planned and executed intellectual activity as, for instance, opposed to spontaneous speech. Yh, an object-oriented system that implements an explicit theory of deliberate writing, is presented with emphasis on generation as planning, and the revision of plans as an incremental process driven by critics. An attempt to account for various facets of discourse and text structure underlies the development of ``Rhetorical Structure Theory (RST) and the ``schema approach of McKeown's TEXT system. The basic methodological assumptions and suitability for text structure description of both approaches are compared in the contribution of W. Mann. D. Appelt's paper presents a logical possible-worlds formalization of the fundamental notions underlying planning to communicate (the speaker's goals and intentions, actions to indicate reference, etc.). These principles are incorporated in the KAMP system, a hierarchical, critic- based planner which produces complex English sentences that are tailored to what the hearer does and doesn't know including referring expressions as part of an agent's plan. The SUTRA generation system (originally developed as part of the HAM-ANS dialogue system) is designed as a reusable module that works within a larger generation system for German language. Its syntactic knowledge of sentence constituent placement and clause structure, and morphological knowledge of inflection paradigms is considered to be domain-independent and used by meaning-preserving surface transformation rules that map verbalized representation structures (containing a lexically and grammatically specified German language expression) into grammatically well-formed natural language output (S. Busemann). Another German language generation system, VIE-GEN, constitutes a major part of the VIE-LANG dialogue system. E. Buchberger and H. Horacek describe the representation of its linguistic (particularly lexical) knowledge that allows it to be used bi-directionally, i.e. for both parsing and generation. VIE-GEN starts with information given in a structural inheritance network of semantic primitives, and produces fluent text based on the selection of proper concepts from the network, intermediate representations containing word roots and unordered, case- marked constituents, and the application of transformation rules. In particular, the system's capabilities include the formation of conceptual paraphrases, the production of alternative word orderings, and appropriate strategies to create coherent text (e.g., through the generation of anaphora and gapping). The generation of German language as part of the SUSY machine translation system is considered by H.-D. Luckhardt. Generation is here directed by a language-neutral interlingua structure (used as part of the EUROTRA European MT project), which is essentially a deep syntactic structure. Luckhardt discusses the differences in issues and focus between generation in the artificial intelligence and machine translation community, and provides an extensive treatment of the theoretical issues surrounding analyses based on case theory within the formal framework of dependency (valency) grammar. A system that parses Spanish newspaper articles and generates (meaning- equivalent) Japanese stories is presented by S. Ishizaki. It is working from a completely conceptual representation basis (conceptual dependency with MOPs and causal chains organizing larger-scale events) and deals with a number of features in the Japanese language that have no direct counter-parts in English (e.g., strict temporal ordering of events with coordinated tenses). K. Kukich elaborates on the ANA stock market summary generator. It is based on a set of production rules that organize a large phrasal lexicon composed of complex phrasal units (and in this sense these rules constitute a ``clause-combining, rather than a ``sentence building grammar). ANA handles the entire generation process and the conceptualizations underlying it - to the point of starting from the raw statistics of the day's stock transactions, noticing relevant patterns in these data, and determining the appropriate conventional phrases for describing them in natural language. Similarly, P. Jacobs' PHRED system is centered on the use of pre-formed, specialized phrases, and designed in such a way that its knowledge of the English language could be directly shared with the parsing system PHRAN. Jacobs breaks down the generation process into a recursive sequence of fetching, restricting, and interpreting stages, whereby phrases are first retrieved and noticed applicable, then tested and specialized to the particulars of the concepts involved by way of constraint checking, and finally read out as a text. Finally, E. Hovy develops a functional linguistic perspective on the generation process that leads him to conclude that the entire linguistic knowledge (including the formative constraints usually expressed by syntactic rules) should be captured in the lexicon. This lexicalist perspective is formalized as a set of specialist procedures which determine the applicability of ``phrases (fixed in form and noncompositional in meaning) in context. The specialists which constitute a distributed phrasal grammar form the text generation system PAULINE.

Recommendations

Cited in

(24)

This page was built for publication: Natural Language Generation Systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4040438)