📄 Design of a Knowledge-Based Report Generator

Summary

The paper presents knowledge-based report generation, a way to turn numeric databases into natural-language summaries such as daily market reports, weather bulletins or sales overviews. The first application is Ana, a partially implemented stock report generator. It takes Dow Jones stock quotes and writes the opening paragraphs of a market summary. The design has four sequential modules: a fact generator, a message generator, a discourse organizer and a text generator. All modules except the first are written as OPS5 production rules. Generation works with large linguistic units: complete semantic messages, a phrasal lexicon of ready-made phrases, clause-level syntactic categories, and a grammar that combines clauses. Ana produces fluent and varied paragraphs from the same data, and the author shows two different reports for 12 January 1983. Domain knowledge must be rebuilt for each new domain. The text-generation part, however, is about a quarter of the code and can be moved to other domains. This points to a practical path toward tailorable, extensible report generators for periodic summaries.

Key points

• Architecture: The pipeline has four modules: fact generator, message generator, discourse organizer and text generator.
  • The fact generator is a C program. The other three modules are OPS5 rules.
  • Module 2 had 120 rules and module 3 had 16 rules.
  • Module 4 had 109 rules plus a phrasal dictionary of 519 entries.
• Running time: On a lightly loaded VAX 11/780, the four phases took 16 s, 34 s, 24 s and 1 min 59 s.
• Example inference rule: The message generator derives messages from facts. If the Dow closed up but NYSE declines exceeded advances, the system concludes the market was "mixed."
• Text generator tasks: It chooses phrases and syntactic forms (sentence, participial clause, prepositional phrase, etc.). It also handles anaphora for subjects, verb morphology, punctuation and multi-clause sentences.
• Parameter bank: Rhetorical settings (maximum about 30 syllables and 3 messages per sentence) and weights on syntactic forms shape the style. Varying the weights produced the two different sample texts.

Interesting ideas

• I1. Separate deciding what to say from saying it. Data becomes facts, facts become messages, and only then is text produced. Isolating the language layer this way lets it be reused across domains, which helps any data-to-text system such as trade statistics or regional reports.
• I2. Phrasal lexicon instead of word-by-word assembly. Each lexicon entry is a whole domain phrase, stored with its verb forms and tagged with the meaning it expresses. This yields fluent sublanguage text cheaply in narrow domains, at the cost of some flexibility.
• I3. Weighted random choice of syntactic forms. Each sentence-building option carries a weight that can be adjusted. Small parameter changes alter the style, and repeated runs avoid monotonous output, which is useful for generators of recurring reports.
• I4. Default message ordering with exception rules. Messages follow a standard order, but special rules raise the priority of noteworthy events such as record highs. This simple mechanism lets salient information come first.
• I5. The "gear-shifting" view of generation. The system works mainly with whole stored phrases. When no phrase fits, it would fall back to adapting a near match or building a phrase from smaller units. This suggests a fallback hierarchy for template-based generation systems.