Two-Level, Many-Paths Generation

Two-Level, Many-Paths Generation Kevin Knight USC/Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292 knight@isi.edu Vasileios Hatzivassiloglou Department of Computer Science Columbia University New York, NY 10027 vh@cs.columbia.edu Abstract Large-scale natural language generation re- quires the integration of vast mounts of knowledge: lexical, grammatical, and concep- tual. A robust generator must be able to operate well even when pieces of knowledge axe missing. It must also be robust against incomplete or inaccurate inputs. To attack these problems, we have built a hybrid gen- erator, in which gaps in symbolic knowledge are filled by statistical methods. We describe algorithms and show experimental results. We also discuss how the hybrid generation model can be used to simplify current generators and enhance their portability, even when perfect knowledge is in principle obtainable. 1 Introduction A large-scale natural language generation (NLG) system for unrestricted text should be able to op- erate in an environment of 50,000 conceptual terms and 100,000 words or phrases. Turning conceptual expressions into English requires the integration of large knowledge bases (KBs), including grammar, ontology, lexicon, collocations, and mappings be- tween them. The quality of an NLG system depends on the quality of its inputs and knowledge bases. Given that perfect KBs do not yet exist, an impor- tant question arises: can we build high-quality NLG systems that are robust against incomplete KBs and inputs? Although robustness has been heavily stud- ied in natural language understanding (Weischedel and Black, 1980; Hayes, 1981; Lavie, 1994), it has received much less attention in NLG (Robin, 1995). We describe a hybrid model for natural language generation which offers improved performance in the presence of knowledge gaps in the generator (the grammar and the lexicon), and of errors in the se- mantic input. The model comes out of our practi- cal experience in building a large Japanese-English newspaper machine translation system, JAPAN- GLOSS (Knight et al., 1994; Knight et al., 1995). This system translates Japanese into representations whose terms are drawn from the SENSUS ontol- ogy (Knight and Luk, 1994), a 70,000-node knowl- edge base skeleton derived from resources like Word- Net (Miller, 1990), Longman's Dictionary (Procter, 1978), and the PENMAN Upper Model (Bateman, 1990). These representations are turned into En- glish during generation. Because we are processing unrestricted newspaper text, all modules in JAPAN- GLOSS must be robust. In addition, we show how the model is useful in simplifying the design of a generator and its knowl- edge bases even when perfect knowledge is available. This is accomplished by relegating some aspects of lexical choice (such as preposition selection and non- compositional interlexical constraints) to a statisti- cal component. The generator can then use simpler rules and combine them more freely; the price of this simplicity is that some of the output may be invalid. At this point, the statistical component intervenes and filters from the output all but the fluent expres- sions. The advantage of this two-level approach is that the knowledge bases in the generator become simpler, easier to …

Content truncated to 500 words