Instance-based Sentence Boundary Determination by Optimization for Natural Language Generation

Proceedings of the 43rd Annual Meeting of the ACL, pages 565–572, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Instance-based Sentence Boundary Determination by Optimization for Natural Language Generation Shimei Pan and James C. Shaw IBM T. J. Watson Research Center 19 Skyline Drive Hawthorne, NY 10532, USA {shimei,shawjc}@us.ibm.com Abstract This paper describes a novel instance- based sentence boundary determination method for natural language generation that optimizes a set of criteria based on examples in a corpus. Compared to exist- ing sentence boundary determination ap- proaches, our work offers three signifi- cant contributions. First, our approach provides a general domain independent framework that effectively addresses sen- tence boundary determination by balanc- ing a comprehensive set of sentence com- plexity and quality related constraints. Second, our approach can simulate the characteristics and the style of naturally occurring sentences in an application do- main since our solutions are optimized based on their similarities to examples in a corpus. Third, our approach can adapt easily to suit a natural language gen- eration system’s capability by balancing the strengths and weaknesses of its sub- components (e.g. its aggregation and re- ferring expression generation capability). Our final evaluation shows that the pro- posed method results in significantly bet- ter sentence generation outcomes than a widely adopted approach. 1 Introduction The problem of sentence boundary determination in natural language generation exists when more than one sentence is needed to convey multiple concepts and propositions. In the classic natural language generation (NLG) architecture (Reiter, 1994), sen- tence boundary decisions are made during the sen- tence planning stage in which the syntactic struc- ture and wording of sentences are decided. Sentence boundary determination is a complex process that directly impacts a sentence’s readability (Gunning, 1952), its semantic cohesion, its syntactic and lex- ical realizability, and its smoothness between sen- tence transitions. Sentences that are too complex are hard to understand, so are sentences lacking seman- tic cohesion and cross-sentence coherence. Further more, bad sentence boundary decisions may even make sentences unrealizable. To design a sentence boundary determination method that addresses these issues, we employ an instance-based approach (Varges and Mellish, 2001; Pan and Shaw, 2004). Because we optimize our so- lutions based on examples in a corpus, the output sentences can demonstrate properties, such as simi- lar sentence length distribution and semantic group- ing similar to those in the corpus. Our approach also avoids problematic sentence boundaries by op- timizing the solutions using all the instances in the corpus. By taking a sentence’s lexical and syntac- tic realizability into consideration, it can also avoid sentence realization failures caused by bad sentence boundary decisions. Moreover, since our solution can be adapted easily to suit the capability of a natu- ral language generator, we can easily tune the algo- rithm to maximize the generation quality. To the best of our knowledge, there is no existing comprehen- sive solution that is domain-independent and pos- sesses all the above qualities. In summary, our work offers three significant contributions: 1. It provides …

Content truncated to 500 words