A Statistical, Grammar-Based Approach to Microplanning

A Statistical, Grammar-Based Approach to Microplanning Claire Gardent∗ CNRS and Universit´e de Lorraine Laura Perez-Beltrachini∗∗ CNRS and Universit´e de Lorraine Although there has been much work in recent years on data-driven natural language generation, little attention has been paid to the fine-grained interactions that arise during microplanning between aggregation, surface realization, and sentence segmentation. In this article, we propose a hybrid symbolic/statistical approach to jointly model the constraints regulating these inter- actions. Our approach integrates a small handwritten grammar, a statistical hypertagger, and a surface realization algorithm. It is applied to the verbalization of knowledge base queries and tested on 13 knowledge bases to demonstrate domain independence. We evaluate our approach in several ways. A quantitative analysis shows that the hybrid approach outperforms a purely symbolic approach in terms of both speed and coverage. Results from a human study indicate that users find the output of this hybrid statistic/symbolic system more fluent than both a template-based and a purely symbolic grammar-based approach. Finally, we illustrate by means of examples that our approach can account for various factors impacting aggregation, sentence segmentation, and surface realization. 1. Introduction When generating a text, many choices must be made. The content to be expressed must be selected (content selection) and structured (document planning). Content must be distributed into sentences (sentence segmentation). Words (lexicalization) and syntac- tic structures (surface realization) must be chosen. Appropriate referring expressions must be identified to describe entities (referring expression generation). Coordinated and elliptical constructs may be exploited to omit repeated information (Aggregation). ∗ LORIA Campus Scientifique BP 239, F-54506 Vandoeuvre-l`es-Nancy Cedex, France. ∗∗ LORIA Campus Scientifique BP 239, F-54506 Vandoeuvre-l`es-Nancy Cedex, France. E-mail: laura.perez@loria.fr. Submission received: 25 March 2015; Revised version received: 15 January 2016; Accepted for publication: 14 March 2016. doi:10.1162/COLI a 00273 © 2017 Association for Computational Linguistics Published under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license Computational Linguistics Volume 43, Number 1 These decisions interact and are subject to various constraints. Consider for instance the content sketched in Example (1). There are many ways of verbalizing this con- tent1 but the appropriate choice depends on the context. For instance, the elision form (1l-m) is only appropriate in a context where another sell literal is present (e.g., car(x) sell(x,y) sportsCar(y) sell(x,z) trucks(z)). In this case, the repeated sell predicate can be elided (A car dealer selling sports cars and trucks). (1) CarDealer ⊓ ∃sell.(Truck) a. The car dealer should sell trucks. (Canonical Clause) b. It should sell trucks. (Canonical Clause with Pronominal Subject) c. and the car dealer should sell trucks (And S-Coordination) d. and it0 should sell trucks (And S-Coordination with Pronominal Subject) e. The car dealer who should sell trucks (Subject Relative) f. The car dealer (...) and who should sell trucks (And Subject Relative) g. The car dealer (...), who should sell trucks (Comma Subject Relative) h. The car dealer selling trucks (Gerund) i. The car dealer (...) and selling trucks (And Gerund) j. The car dealer (...), selling trucks (Comma Gerund) …

Content truncated to 500 words