Using template-grammars for shake & bake paraphrasing

Using Template-Grammars for Shake & Bake Paraphrasing Michael Carl, Ecaterina Rascu and Paul Schmidt Institut f¨ur Angewandte Informationsforschung Saarbr¨ucken, Germany email:{carl,kati,paul}@iai.uni-sb.de Abstract In this paper we propose an approach to corpus-based generation in a machine translation framework that is similar to shake & bake (Whitelock, 1992). A bag of words is mapped against an automatically induced TL template grammar and a sentence is generated by recursively applying rules that are extracted from the template grammar. A test version of the template grammar is enriched with further lexical and grammatical variation patterns. We show how we induce a template grammar and how it is enriched with additional paraphrasing knowledge. We suggest a framework for weighing and training the template grammars and show that the enriched template grammars produce better paraphrases. 1. Introduction Paraphrasing is frequently used as an instrument to improve the output of various NLP applications. On the one hand, it is used to cope with the highly variable character of natural language in applica- tions like multi-document summarisation (Barzilay et al., 1999) where phrases reporting on the same fact have to be found in input documents. Moreover, in controlled language machine translation (Mitamura and Nyberg, 2001) replacing a source language (SL) item with a paraphrase that is better suited to the requirements of the MT engine boosts the perfor- mance of the system, whereas in information retrieval or question answering (Jacquemin et al., 1997; Ri- naldi et al., 2003) query expansion by means of para- phrases leads to increased retrieval efficiency. On the other hand, paraphrasing is used to control the gen- eration process within NLP applications. It may help produce the best-suited utterance in a given situation (Iordanskaja et al., 1991; Robin, 1994; Dras, 1999). In this paper we describe an approach to corpus- based generation within an MT framework in which paraphrasing is used as a verifying instrument and a means to improve the performance of the system. The approach developed in this paper is a contri- bution to METIS-II. In the EU-project METIS-II, a follow-up to METIS-I1 (Dologlou et al., 2003), the aim is to investigate the possibilities to develop a data-driven MT system using a huge monolingual target language (TL) corpus and a bilingual dictio- nary. While the dictionary is used to map SL items onto the TL, the corpus serves as a model to gener- ate the TL sentences. This parallels with shake & bake (S&B) (Whitelock, 1992). In S&B the bilin- 1http://www.ilsp.gr/metis2/ gual knowledge is exhausted by the equivalence of basic expressions and TL generation is under direct control of the TL grammar. This makes large scale structural reorganisation of the TL possible and over- comes the inherent problems in conventional third- generation transfer-based MT. In these latter systems, TL generation is merely a matter of traversing and printing out intermediate representations which are defined by the structure of the SL text (Whitelock, 1991). The S&B approach also entails that the transfer component and the generation component can be de- veloped and tested …

Content truncated to 500 words