Dependency Based Chinese Sentence Realization

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 809–816, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Dependency Based Chinese Sentence Realization Wei He1, Haifeng Wang2, Yuqing Guo2, Ting Liu1 1Information Retrieval Lab, Harbin Institute of Technology, Harbin, China {whe,tliu}@ir.hit.edu.cn 2Toshiba (China) Research and Development Center, Beijing, China {wanghaifeng,guoyuqing}@rdc.toshiba.com.cn Abstract This paper describes log-linear models for a general-purpose sentence realizer based on de- pendency structures. Unlike traditional realiz- ers using grammar rules, our method realizes sentences by linearizing dependency relations directly in two steps. First, the relative order between head and each dependent is deter- mined by their dependency relation. Then the best linearizations compatible with the relative order are selected by log-linear models. The log-linear models incorporate three types of feature functions, including dependency rela- tions, surface words and headwords. Our ap- proach to sentence realization provides sim- plicity, efficiency and competitive accuracy. Trained on 8,975 dependency structures of a Chinese Dependency Treebank, the realizer achieves a BLEU score of 0.8874. 1 Introduction Sentence realization can be described as the process of converting the semantic and syntactic representation of a sentence or series of sen- tences into meaningful, grammatically correct and fluent text of a particular language. Most previous general-purpose realization sys- tems are developed via the application of a set of grammar rules based on particular linguistic theories, e.g. Lexical Functional Grammar (LFG), Head Driven Phrase Structure Grammar (HPSG), Combinatory Categorical Grammar (CCG), Tree Adjoining Grammar (TAG) etc. The grammar rules are either developed by hand, such as those used in LinGo (Carroll et al., 1999), OpenCCG (White, 2004) and XLE (Crouch et al., 2007), or extracted automatically from annotated corpora, like the HPSG (Nakanishi et al., 2005), LFG (Cahill and van Genabith, 2006; Hogan et al., 2007) and CCG (White et al., 2007) resources derived from the Penn-II Treebank. Over the last decade, there has been a lot of in- terest in a generate-and-select paradigm for sur- face realization. The paradigm is characterized by a separation between realization and selection, in which rule-based methods are used to generate a space of possible paraphrases, and statistical methods are used to select the most likely reali- zation from the space. Usually, two statistical models are used to rank the output candidates. One is n-gram model over different units, such as word-level bigram/trigram models (Bangalore and Rambow, 2000; Langkilde, 2000), or fac- tored language models integrated with syntactic tags (White et al. 2007). The other is log-linear model with different syntactic and semantic fea- tures (Velldal and Oepen, 2005; Nakanishi et al., 2005; Cahill et al., 2007). However, little work has been done on proba- bilistic models learning direct mapping from in- put to surface strings, without the effort to con- struct a grammar. Guo et al. (2008) develop a general-purpose realizer couched in the frame- work of Lexical Functional Grammar based on simple n-gram models. Wan et al. (2009) present a dependency-spanning tree algorithm for word ordering, which first …

Content truncated to 500 words