A Working Report on Statistically Modeling Dative Variation in Mandarin Chinese

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1236–1244, Beijing, August 2010 A Working Report on Statistically Modeling Dative Variation in Mandarin Chinese Yao Yao University of California, Berkeley Department of Linguistics yaoyao@berkeley.edu Feng-hsi Liu University of Arizona Department of East Asian Studies fliu@u.arizona.edu Abstract Dative variation is a widely observed syn- tactic phenomenon in world languages (e.g. I gave John a book and I gave a book to John). It has been shown that which surface form will be used in a dative sen- tence is not a completely random choice, rather, it is conditioned by a wide range of linguistic factors. Previous work by Bresnan and colleagues adopted a statis- tical modeling approach to investigate the probabilistic trends in English dative alter- nation. In this paper, we report a similar study on Mandarin Chinese. We further developed Bresnan et al.’s models to suit the complexity of the Chinese data. Our models effectively explain away a large proportion of the variation in the data, and unveil some interesting probabilistic fea- tures of Chinese grammar. Among other things, we show that Chinese dative varia- tion is sensitive to heavy NP shift in both left and right directions. 1 Introduction 1.1 Overview In traditional linguistic research, the study of syn- tax is most concerned with grammaticality. Sen- tences are either grammatical or ungrammatical, and syntactic theories are proposed to explain the structural features that cause (un)grammaticality. Meanwhile, little attention has been paid to the relative acceptability of grammatical sentences. If two sentences are both grammatical and basically express the same meaning, are they equally likely to occur in the language? The answer is proba- bly no. For example, in English, the sentence I have read that book is much more frequent than That book I have read. The latter topicalized sen- tence is only used when the entity denoted by That book is in focus. This indicates that the choice of surface sentence form is not entirely random, but conditioned by some factors including infor- mation status. Thus, instead of categorizing sentences as grammatical or ungrammatical, a better way to express the degree of grammaticality would be to use a likelihood continuum, from 0 to 1, where un- grammatical sentences have zero likelihood and grammatical sentences fall somewhere between 0 and 1, with some being more likely than oth- ers. The idea of associating linguistic forms with various probabilities has been around for a while (see Jurafsky, 2003 and Manning, 2003 for an ex- tensive review). Recent psycholinguistic research has shown that just like grammaticality, the likeli- hoods of sentence forms are also part of the user’s linguistic knowledge. Sentences with high proba- bilities are in general easier to comprehend and produce, and their production is more prone to phonetic reduction (Bresnan, 2007; Gahl and Gar- nsey, 2004; Levy, 2008; among others). The fa- mous example of garden path sentences also ex- emplifies the difficulty of comprehension in low- probability sentence forms. If we accept the premise …

Content truncated to 500 words