A Quantitative Analysis Of Different Text Type Features In English Under The Framework Of Dependency Grammar

Dependency Grammar describes binary,asymmetric and labeled syntactic relations between words in a sentence.Dependency distance is considered as an important concept in the field of dependency grammar,which refers to the linear distance between two syntactically related words in a sentence.Substantial studies have demonstrated that mean dependency distance(MDD)can detect syntactic feature of a sentence in linear dimension.Additionally,the hierarchical structure of a sentence is also a significant feature of human language,but previous studies haven’t pored over the DD in hierarchical dimension.Therefore,in terms of MDD,a new quantitative indicator,mean hierarchical dependency distance(MHDD),is proposed and employed here to manifest syntactic features in hierarchical dimension.In terms of the foregoing two indicators,the present study takes a quantitative approach to investigate the syntactic features of different text types in English from both linear and hierarchical dimensions,based on the data source selected from 12 text types of Freiburg-Brown corpus of American English.To examine the progressive properties of MDD and MHDD in 12 text types with the change of sentence length,we randomly selected 360 sentences from each text type,with sentence length ranging from 5 to 40 words(10 sentences per sentence length).The entire research treebank has a total of 4320 sentences and 97200 words.All sentences are first parsed by Stanford Parser and then manually checked corresponding to the Prague Dependency Treebank annotation scheme.Specifically,the current study intends to address the following three questions:(1)What are the distributions of MDD and MHDD of 12 text types?(2)What is the relationship between MDD and MHDD of 12 text types?(3)Do the distributions of dependency relations in terms of MDD or MHDD distinguish different text types?The present study yields the following results:(1)The distributions of MDD and MHDD of 12 text types form a fluctuating continuum.MHDD reflects the syntactic complexity of different text types in the same fashion as that of MDD.But it is not appropriate to generalize the syntactic complexity of text types merely based on distributions of MDD and MHDD,except for those on the polar points.(2)There is a significant high positive relationship between MDD and MHDD in 12 text types.MDD and MHDD of 12 text types all obey certain power function with different power exponents.The growth rate of MHDD with MDD displays a ‘narrative vs.expository’ distinction in 12 text types.The significant high positive correlation of MDD and MHDD in 12 text types may first lie in the overlap of numerator and denominator of the two algorithms.The fact that the largest proportions of data points in MDD and MHDD remain at low spectrum further demonstrates their significant high positive correlation.(3)Kruskal-Wallis test demonstrates the possibility of generalizing the syntactic complexity of 4 macro-domains if the 12 text types are categorized into 4 macro-domains as they are in Frown.On the cline of expository to narrative domains,all dependencies in academic are constrained to short-to-middle-range DDs and beneath relatively low hierarchies,indicating its preference for nominal style,flatter structures and denser style of writing.Non-fiction may not be that nominalized as academic,as proportions of some modifying dependencies climbs to higher hierarchies.Nominal information mainly characterized by dependencies of noun modifiers in academic is expressed by clausal dependencies in news.All dependencies in fiction show an apparent downswing trend to short DDs and low hierarchical levels.Whatever the text type is,DDs and HLs less than 10 always dominate the most majority of all dependencies.The present study has demonstrated the feasibility of measuring DD hierarchically,and systematically investigated the syntactic features of different text types on the correlated linear-hierarchical distance.The syntactic features of 4 macrodomains have been generalized based on the percentages of key dependencies both linearly and hierarchically.It is hoped that the methods and the findings of this study may shed some light on future quantitative researches on textual features or language families.