“The Analysis of Classical Greek and Latin Compositional Word-Order Data”

Fiona J. Tweedie Department: Statistics University of Glasgow fiona@stats.gla.ac.uk Bernard D. Frischer Dept. of Classics University of California frischer49@aol.com

Introduction

A recent paper by Frischer et al. (1999) examines the position of the direct object and its governing verb in works in Classical Greek and Latin. Through statistical analysis of a great number of sentences, one long suspected difference between Latin and Greek word-order was confirmed, and the ramifications of this observation were explored for some possible cases of word-order transference between Latin and Greek. The difference between the languages concerns the positioning of the accusative direct object with respect to the verb governing it. That there is a difference in the Greek and Latin distributions is no surprise: Classical linguists have long observed that Latin has a greater tendency to place the verb at the end of the clause than does Greek. From this fact alone one might predict that the direct object in Latin is more likely to precede than to follow the verb on which it depends than is the case in Greek. This prediction was tested empirically by tabulating the direct object distributions in sixty passages written by fifteen Latin and ten Greek prose authors. Each passage was randomly selected in the text of an author. Analysis was based on the first one hundred direct objects in the accusative case that were encountered in a passage, and a tabulation was done of those that occurred before and those after the governing verbs. With remarkable consistency, the texts in our sample clumped into a Latin and a Greek cluster, offering strong empirical and statistically significant proof that the position of the direct object with respect to its governing verb differed in Greek and Latin prose. Five Greek texts by two authors writing in Greek turned out to be anomalous, fitting firmly into the Latin group. Four of the texts were written by Cassius Dio; the fifth is the Greek translation of the Emperor Augustus' Res Gestae, a bilingual version of which survives, and whose original is known to have been written in Latin. In considering the results, including the anomalous texts, the study showed that native language did not necessarily have an effect on a writer's placement of the direct object; nor did the language of an important literary or historical source. Native Greek authors writing in Latin respected Latin word order; Romans writing in Greek generally conformed to Greek practice. The study suggested that some but not all explanations for the data are linguistic. On the linguistic level, it was the greater consistency of Latin SOV word-order that helped the Latin pattern to prevail over the more flexible Greek positioning of the verb and direct object. This was true not only for Roman authors writing Latin with a Greek source before them (like Aulus Gellius or Cicero) but also for a Greek author like Ammianus Marcellinus writing in Latin. It was evidently normally easy for both Greeks and Romans to recognize and to respect the tendency of Latin to place the verb at the end of the clause. On the other hand, in the interesting case of the Greek translation of the Res Gestae and other official documents, where the Roman chancellery's habit of translating Latin into Greek through quasi-relexification was seen, the study proposed an explanation based either on Roman scrupulosity in legal matters or on a sociological factor of linguistic hegemony. Finally, in the case of Cassius Dio there was seen the operation of a psycholinguistic or sociolinguistic cause for word-order transference: Dio's conscious or unconscious presentation of himself as a Roman. The data sets used by Frischer et al. list for each text sample the number of direct objects that occur before the governing verb in main clauses (MCB) and other clauses (OB) as well as the number of direct objects that occur after the governing verb in main clauses (MCA) and other clauses (OA). In most cases the total number of direct objects is 100. There are a few samples where only 99 sentences were examined, the data have been rescaled to have a total of 100. Frischer et al. (1999) treat the number of clauses in each of the categories as separate, independent random variables. Quantitative linguists and stylometricians may not be aware of a well-known problem that affects the "standard" statistical analysis of compositional data, that is, when dealing with data that adds up to a known total. This presentation aims to explain the problem, present one approach that succeeds in solving it; and apply this approach to such statistical techniques such as principal components analysis, cluster analysis and discriminant analysis. The conclusions of Frischer et al. (1999) will be re-examined.

Compositional Data Analysis

The statistical analysis detailed in the previous section produces easily interpretable results; separate clusters of Latin and Greek authors. However, the authors ignore a constraint on the data, that the number of clauses examined is always one hundred. This constraint has far-reaching implications, Aitchison (1986) shows that that interpretation of the covariance matrix of data with a constrained sum is fraught with problems. Thus any analysis of the data which involves the covariance matrix is also suspect. This includes principal components analysis, discriminant analysis and certain versions of cluster analysis. We shall use compositional data analysis analogues of these techniques to investigate these data. Logcontrast principal components analysis and logcontrast cluster analysis broadly confirm the results from the crude analysis and add refinements; the Graecian nature of Varro's word-ordering and the highlighting of genre differences with Tacitus appear to merit further investigation. In addition, some Greek texts of Plutarch and Marcus Aurelius appear more Latinate than had been previously noticed. Discriminant analysis results in these Greek texts being classed as Latinate, while Tacitus' Agricola is classified as Graecian. In addition, the texts by Cassius Dio and the Greek version of the Res Gestae are classed as Latinate. Further work is currently being undertaken on other works by Varro which appears to indicate that his word-order straddles the Greek-Latin boundary. This will be reported in detail at the conference.

References:

J. Aitchison. The Statistical Analysis of Compositional Data. Monographs in Statistics and Applied Probability. London: Chapman and Hall, 1986.

B. D. Frischer R. Andersen S. Burnstein J. Crawford H. Dik R. Gallucci A. Gowing D. Guthrie M. Haslam D. I. Holmes V. Rudich R. K. Sherk A. Taylor F. J. Tweedie B. Vine. “Word-order transference between Latin and Greek: The relative position of the accusative direct object and the governing verb in Cassius Dio and other Greek and Roman prose authors.” Harvard Studies in Classical Philology. 1999. : .