Data Analysis RESEARCH METHODOLOGY

44 b. get = get in preparing something : getting the dinner ready for the kids. c. get = get in abstract sense : get that a lot. After re-checking, omitting, and re-selecting a hundred instances from each verb representing a semantic class, four hundred instances in total were obtained. Those four hundred instances become the data set for this research.The data set of the benefactive constructions is presented in Appendix 1.

3.3 Data Analysis

In the process of data analysis, first the researcher annotated the 400 instances of the data set with fourteen predicting features. The nominal features were coded into binary variables, for example beneficiary = animate is 1 and beneficiary = inanimate is 2. The same was done to the construction used: if the construction is double object construction, it received value 1, if it was benefactive PP, the value was 2. Separately, the ordinal feature with a continuous scale namely syntactic complexity is treated as the way it is, which is ordinal. The log scale is obtained by subtracting the theme length from beneficiary length. See Appendix 2 for the coded data set. The second step, the researcher processed the annotated data of benefactive instances in SPSS. The binary and continuous scaled variables were input into SPSS. The data was then analyzed using mixed-effect binary logistic model and simple binary logistic model. The regression models found can be employed to first, establish the features significant to the choice of benefactive and second, find the preference of the instances carrying the features. Beside p value significance, standard error SE, and odds ratio expB, the result of 45 mixed-effect binary logistic model provides intercept constant, and the coefficient of each feature which can be developed into the probabilistic model formula. The regression models were carried out using SPSS.17 program. To do the mixed-effect binary regression analysis, in the .sav file of SPSS containing all the variables ditransitivity and the fourteen linguistic features click Analyze Regression Binary Logistic then put the ditransitivity of benefactive construction into the dependent variable string and the fourteen linguistic features into the covariates string. Then, click save put a check mark on probabilities continue. The following step was clicking option put a check mark on CI for expB: 95 continue OK. The results of the computation process were presented in.spv file, the output of SPSS statistics viewer. The output provides the statistical data of the model accuracy, chi-square distribution, -2 log likehood, along with the coefficient, standard error, p-value, expB, and CI 95 of each feature. The process of simple binary logistic regression analysis is similar to the steps of mixed-effect binary logistic regression above, yet in the string covariates, each single feature was put one after another for every analysis. The probabilistic regression formula used in this model is the logit model. The formula utilized the intercept constant and the coefficients of the linguistic features. The model formula is written as follow, in which p represents probability of the occurrence of PP construction, β represents the coefficient of each feature. constant+ β1 x1+ β2 x2+ β3 x3+… + βn xn e In p constant+ β1 x1+ β2 x2+ β3 x3+… + βn xn 1+e 46 expconstant+ β1 x1+ β2 x2+ β3 x3+… + βn xn 1+ expconstant+ β1 x1+ β2 x2+ β3 x3+… + βn xn In p = constant+ β1 x1+ β2 x2+ β3 x3+… + βn xn 1 - p When the constant and coefficients were obtained from the mixed-effect binary logistic, the formula above can be completed and was able to predict the occurrence of benefactive construction as presented in Appendix 4. The probabilistic formula gives us the percentage of certain instance to take benefactive PP or double object construction. In addition, from the computed data in the regression models, the research could obtain the model fit accuracy of the corpus probabilistic formula of benefactive construction. See Appendix 3 The third step, the researcher made a new model of corpus probabilistic benefactive formula with exactly the same coding system with the one done by Bresnan. The model coded all the features to binary data except for syntactic complexity. The values of the nominal variables which were coded with binary system with number 1 and 2, were then re-coded so that they reflected the presence of one of the feature value, e.g. themebeneficary=animate is 1, or the absence of the same feature value e.g. themebeneficiary=animate is 0. The only ordinal variable, which is syntactic complexity, is kept the way it was. The next process of data processing in SPSS and formulating the logit probabilistic benefactive formula was the same with the second step of data processing above. The result of the mix-effects binary regression model is presented in Appendix 7. 47 The last step, the researcher compared the direction of significance and the size of effect of the features toward benefactive and dative construction. The the significant features for benefactive were put side by side to the significant features of dative. The six shared relevant features were then deeply analyzed based on the coefficient and odds ratio. The coefficient shows the direction of the effect and the odds ratio tells us the size of the effect. Furthermore, the six significant features were taken out from the full model and were put into the partial probabilistic models. Thus, for benefactive cases, the research possessed both the data set and the probabilistic formula, and for dative cases, the research only owned the probabilistic formula. As the researcher did not have the access to the dative data set of Bresnan, a new dative data set was built from TIME Magazine corpus. Doing this, the researcher acquired both benefactive data set and benefactive probabilistic formula along with dative data set and dative probabilistic formula. The logit probabilistic formulas were then cross-applied to the data. The probabilistic formula of dative was tested on benefactive data set, so was the benefactive probabilistic formula tested on dative data set. This method would be able to provide the data of the interchangeability of the dative and benefactive. See Appendix 8 for the complete data set, coding process and predicting DO realization process.

3.4 Triangulation