get me a stove indefinite theme get me the stove less probable gets you a four-course meal indefinite theme plays a card for her non-local beneficiary plays her a card less probable fixes tea for the older boys and himself plays us local beneficiary stran

72 The result in the table of variables in the equation shows that the feature pronominality of theme possesses negative coefficient B of -1.340. It means that the non-default definite theme favors the default benefactive PP construction. The size of the effect of the feature is explainable through the odds ratio expB of 0.262. It suggests that definite theme is 0.262 times to take non-default double object construction or around 4 times likely to take benefactive PP construction. Given 100 occurrences of definite theme, 26 will likely take double object construction, while 74 instances favor benefactive PP construction. The 95 confidence interval CI supports the claim, showing that definite theme tends to choose double object construction between 0.105 to 0.656 times. It means that definite theme tends to take benefactive PP construction between almost 2 times and more than 9 times. Equally, if the theme is indefinite, the sentence tendency to appear in the double object construction is bigger than if the theme is definite. The coefficient B, odds ratio expB, and 95 CI above are presented in Table 4.3 See also Appendix 3. The examples below illustrate the probability of the occurrence of instances possessing indefinite and definite theme. 13

a. fixed the dinner definite theme for her family

more probable

b. fixed her family the dinner less probable

14

a. make the frames definite for Dodge Ram pickup more probable b. make Dodge Ram pickup the frames

less probable 15

a. get me a stove indefinite theme

more probable

b. get me the stove less probable

16

a. gets you a four-course meal indefinite theme

more probable

b. gets you the four-course meal less probable

73 This result of the effect of definiteness of theme confirms the finding of the earlier studies on alternations. As definite theme favors benefactive PP construction, it simply agrees the claim that definite argument precedes indefinite one in alternations Bock Irwin 1980; Thompson 1990; Bock et al 1992; Hawkins 1994; Collins 1995; Arnold et al. 2000; Snyder 2003; Gries 2003 in Bresnan 2007 The benefactive PP puts definite argument, which is the theme, in the postverbal position preceding the beneficiary whether it is definite or not. Whereas beneficiary and theme should be in complement Bresnan 2007, the result suggests that the definiteness of beneficiary does not significantly affect the choice of benefactive construction. The p value of the feature definiteness of theme of 0.237 is insignificant. Its 95 confidence interval CI is too way spread. It has lower limit 0.014 and the upper limit of 2.856. The coefficient is negative of -1.598 which cannot be said to be in balance with the definiteness of theme. Ideally, when the definiteness of theme possesses negative coefficient, the definiteness of beneficiary in contrary should have positive coefficient. This way, the idea of definite argument preceding indefinite argument will stay correct. However, this phenomenon is explainable since the feature of definiteness of beneficiary is insignificant to the choice of benefactive construction. When tested with one on one logistic regression toward benefactive construction, the feature definiteness of theme has relatively weak effect and doubtfully accurate. The classification table of the feature definiteness of theme toward the choice of benefactive alternation show 68,5 of the model accuracy. 74 Ignoring the other features, the feature definiteness of theme along with the intercept are able to predict the occurrence of the benefactive PP or double object construction with 68,5 accuracy. In addition, the omnibus and model summary tables suggest that the single feature definiteness of theme is quite weak to determine the choice of benefactive alternation. The evidence shows that the single feature of definiteness of theme reduces the -2 log likehood of 84.842. The model with intercept has -2 log likehood of 535.883, while after adding the variable of definiteness of theme the -2 log likehood decreases to 451.040. Applying the variable of definiteness of theme in the model will be able to explain the choice of benefactive construction as Nagelkerke R square suggests of 0.259. The very small number of Nagelkerke R square tell us that when appears alone, definiteness of theme is scarcely able to explain the choice of benefactive construction. The model fit accuracy, -2 log likehood, and Nagelkerke R square are presented in Appendix 3. The coefficient and the odds ratio of the feature definiteness of theme is stable at around -2 and 0.2 to -2.223 and 0.108 from the ones when it was applied with the other thirteen features. Variables in the Equation of Theme Definiteness B S.E. Wald df Sig. ExpB 95 C.I.for EXPB Lower Upper Step 1 a themedefiniteness -2.223 .277 64.512 1 .000 .108 .063 .186 Constant 4.191 .506 68.465 1 .000 66.079 a. Variables entered on step 1: themedefiniteness. Table 4.11 Variable definiteness of theme removed from the full model 75

4.1.5 Person of Beneficiary

In the model, the nominal value of person of beneficiary was coded to categorical variable beneficiary=local is 1, and beneficiary=non-local is 2 See Appendices 1 and 2 for the detailed identification and annotating process of the benefactive data. The default of the feature is non-local, whereas the default of the benefactive construction is benefactive PP. The p value of the feature person of beneficiary reaches the number 0.007 See Table 4.3 and Appendix 3 which is less than the conventional significance level 0.05. Thus, the feature person of beneficiary is proven to be relevant to the choice of benefactive construction. The crosstabulation table of person of beneficiary toward ditransitivity shows that 74.8 of non-local beneficiary take benefactive PP construction, only the rest 25.2 of non-local beneficiary take double object construction. Conversely, 80.4 of local beneficiary take double object construction, and the rest 19.6 take benefactive PP construction. Both statistics say that local argument tends to come first. 76 Person of Beneficiary toward Ditransitivity Crosstabulation Ditransitivity Total benefactive constructio prepositiona l Person of Beneficiary local beneficiary Count 82 20 102 within Person of Beneficiary 80.4 19.6 100.0 within Ditransitivity 52.2 8.2 25.5 of Total 20.5 5.0 25.5 non-local beneficiary Count 75 223 298 within Person of Beneficiary 25.2 74.8 100.0 within Ditransitivity 47.8 91.8 74.5 of Total 18.8 55.8 74.5 Total Count 157 243 400 within Person of Beneficiary 39.3 60.8 100.0 within Ditransitivity 100.0 100.0 100.0 of Total 39.3 60.8 100.0 Table 4.12 Crosstabulation of person of beneficiary toward ditransitivity The result shows that the feature givenness of beneficiary possesses positive coefficient B of 1.090. It means that the default non-local beneficiary favors the default benefactive PP construction. The size of the effect of the feature is explainable through the odds ratio expB of 2.974. It suggests that non-local beneficiary is almost 3 times likely to take benefactive PP construction. The 95 confidence interval CI supports the claim, showing that non-local beneficiary tends to choose benefactive PP construction between 1.340 to 6.597 times. Conversely, if the beneficiary is local, the sentence will tend to appear in the double object construction. The coefficient B, odds ratio expB, and 95 CI above are presented in Table 4.3 See also Appendix 3. The examples below 77 illustrate the tendency of instances with non-local and local beneficiary toward benefactive construction. 17

a. plays a card for her non-local beneficiary

more probable

b. plays her a card less probable

18

a. fixes tea for the older boys and himself

more probable non-local beneficiary b. fixes the older boys and himself tea less probable 19

a. plays us local beneficiary strange trick

more probable

b. gets a decent wage for them less probable

20

a. fix yourself local beneficiary something to eat

more probable

b. fix something to eat yourself less probable

This result of the effect of person of beneficiary confirms the finding of the earlier study on alternations. As non-local beneficiary favors benefactive PP construction, it simply agrees the claim made by Bresnan et al. 2001 saying that in dative alternations, non-local recipient favor prepositional dative construction. It puts the non-local recipientbeneficiary at the very end of the sentence. This feature person of beneficiary does not have a comparison of feature dealing with theme. The feature person of beneficiary stands alone and there is not any other feature namely person of theme. It is simply because during the research, the theme which is person is not found. Thus, theme cannot be categorized into local or non-local. While the feature person of beneficiary combines with the intercept only, it has stronger effect toward the choice of benefactive construction. When the 78 feature of person of beneficiary is tested in one on one binary logistic model, the classification table of the feature person of beneficiary toward the choice of benefactive alternation shows 76.3 of the model accuracy. Ignoring the other features, the feature person of beneficiary and the intercept are able to predict the occurrence of the benefactive PP or double object construction with 76.3 accuracy. In addition, the omnibus and model summary tables support relevance of person of beneficiary by giving the evidence that this single feature reduces the -2 log likehood of 98,673. The model with intercept only owns -2 log likehood of 535.883, while after adding the variable of person of beneficiary the -2 log likehood decreases to 437.209. Applying the variable of person of beneficiary in the model will be able to explain the choice of benefactive construction as Nagelkerke R square suggests of 0.296. Only 29.6 of the cases of benefactive alternation occurrences are explainable through intercept and the feature person of beneficiary. The model fit accuracy, -2 log likehood, and Nagelkerke R square are presented in Appendix 3. The coefficient and the odds ratio of the feature givenness of beneficiary increase noticeably to 2.501 and 12.191 from the ones when it was applied with the other thirteen features. Variables in the Equation of Person of Beneficiary B S.E. Wald df Sig. ExpB 95 C.I.for EXPB Lower Upper Step 1 a personofbeneficiary 2.501 .283 78.154 1 .000 12.191 7.002 21.223 Constant -3.912 .516 57.394 1 .000 .020 a. Variables entered on step 1: personofbeneficiary. Table 4.13 Variable person of beneficiary removed from full model 79

4.1.6 Syntactic Complexity

The last important significant feature toward the choice of benefactive alternation is syntactic complexity or in another word length different . In the model, syntactic complexity is scale valued, thus there is no categorical values for this feature See Appendices 1 and 2 for the detailed identification and annotating process of the benefactive data. There is no default of syntactic complexity, whereas the default of the benefactive construction is benefactive PP. However, the feature of syntactic complexity can be seen as log scale obtained by substracting the beneficiary length with the theme length. When the beneficiary is longer than theme, positive log scale will be achieved. In contrary, when the theme is longer, the log scale results in negative. The p value of the feature syntactic complexity reaches the number 0.000 See Table 4.3 and Appendix 3 which is less than the conventional significance level 0.05. Thus, the feature syntactic complexity is proven to be relevant to the choice of benefactive construction. The crosstabulation table of syntactic complexity toward ditransitivity shows stable effects. All of the negative log-scaled instances ranging from -6 to - 12 take double object construction. 75 of -5 log-scaled instances take double object construction, and the rest 25 take benefactive PP construction. 80 of -4 log-scaled instances take double object construction and the rest 20 take benefactive PP construction. 73.7 of -3 log-scaled instances take double object construction, while the rest 26.3 take benefactive PP. 68,9 of -2 log-scaled instances take double object construction, while the rest 31.1 take benefactive PP construction. 58,4 of -1 log-scaled instances take double object construction, 80 whereas the rest 41.6 take benefactive PP. Interestingly, 81.1 of 0 log-scaled instances take benefactive PP construction and the rest 18.9 take double object construction. 98.2 of +1 log-scaled instances take benefactive PP, and the rest 1.8 take double object construction. All of instances with +2 log scale until +7 log scale take benefactive PP construction. Statistically, positive and zero log- scaled instances favor benefactive PP construction, while negative log-scaled instances favor double object construction. See the figure below to see the detailed information of occurrences. Figure 4.2 Tabular data showing distribution of syntactic complexity in benefactive alternation The result also shows that the feature syntactic complexity possesses positive coefficient B of 0.597. It means that when the log scale is positive, it favors the default benefactive PP construction. On the other hand, when the log 81 scale is negative, the construction seems to appear in double object construction. The size of the effect of the feature is explainable through the odds ratio expB of 1.817. It suggests that positive log-scaled instance is almost 2 times likely to take benefactive PP construction. The 95 confidence interval CI supports the claim, showing positive log-scaled instance tends to choose benefactive PP construction between 1.319 to 2.501 times. Conversely, if the log scale is negative, the sentence will tend to appear in the double object construction. The coefficient B, odds ratio expB, and 95 CI above are presented in Table 4.3 See also Appendix 3. The examples below illustrate the probability of the choice of benefactive construction based on the length different between beneficiary and theme. 21 a. make room for that rosy-cheeked girl. more probable b. make that rosy-cheeked girl room less probable theme: room 1 beneficiary: that rosy-cheeked girl 1 2 3 4 log scale= beneficiary length - theme length = 4 – 1 = 3 positive 22

a. fixing lunch for her five-year-old son more probable

b. fixing her five-year-old son lunch less probable theme: lunch 1 82 beneficiary: her five-year-old son 1 2 3 4 5 log scale= beneficiary length - theme length = 5 – 1 = 4 positive 23 a. played us no such ugly trick more probable b. played no such ugly trick for us less probable theme: no such ugly trick 1 2 3 4 beneficiary: us 1 log scale= beneficiary length - theme length = 1 - 4 = -3 negative 24 a. make himself an oil lamp more probable b. make an oil lamp for himself less probable theme: an oil lamp 1 2 3 beneficiary: himself 1 log scale= beneficiary length - theme length = 1 - 3 = -2 negative 83 This result of the effect of syntactic complexity confirms the finding of the earlier studies on clause word order and on the dative alternation. Quirk et al. 1972 mentioned what is so called principle of end weight, in which language users are inclined to place the more complex elements at the end of the utterance. Additionally, the result of this study verifies the claim of the recent study on predicting word order that relative syntactic complexity is vital Hawkins 1994; Arnold et al. in Bresnan 2010. While the feature syntactic complexity combines with the intercept only, it has a little bit stronger effect toward the choice of benefactive construction. When the feature of syntactic complexity is tested in one on one binary logistic model, the classification table of the feature syntactic complexity toward the choice of benefactive alternation shows 73.5 of the model accuracy. Ignoring the other features, the feature syntactic complexity and the intercept are able to predict the occurrence of the benefactive PP or double object construction with 73.5 accuracy. In addition, the omnibus and model summary tables support relevance of syntactic complexity by giving the evidence that this single feature reduces the -2 log likehood of 152.497. The model with intercept only owns -2 log likehood of 535.883, while after adding the variable of syntactic complexity the -2 log likehood decreases to 383.385. Applying the variable of syntactic complexity in the model will be able to explain the choice of benefactive construction as Nagelkerke R square suggests of 0.429, 42.9 of the cases of benefactive alternation occurrences are explainable through intercept and the feature syntactic complexity. The model fit accuracy, -2 log likehood, and Nagelkerke R square are presented in Appendix 3. The coefficient and the odds ratio of the feature 84 givenness of beneficiary increase slightly to 1.003 and 2.726 from the ones when it was applied with the other thirteen features. Variables in the Equation of Syntactic Complexity B S.E. Wald df Sig. ExpB 95 C.I.for EXPB Lower Upper Step 1 a syntacticcomplexity 1.003 .114 77.110 1 .000 2.726 2.179 3.410 Constant 1.229 .159 59.846 1 .000 3.417 a. Variables entered on step 1: syntacticcomplexity. Table 4.14 Variable syntactic complexity removed from full model To conclude with, adding the six significant variables has reduced the -2 log likehood by 328,175 with 6 degree of freedom. The -2 log likehood is a measure of how well and accurate the model explains variations in the outcome of interest. Thus, the table of effect of the significant features to the choice of benefactive construction below can explain the size of each feature to increase the accuracy of the model. The model accuracy or goodness-of-it will decrease in a certain extent when a feature is omitted. The figure below shows the features‟ significance by removing each feature from the full model of probabilistic binary logistic regression model. 85 Figure 4.3 Effect sizes of the significant features to the choice of benefactive construction Some features possess strong significance level toward the choice of benefactive alternation, yet they need to be in the equation and bounded to each other to reach the possible highest accurate model. Partially, givenness of beneficiary has 83.8 of model accuracy, animacy of beneficiary has 63.5 model accuracy, pronominality of theme has 60.8 model accuracy, definiteness of theme has 68,5 model accuracy, person of beneficiary has 76.3 model accuracy , and syntactic complexity has 73.5 model accuracy. When the six significant features are in the equation and bounded to each other they are able to reach 90 model accuracy. The model accuracy declines 1 from the model when all features are included. Eight features appear not to be powerful enough to significantly affect the choice of benefactive construction. However, the eight features hold 1 relevance to the choice of benefactive construction. The logistic 86 model predicting accuracy decreases 1, from 91 to 90 when the eight insignificant features are omitted See also Appendix 3. The P-P plots of the two models are shown below. We can see that the curves are similar to each other. Figure 4.4 Model plots of observed against estimated responses. Model A employs fourteen features, Model B employs six features The formula of probability corpus derived from logistic model below omits the eight insignificant features, making used only six significant features with p value less than 0.05. The formula is obtained from multivariable binary logistic regression model. 87 Variables in the Equation B S.E. Wald df Sig. ExpB 95 C.I.for EXPB Lower Upper Step 1 a syntacticcomplexity .573 .154 13.844 1 .000 1.774 1.312 2.400 beneficiaryanimacy 2.249 1.019 4.869 1 .027 9.480 1.286 69.903 beneficiarygivenness 3.563 .649 30.114 1 .000 35.283 9.882 125.980 themepronominality -1.909 .773 6.092 1 .014 .148 .033 .675 personofbeneficiary 1.090 .387 7.916 1 .005 2.975 1.392 6.358 themedefiniteness -1.369 .445 9.486 1 .002 .254 .106 .608 Constant -1.766 1.797 .966 1 .326 .171 a. Variables entered on step 1: syntacticcomplexity, beneficiaryanimacy, beneficiarygivenness, themepronominality, personofbeneficiary, themedefiniteness. Table 4.15 Table of six variables in the equation The formula is as follow: constant+β1x1+β2x2+β3x3+… +βnxn e In p constant+ β1x1+β2x2+β3x3+… +βnxn 1+e expconstant+ β1x1+β2x2+β3x3+… +βnxn 1+ expconstant+ β1x1+β2x2+β3x3+… +βnxn In p = constant+ β1x1+β2x2+β3x3+… +βnxn 1 - p = -1.766+0.573syntacticcomplexity a+2.249beneficiaryanimacyb+3.563benefici arygivennessc- 1.909themepronominalityd+1.090personofben eficiarye-1.369theme definitenessf 88 The formula works as follow: 25 a. I call a woman a christian, that works from morning till night, to get bread for her family, and is never heard to scold or mutter. b. I call a woman a christian, that works from morning till night, to get her family bread, and is never heard to scold or mutter. verb : get theme : bread : non-pronound; indefinitef 1 beneficiary : her family : animateb; non-givenc; non-locale 1 2 log scale= 2-1=+1a constant+βx1+βx2+βx3+… +βxn e In p constant+βx1+βx2+βx3+… +βxn 1+e expconstant+βx1+βx2+βx3+… +βxn 1+ expconstant+βx1+βx2+βx3+… +βxn -1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 1-1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 89 -1.766+0.5731+2.2491+3.5632 -1.9092+1.0902-1.3692 1-1.766+0.5731+2.2491+3.5632 -1.9092+1.0902-1.3692 = 0.97825 The result shows the tendency of the instance to take benefactive PP construction. The number shows that it is more than 0.5 to take benefactive PP, so the model predicts that the occurrence is in benefactive PP construction. In this instance, the predicted construction matches the observed construction. In the observed data, the following instance appears in the double object construction. The model will try to predict the construction using the six significant features. 26 a. of beer with their father, who had been an honest Dutch cooper, and got himself a comfortable fortune, which was all his children seemed to have inherited from b. of beer with their father, who had been an honest Dutch cooper, and got a comfortable fortune for himself, which was all his children seemed to have inherited from verb : got theme : a comfortable fortune : non-pronound; indefinitef 1 2 3 beneficiary : himself : animateb; givenc; non-locale 1 90 log scale= 1-3= -2 a constant+βx1+βx2+βx3+… +βxn e In p constant+βx1+βx2+βx3+… +βxn 1+e expconstant+βx1+βx2+βx3+… +βxn 1+ expconstant+βx1+βx2+βx3+… +βxn -1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 1-1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f -1.766+0.573-2+2.2491+3.5631 -1,9092+1,0902-1,3692 1-1.766+0.573-2+2.2491+3.5631 -1.9092+1.0902-1.3692 = 0.18603 The result shows that the tendency of the instance to take benefactive PP construction is 0.18603 which is less than 0.5. Thus, the model predicts that the occurrence is in double object construction. Once again, the predicted construction matches the observed construction. The figure below shows the process of predicting the expected PP realization. In this model, the column „probability to take PP‟ shows the percentage of the instance tendency to take benefactive PP construction. When the probability is more than 0.5, the instance tends to take benefactive PP 91 construction. Conversely, when the value is less than 0.5, the instance tends to take double object construction. The column expected is the nominal values predicted PP and DO which are converted into numbers. The column observed is taken from the observed construction of ditransitivity. When the expected and observed matches, the accuracy of prediction is written accurate. It works vice versa. Figure 4.5 Excel processing of PP realization prediction Compared to the previous model, in which, all fourteen features are included, this model has higher -2 log likehood, slightly less chi-square statistic, and slightly less Nagelkerke R square. The model with fourteen features has -2 log likehood of 193.177 while this model with six features has -2 log likehood of 207.708. The chi-square statistic of the model with fourteen features is 342.706, whereas this model has chi-square statistic of 328,175. The model with fourteen features has Nagelkerke R square of 0.780, while this model has Nagelkerke R 92 square of 0.758. It tells us that the new model, indeed, slightly less accurate than the previous one. Yet, the declining of the accuracy is hardly noticeable since it only helps to correctly predict an instance more out of 100 occurrences. With the new model which considers only six features and an intercept, the new model is able to accurately predict 90 out of 100 instances. The model is tested in some ways and proven to be valid. To test the internal validity of the formula and to find the accuracy of the model to the unseen data, the ten-fold validation is used. The researcher breaks the data set into 10 folds of instances, and then new model is made. The model is computed from 90 of the total data. The computed data consists of the second until the tenth folds. The corpus probability formula which is obtained then is used to test the rest 10 instances, which is the fold number one. The next model is made from the folds one, three until tenth. The formula is tested on fold number two. The test works this way until ten models out of the 10 folds were made. The accuracy is derived from the mean of accuracy of the 10 models. This test will show how well the model can predict the unseen data within the internal data set. The result shows that it still has relatively high percentage of overall model accuracy 89.5 compared to the majority baseline of 70.1. The first model reaches 82.5 accuracy when it is tested to first fold. Then the following models consecutively show the accuracy of 87.5, 85, 92.5, 95, 85, 87.5, 95, 90, and 95. The mean of the accuracy of those ten models is 89.5. See the figure below see also appendix 5: 93 Figure 4.6 The percentage of the ten-fold cross-validation accuracy In addition, to test the external validity of the model, the model is used to predict the occurrences of instances in other data set. The model is applied in 40 instances which are taken from Time Magazine Corpus and is able to correctly predict 34 out of 40. It shows that the external validity of the model is still relatively high, which is 85, when applied to other similar situations in real world, in this case in Time Magazine Corpus situationsee appendix 6.Two examples are brought up from the Time Magazine Corpus data set. 27 a. do an he says he wants to study bugs-dats nuts-and Ive just got him a job at Central Islip. Dats because Im in de Marshall b. do an he says he wants to study bugs-dats nuts-and Ive just got a job for him at Central Islip. Dats because Im in de Marshall 94 verb : got theme : a job : non-pronound; indefinitef 1 2 beneficiary : him : animateb; givenc; non-locale 1 log scale= 2-1= -1 a constant+βx1+βx2+βx3+… +βxn e In p constant+βx1+βx2+βx3+… +βxn 1+e expconstant+βx1+βx2+βx3+… +βxn 1+ expconstant+βx1+βx2+βx3+… +βxn -1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 1-1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f -1.766+0.573-1+2.2491+3.5631 -1.9092+1.0902-1.3692 1-1.766+0.573-1+2.2491+3.5631 -1.9092+1.0902-1.3692 = 0.28843 The result shows the tendency of the instance to take benefactive PP construction is 0.28843. The number shows that it is less than 0.5 to take 95 benefactive PP, so the model predicts that the occurrence is in double object construction. In this instance, the predicted construction matches the observed construction. In this very case, the model is externally valid. 28 a. week before the committee. He must explain, among other things, where he got the money for his stock market operations. b. week before the committee. He must explain, among other things, where he got his stock market operations the money. verb : got theme : the money : non-pronound; definitef 1 2 beneficiary: his stock market operations : inanimateb;non-givenc;non-locale 1 2 3 4 log scale= 4-2= +2 a constant+β1x1+β2x2+β3x3+… +βnxn e In p constant+β1x1+β2x2+β3x3+… +βnxn 1+e expconstant+β1x1+β2x2+β3x3+… +βnxn 1+ expconstant+β1x1+β2x2+β3x3+… +βnxn -1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 1-1.766+0.573a+2.249b+3.563c -1.909d+1.090e-1.369f 96 -1.766+0.5732+2.2492+3.5632 -1.9092+1.0902-1.3691 1-1.766+0.5732+2.2492+3.5632 -1.9092+1.0902-1.3691 = 0.99966 The result shows the tendency of the instance to take benefactive PP construction is 0.99966. The number shows that it is more than 0.5 to take benefactive PP, so the model predicts that the occurrence is in benefactive PP construction. In this instance, the predicted construction matches the observed construction. Once again, the model predicts accurately and is proven to be externally valid.

4.2 Features Relevant on Dative and Benefactive