Classification of the Causees
The same models were tested on the Causee nouns. Neither of them showed much predictive power, as displayed in Figure 3, although the predictive power slowly grows with the number of clusters. This is not surprising: the higher the granularity, the better the individual observations are fitted, but at the cost of parsimony. This poor predictive power (maximum C = 0.74) corroborates the previous studies, according to which the inherent semantic classes of the Causee are largely irrelevant, in contrast with the thematic role of the participant in bringing about the effected event.
Figure 3: The Causee: predictive power of two models, for different number of classes
Classification of the Effected Predicates
Finally, let us consider the Effected Predicate. Figure 4 shows the predictive power of 16 models. According to the analysis, the best-performing model was 23syn (the upper line), the model with information about the subcategorisation frames based on 23 syntactic relations without any additional information, although some other models were more successful for a very small number of classes, as one can see from Figure 4. It is interesting that the model 9richsubcat, which was the leader when the number of clusters was very small (C = 0.68 with 5 classes), contained information about the subcategorisation frames enriched with the information about the prepositions and semantic noun classes. As the number of clusters grew, the leadership was taken over by other models.
Figure 4: The Effected Predicate: predictive power of sixteen models, for different number of classes
Namely, the next two leaders (for 10 and 15 clusters) were based on the subcategorisation frames enriched with the information about the prepositions (9relprep with 10 clusters, C = 0.73) and semantic noun classes (23sclass with 15 clusters, C = 0.79). From 20 verb classes on, the simple subcategorisation frames based on 23 dependencies without additional information yielded the best results (23syn with 20 clusters, C = 0.83), although the model was closely followed by 9relprep, which involved subcategorisation frames based on 9 syntactic relationships and prepositions, especially for large numbers of clusters. The bag-of- words models performed on average worse than most other models, although the large window model (BOW15) was slightly more successful than the small window one (BOW4). Another poorly performing model was 9sclass (subcategorisation frames based on only 9 syntactic relations, enriched with the information about noun classes). All this suggests that the abstract constructional information is vital for a successful verb classification, and that this information should be very detailed and go beyond the main arguments. We can also conclude that the semantic and prepositional information does not add much, and even makes the classification less successful, when we have 20 and more classes.
The increase in the predictive power was gradual, so that the optimal classification into 35 clusters (C = 0.88) - the number after which the predictive power does not increase substantially - was medium-grained. Only three of these clusters contain verbs that predominantly co-occurred with doen in the test sample. These classes are interpretable as verbs of qualitative and configurational change of state (e.g. herleven ‘come to life again’, kantelen ‘tip over’, smelten ‘melt’, verslappen ‘weaken’, vervagen ‘fade’), verbs of quantitative change along a scale (e.g. stijgen ‘go up’, dalen ‘go down’, groeien ‘grow’, zakken ‘fall’) and verbs of various mental processes, emotions and beliefs (e.g. denken ‘think’, vermoeden ‘suppose', geloven ‘believe', besluiten ‘decide, conclude', vrezen ‘fear', hopen ‘hope').
Most of the clusters showed a higher proportion of laten, as was the case with the Causer (see section 5.1). One of them was the cluster containing predominantly verbs of communication: schrijven ‘write’, uitleggen ‘explain’, adviseren ‘advise', vertellen ‘tell'. Another cluster contained many verbs of change of possession and possessional deprivation: geven ‘give', ontnemen ‘take, rob', leveren ‘deliver’, verkopen ‘sell’. Yet another one mainly consisted of verbs of searching, active perception and testing, e.g. onderzoeken ‘explore’, bekijken ‘have a look (at)’, toetsen ‘test’. Most of these imply a volitional human Causee and a human Causer, who represents an authority and gives orders (10). The presence of such semantic frames and scenarios of causation implies that the combinations of the three slot fillers are not arbitrary. From this follows that the combined effect of the semantic classes of the Causer, Causee and Effected Predicate on the choice of the auxiliary is probably not additive. We will come back to this observation later.
(10) Obama laat alle kerncentrales VS onderzoeken.
Obama lets all nuclear-stations US check
‘Obama orders to check all nuclear power stations in the US.’
Another cluster contains verbs related to putting and bringing (leggen ‘lay’, zetten ‘set’, stellen ‘put’, brengen ‘bring’), which imply an active Causee that brings about a change in the location or position of another entitity. Yet another one had verbs of motion, such as draaien ‘turn round’, glijden ‘slide’, rijden ‘ride’, vliegen ‘fly’, which also involve a certain degree of autonomy on the part of the Causee.
However, many other classes were more difficult to interpret. For instance, one of the clusters with a very high proportion of laten contained verbs of perception (zien ‘see’, voelen ‘feel’), the verbs weten ‘know’, kennen ‘know, be acquainted’, maken ‘make’, doen ‘do’, blijken ‘appear’, schijnen ‘shine, appear’, zijn ‘be’, worden ‘become’ and hebben ‘have’. It is difficult to say why these verbs go together. Probably this is an effect of the overall high frequency and broad constructional repertoire of these verbs. In this respect, this cluster was similar to another one, which contained the semi-auxiliary or light verbs prototypically related to motion or position: gaan ‘go’, zitten ‘sit’, staan ‘stand’, vallen ‘fall’, komen ‘come'. Note that many of these verbs in the vague clusters are strongly associated with laten (see section 2).
In general, many of these classes strikingly resemble Levin’s (1993) classification. This similarity can be explained by the fact that both Levin’s “alternations” and subcategorisation frames reflect the distribution of verbs in constructions. However, our approach is finer-grained (more possible constructions are examined), probabilistic and treats every construction on its own, not as a part of an alternation pair.