Preposition Stranding in English: Predicting Speakers’ Behaviour

meaning of the preposition 0.014 length of the extracted phrase 0.036 spatial meaning of the preposition 0.04 agentivity of the head noun 0.104 phrasal-prepositional verbs 0.114 frequency of the preposition 0.115 barrierhood of the extracted phrase 0.119 prepositional verbs 0.126 concreteness of the head noun 0.132 copula as verb 0.153 entrenchment of the head noun 0.165 intransitive verbs 0.165 length of the preposition 0.218 according to the low factor loadings (-0.223 ≤ loading ≤ 0.223), 11 these variables do not discriminate significantly between the two constructions modality 0.382 high/low value ⇒ SC/PPC Table 5: Factor Loadings of the Discriminant Analysis It is obvious that, of all variables investigated, the bridging structure, the verb and the modality influence PS most strongly. The hypothesis of the influence of processing effort on the choice of construction seems to be borne out since the length and the barrierhood of the bridging structure relate straightforwardly (along the lines discussed in section 1.1) to the morphosyntactic and semantic processing effort respectively necessary for the production of the utterance. As to the influence of transitive verbs on PS, one might wonder whether this finding supports the role of processing put forth, but there is an obvious explanation for that, too: as opposed to all other kinds of verbs investigated here, transitive verbs require a direct object, i.e. at least an additional NP. This NP will obligatorily add to the length and the barrierhood of the bridging structure as in, say, To whom did John give [NP the book]? or Who did John give [NP the book] to? and thereby yield a preference for the PPC. A look at our data supports this hypothesis; consider Table 6. Transitive (111 sentences) Not transitive (190 sentences) Total LENGTH_BS: Mean (Std. dev.) 10.9 (7.7) 6.5 (6.4) 8.1 (7.2) BARRIER_BS: Mean (Std. dev.) 4 (2.9) 2.5 (2.7) 3 (2.9) Table 6: The Effect of Transitivity on LENGTH_BS and BARRIER_BS The average length and barrierhood of the bridging structure is much higher for transitive verbs than for non-transitive verbs; the differences are, according to Welch’s t test, highly significant and the influence of transitive verbs can, thus, be explained in terms of processing effort. The effect of verb voice on PS is more difficult to relate to processing cost: when the main verb is in the passive, we find SC significantly less than expected. At this preliminary stage, I can only suggest somewhat tentatively that the noncanonical passive is more difficult to process than the canonical active so that both passive and SC is avoided by speakers. Admittedly, compared to the other more solid arguments, this is fairly vague and requires further investigation. The strong influence of the modality, however, is most probably not due to a causal influence on processing – rather, it is more likely due to writers’ prescriptive knowledge/awareness (never use a preposition to end a sentence with!). 4. Summary / Conclusions We have seen how the analysis of syntactic variation can benefit from the use of rigorous corpus-based and (multifactorial) statistical investigation. While such techniques to analysing variation data were quite common in the 70s (cf. the notion of variable rules employed by Cedergren, Labov, Sankoff and others), nowadays the analysis of variation does not (at least to my mind) utilise the power of these techniques frequently enough. This is all the more surprising since even introductory textbooks (!) to corpus linguistics as well as other publications have argued time and again that monofactorial studies often do not suffice: [...] straightforward significance or association tests, although important, cannot always handle the full complexity of the data. The multivariate approaches [...] offer a way of looking at large numbers of interrelated variables and discovering or confirming broader patterns within those variables. (McEnery and Wilson 1997:82) Although linguists ... typically do not use statistical techniques, the approach just illustrated fits conceptually with correlational models using multiple regression analyses ... [i.e.,] with a more complex design we can obtain information that is not readily available by armchair analysis. (Bates and McWhinney 1982:181) In this respect, I would thus argue that, methodologically at least, there is a great deal that we as linguists can learn from other behavioural sciences as far as data collection, hypothesis testing and exploratory statistical techniques are concerned. I would also hope that a shift to more rigorous testing of the sort detailed above would render linguistic findings more objective and reliable than has been the case in the preceding 40 years of predominantly intuitive/introspective analyses of acceptability/grammaticality judgements (cf. Schütze 1996 for a similar line of reasoning, though not in the direction of multifactorial corpus analyses). In the case at hand, the most crucial determinants of PS seem to be the processing effort associated with the two word orders and the knowledge of prescriptive grammar rules. On a more general note, the findings concerning processing effort lend themselves to being integrated into psycholinguistic theories based on interactive activation networks such as Bates and MacWhinney’s (1982, 1989) Competition Model, where variables with different constructional preferences compete with each other: the notion of interaction as dealt with in section 3.2 operationalises the notion of conflict validity, the prior probabilities of the two constructions in the LDA/CART analyses correspond to resting levels / baseline activations, and the variables’ weightings could readily be interpreted as association strengths between variables and the constructional choice. However, further research is necessary to integrate more of the previous findings into psycholinguistic theory.