Sentence Completion
7.2 Sentence Completion
7.2.1 Introduction
The sentence completion study was designated to confirm that referential context would significantly influence RC attachment in a two-site ambiguity of the form NP1-NP2- RC. This used the method employed by Desmet et al. (2002), which was itself based upon Zagar et al. (1997). Desmet et al.’s materials were obtained 1 and translated into Engish by a professional translator. This provided a framework for the materials in the present study. They were then edited.
Context paragraphs, which preceded the target sentences, were either (i) neutral, (ii) early closure-biasing or (iii) late closure-biasing. Neutral contexts contained either no referents (twenty-seven instances) or only one possible referent for both NPs in the target sentence (three instances). Early closure-biasing contexts contained two referents for NP1 and one for NP2. Late closure-biasing context contained two referents for NP2 and one for NP1. There were ninety individual context paragraphs.
The design comprised thirty pairs of target sentences. Each target sentence took the form NP1 of NP2 RP VERB PP (e.g. The police interrogate the advisor of the politicians who speaks with a soft voice). Every target sentence was disambiguated at the verb, which agreed in number with either NP1 or NP2. In half the sentences, NP1 was plural and NP2 singular. This was reversed for the other half. The NP1-attaching and NP2-attaching versions of the sentences were identical but for the verb, which was either singular or plural. Verb-agreement disambiguation is useful because it is purely grammatical. For the sentence completion study, the verb and its prepositional phrase (e.g. speaks with a soft voice) were omitted. A participant’s verb choice would indicate their attachment preference. (See Appendix Three for a complete list of materials.)
There is empirical support for NP2 attachment in English RC ambiguities (e.g. Corley, 1995; Cuetos Mitchell, 1988; Mitchell Cuetos, 1991b; Gilboy et al., 1995). With the referential context manipulations here, we should expect the neutral context to influence this little, the early closure-biasing context to reduce or reverse NP2 preference, and the late closure-biasing context to slightly increase NP2 preference.
1 With the kind permission of the first author, Timothy Desmet.
7.2.2 Method
Participants Eighteen (18) participants were recruited from a first-year undergraduate psychology
course. Only native speakers of International English were recruited. Materials
Thirty target sentences were constructed using translated materials from Desmet et al. (2002). These sentences took the form NP1 of NP2 RP, e.g.:
(55) Amnesty International honours the lawyer of the campaigners
who is concerned about individual freedom.
Sentences were preceded by one of three contexts: neutral (no referents for NP1 or NP2, or, rarely, one possible referent for both), early closure-biasing (two possible referents for NP1), late closure-biasing (two possible referents for NP2). Participants saw only one version of each context paragraph (neutral or early or late) according to a Latin-square design. Thus, there were three master questionnaires. Items within each of the three questionnaires were randomly ordered. Here is an example of a neutral context paragraph (which precedes (55)):
(56) There is international concern about the abuse of human rights. In
South America, a taskforce is set up to recognize those people who contribute to upholding human rights. It makes recommendations to well-known agencies.
There were 30 experimental items in a given questionnaire. To distract participants from the experimental manipulation, 40 filler paragraphs were constructed that bore a surface resemblance to the experimental items and were of comparable length. These paragraphs truncated pseudo-randomly during the fifth sentence.
Materials were presented using a paper questionnaire.
Procedure Participants completed the questionnaires in a quiet room. Written instructions were
provided on the first page. Participants were asked to provide ‘[the] first plausible continuation that comes to mind’. To ensure that the verb chosen by the participant would disambiguate the RC, participates were asked ‘to make the continuation similar in style and tense to the original’. An example continuation was provided.
7.2.3 Results
For each participant, a NP1 bias (expressed as a percentage) was calculated. Table 20 shows the NP1 bias for continuations that followed neutral, early closure-biasing and late closure-biasing paragraphs.
Table 20: Mean NP1 bias as a function of context paragraph
Mean Range SD
When the context was neutral, NP1 bias was 34.1. This is consistent with studies that show a Late Closure bias in two-site RC ambiguities (see Mitchell Brysbaert, 1998). Following an early biasing-context, NP1 preference is increased by more than
30. This is in the direction we would expect. It might have been the case that the late- closure context would influence continuation in much the same way as the neutral context, but in fact it it increases the late closure bias by nearly ten per cent, so that 75 of participants favour an NP2 continuation when the target sentence is preceded by such
a context.
Three planned, pairwise T-tests were performed to confirm that the differences between the contexts are reliable. The results are summarized in Table 21.
Table 21: T-tests showing the differential effect of context on NP1 bias
T-test
T df p (two-tailed)
Neutral vs. Early
context Neutral vs. Late
context Early vs. Late
context
Table 21 shows a reliable difference between neutral and early contexts (T 17 = 5.07, p < 0.001). Similarly, the difference between early and late contexts is reliable (T 17 =
6.77, p < 0.001). There is a marginally unreliable difference between neutral and late
contexts (T 17 = 2.05, p = 0.056), though this is not surprising given the tendency
towards Late Closure in English.
7.2.4 Discussion
The sentence completion study has confirmed the expected referential biases of the neutral, early closure-biasing and late closure-biasing contexts. For the sentence completion component of Zagar et al.’s (1997) study, early closure-biasing context produced an 89 NP1 bias (NP1 being the general tendency in the resolution of French two-site RC constructions), whereas their late closure-biasing context reduced this to
85. Though this difference was significant, it is a swing of only 4. For Desmet et al. (2002), the late closure biasing context swung NP1 preference (NP1 also being the general tendency in Dutch) from 83 following early contexts to 39, which is a swing of 44. In the study presented here, where English readers generally prefer Late Attachment, late biasing-contexts produced an NP1 bias of 25 and early biasing- contexts 65, which is a swing of 40. Like Desmet et al., the difference is significant at the five per cent level.
Thus, the materials are sufficiently biasing in terms of their direction and strength. We can now proceed examine whether their effect in an on-line study may vary with reading span.