Since the early 1990s, there has been a debate on the universality of locality in sentence processing (i.e., the preference to associate a word or phrase to the closest possible word). Studies across various languages have investigated ambiguous relative clauses that can be attached to either of two nouns to determine the types of languages in which locality is violated. We report a corpus count and a questionnaire in Thai indicating that intra-sentential contexts can obscure locality. Two reading experiments controlling for context are also reported in support of locality in Thai. The finding that context distorted locality raises the possibility that previous reports of locality violations in various languages may be reduced to contextual effects.
As sentences are read word by word, there is a preference to
attach a new word (or phrase) to the closet possible word
(1) John met the friend of the teacher who was in
Germany
Prima facie evidence against locality comes from reports
suggesting that N1 is favored in the equivalent of (1) in
various languages
One problem with previous studies is that they often
discussed RC attachment as a purely syntactic phenomenon.
However, the surrounding context can affect the intended
meaning of the RC as readers expect clauses to be coherent
and thus prefer attaching the RC to N1 if it provides a
reason or justification for the statement in the matrix clause
Text coherence may be achieved in other ways. Even when surrounding context does not require a causal explanation or justification, it may still affect RC attachment. For example, based on world knowledge, the RC is likely to be attached to N1 (friend) in there was a wake for the friend of the teacher who died yesterday.
In this paper, we suggest that locality preferences can be obscured by contextual factors. Therefore, we report results factoring out preferences stemming from the intended meaning of the context surrounding the RC. A corpus count and a questionnaire provide evidence for contextual effects in RC attachment in Thai. Moreover, two reading experiments support the claim that locality is obeyed in Thai when contextual effects are kept under control.
Thai is an SVO (subject-verb-object) language. The word order of the target construction is N1 of N2 RC. There are no plural markers or morphological agreement, thus ambiguity resolution is often based on plausibility.
Although previous results in Thai provided support for a
non-local N1 attachment preference
Since coherence is important in writing
A total of 4,800 instances of kh ɔ̌ŋ “of” followed by thî :
“that” with up to three intervening words were randomly
selected (Siriwittyakorn et al., 2014, for details) from the six
writing genres of the Thai National Corpus
There were 2,109 instances of the target construction (N1 of N2 RC). Instances were further eliminated if RC attachment was ambiguous (353 instances, 16.74%); if one of the head nouns was a pronoun, a proper name or a biasing noun (e.g., khōn ‘person’, sìŋ ‘thing’; 769 instances, 36.46%); or if they were repetitions (14 instances, 0.66%).
We report the results for the remaining 973 instances coded according to attachment (N1 or N2) and lexical features of N1 and N2 (animacy: animate or inanimate; and concreteness: concrete or abstract; e.g., animate-concrete: man, animate-abstract: government, inanimate-concrete: house, inanimate-abstract: goodness).
More crucially, because we were interested in the influence of the surrounding context, instances were also classified according to disambiguating point. If attachment was resolved within the complex NP (i.e., N1 of N2 RC), it was coded as internally-disambiguated (e.g., “voice of man that was uttered”). If context surrounding the complex NP was needed to determine attachment, it was coded as externally-disambiguated (e.g., in “The writer used only words that have beautiful sound to create rhyme of word that arouses listeners’ emotion,” “rhyme” is more likely to be associated with the “beautiful sound” mentioned in the matrix clause, making it more likely to arouse listeners’ emotion and, hence, more likely to be modified by the RC).
Externally-disambiguated instances only involved
intrasentential contexts
Two native Thai speakers coded all instances independently. Disagreements (5.33%) were settled after discussion with a third native Thai speaker.
Overall (i.e., for internally- and externally- disambiguated instances), there was no reliable difference in attachment (N1 attachment: 460, 47.28%; N2 attachment: 513, 52.72%; where clear, only numbers for N1 are reported from here on). However, when restricted to internally-disambiguated instances (i.e., when attachment did not depend on the surrounding context), the bias towards N2 attachment was reliable (N1: 45.11%; χ2 (1) = 8.12, p = .004; see Table 1).
For the externally-disambiguated instances, N1 attachment was more frequent than N2 attachment (N1: 67 instances, 67.00%; χ2 (1) = 10.9, p = .001). The interaction between attachment (N1 or N2) and point of disambiguation (internal or external) was also reliable (χ2 (1) = 16.37, p < .001).
For animacy and concreteness, results are reported for the
internally-disambiguated instances (trends are the same
when externally-disambiguated instances are included; there
was no interaction between animacy-concreteness and
context; all ps >.10). The trends in Table 1 replicate the
effects of animacy and concreteness in Dutch
Moreover, like in Dutch, there were few instances when both nouns were animate-concrete (for Thai, N1: 9; N2: 12). However, numerical trends were in the opposite direction in Dutch (out of 1,065 instances, N1: 19; N2: 10), but a direct comparison with our results is difficult, since the Dutch counts did not differentiate between internally- and externally-disambiguated RCs.
Context was found to be a factor that can obscure the N2
concrete 4-4 0-2 27-47* 49-122* 80-175*
abstract 0-1 0-0 9-3 45-69* bias, which was significant only when surrounding-context effects were excluded. The fact that context often favored N1 (67% of externally-disambiguated instances) is not surprising. To increase text coherence, writers may prefer N1 attachment as it is the head of the complex NP and is part of the outer clause (e.g., the matrix clause).
In comprehension, readers have been shown to prefer
texts to be coherent using clausal relations such as causality
and justification
A questionnaire study is reported providing evidence that intra-sentential contexts not involving causal relations can affect RC attachment preferences during comprehension.
Participants Sixteen Native Thai speakers aged between 20-30 volunteered to participate. One of them was a graduate student in linguistics who was unaware of the purpose of the study. All participants signed consent forms. Stimuli To investigate the influence of context, 12 corpus fragments (N1 of N2 RC), not involving causality or justification relations, were shown in isolation or with intrasentential context (i.e., the entire corpus sentence). An example translated into English is shown in (2).
(2) a. In isolation: “Rhyme of word that arouses listeners’ emotion” b. With context: “The writer used only words that have beautiful sound to create rhyme of word that arouses listeners’ emotion.”
The native Thai speakers who coded the corpus sentences (see previous section) agreed that N2 (“word”) would be favored in (2a), but N1 (“rhyme”) should be favored in (2b) as “rhyme” is more closely related with “beautiful sound.”
The twelve corpus segments for which reversals were most consistent according to a pre-test were included in the questionnaire (e.g., (2)). According to the pre-test, four items had a bias for N1-attachment when read in isolation, and a bias for N2-attachment when context was included (i.e., N1-N2 items). In the remaining eight items, the trend was in the opposite direction (i.e., N2-N1 items). Procedure and Analyses Participants were asked to choose between two paraphrases for the two possible attachments (e.g., “rhyme arouses listeners’ emotion” and “word arouses listeners’ emotion”). Order of the paraphrases was included as a factor together with context (with context or in isolation). The 12 test sentences were distributed into four lists following a Latin Square design. Each participant saw one list with 66 fillers.
Six fillers had one correct paraphrase. Since all
participants answered these items correctly, results reported
include all participants’ data. Analyses were conducted on R
version 3.0.2
In two items, a portion of the RC was incorrectly shown as part of the matrix clause; and in two items, there were more than two possible host sites. Therefore, only the results for the remaining eight items are reported (trends including all 12 items were similar to those reported).
All means reported are averages over participants.
As predicted there was a main effect of context as the preferred attachment site in isolation (77.08%) was less preferred when context was provided (16.67%; p < .001).
The effect of context was qualified by an interaction with
type of item (N1-N2, or N2-N1: p = .013) but this is not of
interest as it only indicates that the effect was stronger in the
N1-N2 items
The change in preferences indicates that context can affect RC-attachment during comprehension. To achieve coherence, readers do not expect only causality and justification. They also used their world knowledge to relate the meaning of the RC to that of the matrix clause.
In this paper, we are not directly addressing the relation
between frequency in corpora and preferences in
comprehension, but some of the data used to support such a
frequentist explanation should be reexamined. For example,
although in a completion questionnaire using corpus
fragments
In short, studies investigating RC attachment should take
the influence of context into consideration. This is true for
work using corpora, given the rich contexts that often
precede the target construction. But it is also true for
experiments showing individual sentences in isolation given
that intra-sentential context can be a crucial factor affecting
attachment
Although the N2 preference observed in the internallydisambiguated instances of the corpus is compatible with locality, lexical factors other than animacy and concreteness may have contributed to the N2 preference. To provide clearer evidence for locality in Thai, we conducted reading experiments controlling for contextual and lexical factors, and report them as Experiments 2 and 3.
An off-line task (i.e., overall preferences after the sentence was read) was used to determine whether readers prefer N1 as the attachment site of the RC when the RC and the matrix clause are unrelated. Participants Eighteen native Thai speakers volunteered to participate in the experiment. Three of them had taken an introductory course in linguistics. All participants in this and the subsequent experiment and in the norming questionnaires were undergraduate students at Chulalongkorn University. They all signed a consent form. Stimuli Twenty-four ambiguous sentences in which an RC can be attached to either of the two nouns were created. To avoid animacy and concreteness confounds, and to make the results comparable to previous reports for other languages, the two nouns (N1 and N2) were common human nouns. An example is given in (3).
(3) khó:t kh ɔ̌:ŋ nák wîŋ thî: wâ:t r:ûp sǔ aj coach of runner that draw picture beautifully kāmlāŋ -cà ʔ ʔɔ̀ :k baùt FUTURE become-a-monk “The coach of the runner that is good at drawing is going to become a monk.”
Matrix clauses unrelated to the RCs were created (e.g., in (3), there is no relation between being good at drawing and becoming a monk). Five native Thai speakers confirmed that they could not find a relation between the topics in the RC and the matrix clause. None of the five speakers participated in any of the experiments reported here.
The two interpretations (e.g., “coach is good at drawing,” and “runner is good at drawing”) were equally natural according to a norming questionnaire in which a new group of 30 native Thai speakers rated the plausibility of the two interpretations on a five-point scale (Wilcoxon: all ps > .10). Procedure and Analyses Test items were shown in a fixed random order interspersed with 60 fillers so that at least one filler intervened between two test items.
In order to obscure the purpose of the experiment, 28 fillers were ambiguous sentences, some of which contained the word thî: as an RC marker of an unambiguous RC, a complementizer or a preposition. For 46 fillers, the question had only one possible answer to verify that participants were paying attention (all participants scored over 95%).
Sentences were shown individually without line breaks on a computer monitor. After each sentence, a question was displayed on a new screen. This procedure was adopted to prevent participants from consulting previous items or rereading the sentence when answering the question and thus, noticing the ambiguity. Each question was followed by two alternatives with the order counterbalanced across items. For the test items, the question was about attachment (e.g., “Who is good at drawing?”).
How often each participant chose N1 attachment was
included as the dependent variable. Both by-subject and
byitem analyses were conducted on R version 3.0.2
However, it is conceivable that this is not evidence for a locality preference but an unintended effect of the matrix clauses used. Readers may avoid attaching the RC to N1 when it is unrelated to the matrix clause so as to avoid two unrelated clauses referring to the same entity. To address this possibility an on-line experiment was conducted.
An on-line experiment was conducted to show that there is an N2 preference before readers can determine whether the clauses are coherent (i.e., before the matrix clause is read).
Participants A new group of 42 native Thai speakers volunteered to participate in the experiment. Thirteen of them had taken an introductory course in linguistics. Stimuli In the 24 pairs of test items used (see (4) for an example), the RC modified the subject so that the matrix clause would not contaminate the reading times to the RC. (4) a.
N1 attachment lǎ:nchā:j | kh ɔ̌:ŋ | khnūjǐŋ | thî: | ph ə̂ :ŋ jà: | nephew of duchess that just divorce kàp ʔànōŋ | mɨ̂ a ʔā:thít thî:l ɛ́:w | chô:p pāj with Anong(f) when week past like go thîaw | thî: chīaŋmàj travel at Chiang Mai “The nephew of the duchess that got divorced from Anong(f) last week likes traveling to Chiang Mai.” b. N2 attachment lǎ:nchā:j | kh ɔ̌:ŋ | khnūjǐŋ | thî: | ph ə̂ :ŋ jà: | nephew of duchess that just divorce kàp jōŋjtú | mɨ̂ a ʔā:thít thî:l ɛ́:w | chô:p pāj with Yongyut(m) when week past like go thîaw | thî: chīaŋmàj travel at Chiang Mai “The nephew of the duchess that got divorced from Yongyut(m) last week likes traveling to Chiang
Mai.”
All crucial nouns (N1 and N2) were common human nouns. Disambiguation was based on plausibility (e.g., involving gender stereotypes; “f” and “m” in the glosses in (4) indicate the gender of the preceding noun; e.g., in (4a), the RC modifies the nephew as only a man and a woman can get divorced according to current Thai laws).
To confirm the plausibility biases for each RC, a norming questionnaire was conducted with a new group of 47 native Thai speakers. The results indicated that the two plausible interpretations (e.g., for the nephew and Anong (f) to get divorced, and for the duchess and Yongyut (m) to get divorced) were equally natural, and the two implausible interpretations (e.g., for the nephew and Yongyut (m) to get divorced, and for the duchess and Anong (f) to get divorced) were equally implausible (all ps > .10).
Word and bigram frequencies for the disambiguating
words (e.g., “Anong” in (4a), “Yongyut” in (4b)) obtained
from the Thai National Corpus
Participants read sentences one region at a time by
pressing the space bar. After each sentence, a
comprehension question was shown on a new screen. The
question did not query about attachment to avoid drawing
participants’ attention to the point of the experiment. To fit
the width of the screen, sentences were broken into two
lines. For the test items, the nouns and the RC were always
shown together on the first line
Analyses were performed on R 3.0.2
For all regions, attachment, Latin-Square list and their
interaction were set as fixed factors. Since the
disambiguating words differed (e.g., “Anong” in (4a) and
“Yongyut” in (4b)), their lengths, log frequencies (word and
bigram), and scores from the norming questionnaire were
included as additional fixed factors for the critical region.
Random intercepts were included for participants and items.
Because of convergence limitations, only attachment was
included as by-participant random slope, and attachment,
Latin-Square list and their interaction as by-item random
slopes. Outliers beyond three standard deviations were
removed
Comprehension Accuracy Overall (including test items and fillers) question-response accuracy was 99.04%. None of the participants scored less than 94%.
Accuracy for both conditions was high, but N1 attachment (99.80%) was marginally higher than N2 attachment (98.81%) (The results for both by-subjects and by-items are the same: Wilcoxon Vs = 24, ps = .073; bysubject medians for both N1 and N2 attachment: 12; by-item medians for both attachments: 21). Attaching to N1 may reactivate the representation of this noun, thus making it easier for participants to answer the questions, which always included N1 and the matrix predicate.
Reading Times In the critical region, N2 attachment was read faster than N1 attachment (β = -37.74, p = .049) and attachment did not interact with any of the other factors. In the remaining regions, there was no reliable effect of attachment or interaction with list (main effects of list are not of theoretical interest and are not reported).
The reading-time results confirmed that with animate nouns, N2 was the preferred attachment site. The results are compatible with those of the off-line task in Experiment 2, confirming the locality preference in attachment in Thai. Because the preference was observed before the matrix predicate was read, we can be confident that the present result was not affected by readers trying to make attachment coherent with the matrix clause.
The corpus count and Experiment 1 indicate that context affected attachment both in production and in comprehension. With intra-sentential contexts factored out, the N2 preference in the corpus count was compatible with locality. The effect of locality was further confirmed by offline and on-line tasks (Experiments 2 and 3).
Previous results were primarily concerned with causality
and justifications between clauses in comprehension
We emphasized coherence but there are many ways that
context may affect preferences as sentences are read
Further studies are needed but if previously-reported N1 attachment preferences in various languages can be ascribed to context, then a local attachment preference can be held as a universal principle, without the need for cross-linguistic parameterizations in the way people process sentences.
In a recent proposal,
Similarly, the items of some studies reporting a non-local
attachment preference could not be interpreted as pseudo
RCs
The results that pseudo-RCs have been claimed to explain, may be reduced to contextual effects where text coherence favors attaching the RC to N1 to make its time reference overlap with the time of the matrix clause.
We examined contextual effects in RC attachment in Thai. A corpus count and a questionnaire indicated that context affected attachment. In the corpus results and two reading experiments, there was an N2-attachment preference when contextual effects were excluded. This is compatible with the assumption that locality is a universal parsing principle, which is modulated by context and lexical features such as animacy and concreteness.
This work is a part of the first author’s doctoral dissertation.