Processing Chinese object-topicalization structures in simple and complex sentences

1.1 Background: topic-comment structure in Chinese

Since Li and Thompson’s (1976) seminal work, subject prominence and topic prominence have been used as typological parameters to distinguish the basic sentence structure of languages. English is subject-prominent because it stresses on the relation between subject and predicate, whereas Chinese is topic-prominent (Chao 1968; Li and Thompson 1976, 1981), with a variety of structures built on the relation between topic and comment, as shown in (1).

The above topic-comment structures can be subcategorized into two types: (i) GAPPED topic structure (1a–b), where the topic noun phrase (NP) bears a co-referential relation with an empty category (marked as ‘e’) in the sentence, and (ii) NON-GAPPED topic structure or “Chinese-style topics” (Chafe 1976), where the topic NP bears a part-whole relation (1c–d) or no direct relation at all (1e–f) with the following NP. It is precisely the “aboutness” condition between the topic NP and the rest of the clause that defines the topic-comment structure (Gundel 1985; Gundel and Fretheim 2006).

One distinct feature of the topic-comment structure in Chinese is its wide distribution in daily life, in contrast to the restricted use in English, as evidenced by the existing few corpus studies on the gapped topic structure. In Liu and Gao’s (2010) corpus of vernacular Beijing dialect, 992 tokens of gapped topic structures were found, accounting for 7.6 % of the total 13,000 clauses. Yet in Gregory and Michaelis’s (2001) switchboard telephone corpus, only 44 tokens of English gapped topic structures were found, accounting for 0.13 % of the total 32,805 clauses. Furthermore, in Wang and Li’s (2016) examination of Chinese versus English dialogues – both lasting 5.5 h with similar themes, the object-topicalization structure (i.e., ex. [1a]) is 42.4 times more frequent in spoken Chinese (297 tokens) than in spoken English (7 tokens). As said, gapped topic structure is only one subtype of topic-comment structures in Chinese. If we turn to non-gapped topic structure – roughly, 1.99 times more frequent than gapped topic structure in Zhang’s (2017) spoken corpus of a talk-show program, it is even more widely distributed. In short, given the pervasiveness of topic-comment structures in Chinese (Chao 1968), a Chinese parsing system should be fairly familiar with object-topicalization, a structure of our primary concern.

The other interesting property of Chinese topic-comment structures is that, unlike English, where topics are restricted to main clauses (Hooper and Thompson 1973), elements embedded within a relative clause can also be topicalized in Chinese (Huang et al. 2009: 209–211; Xu and Langendoen 1985; Xu and Liu 1998: 38–41). As shown below, in the canonical SVO structure (2a), the object embedded in the relative clause can be topicalized, forming the object-topicalization sentence (2b). In contrast to the well-formedness of (2b), its English counterpart is ungrammatical.

1.2 Existing processing work on object-topicalization structure in Chinese

Existing psycholinguistic studies on Chinese topic-comment structure, though few, have mainly focused on the object-topicalization structure, namely, Topic-Subject-Verb (TSV). Assuming topicalized object-NPs are derived by movement with a null operator binding an empty category (Huang 1982; Huang et al. 2009; Li 1990, 2000; Shi 2000), these studies have the same logic as memory-based approaches to long-distance dependency processing: Upon encountering the verb, Chinese comprehenders need to retrieve the topic from the working memory and fill it in. Compared to the canonical SVO structure, this gap-filling process in the TSV structure would incur processing costs due to (i) the long distance and (ii) potential intervening competitors that might cause interference in the retrieval/integration process.

However, results from existing work have been mixed. Most studies showed no processing differences at the verb in TSV sentences (self-paced reading: Kuo [2006]; Li and Wu [2020]; cross-modal priming: Cai and Dong [2010]), indicating a lack of locality effect that is not predicted by memory-based theories. To our knowledge, only two studies found processing difficulties at the verb in TSV sentences relative to SVO sentences, indexed by reading time slowdown (Huang and Kaiser 2008) or an enhanced P600 (Yang and Liu 2013). Yet unlike most studies, Huang and Kaiser (2008) used subject-topicalization in conjunction with a parasitic-gap construction. While Yang and Liu (2013) used the object-topicalization structure, their sentence-final constituents were not fully matched, making it hard to compare verbal classifiers (e.g., liang-jiao ‘[lit.] two-foot, [kicked] twice’) to human referents (e.g., someone whose name is Wang Wu). Thus, aspects of syntactic inconsistencies in these two studies undermine the generalizability of their conclusions.

Among the studies that failed to detect any effect at the verb, Li and Wu (2020) actually reported additional findings. Reaction times (RTs) were significantly longer at the position of S in TSV sentences than RTs at the same position in SVO sentences. When analyzing by part of speech, they found that the same lexical verb, though in different positions, was significantly faster in TSV than SVO sentences (373 ms vs. 408 ms, t = 3.81). They also asked Chinese participants to complete a gated sentence-completion task, and found an interesting trade-off between subject-predicate continuations and topic-comment continuations. In the SVO condition, an initial animate noun phrase (i.e., Gate 1, N…) led participants to produce 95.28 % SV continuations but no topic-comment continuations, and an added verb phrase (i.e., Gate 2, NV…) further promoted SVO continuations to 100 %. In the TSV condition, however, the rate of topic-comment continuations was 30.16 % at Gate 1 (i.e., prompted by an initial inanimate N…), but sharply increased to 96.03 % at Gate 2 (i.e., with an added animate noun phrase, NN…). Taking together the production and comprehension data, Li and Wu (2020) reasoned that Chinese participants had a strong bias for interpreting the initial NP as Subject, but upon seeing the second NP in TSVs, they readjusted the subject-reading to a topic-reading fairly quickly, such that they had no difficulty understanding the rest of the sentences, and even processed the verb much faster in TSVs than the verb that occurred earlier in SVOs.

The findings of Li and Wu (2020) appear enlightening, potentially constituting evidence for experience/surprisal-based theories while imposing challenges to memory-based theories in terms of Chinese object-topicalization processing. We note, however, one inherent problem associated with this line of research, namely “comparability”, because TSV and SVO differ in word order. Given that processing research is primarily interested in how parsing varies dynamically in real time, it is perhaps more revealing to examine data by regions than by part of speech. Yet when comparing TSVs to SVOs by regions, Li and Wu (2020) did not regress out extraneous factors related to different parts of speech, nor did they take into consideration a spillover effect from the preceding region. In fact, doing such statistical analysis requires enough data points, whereas they only tested 32 participants. Thus, it is worthwhile to replicate their results with a large sample size.

In sum, existing processing work on Chinese topicalization structures has almost exclusively focused on simple sentences or main clauses, without yielding consistent findings. Furthermore, none of them has considered examining potential retrieval/integration costs in sentences with object-NPs that are extracted from complex relative clauses. More importantly, given their primary concern about syntactic movement, virtually no study was designed to evaluate competing processing accounts. The present study aimed to fill these gaps, recasting the findings in light of two competing processing theories.

1.3 Research goals

We aimed to investigate how Chinese comprehenders process object-topicalization in simple and complex structures. TSV sentences provide us a good testing ground to evaluate memory-based theories and experience-based theories, because they occur frequently, involve non-adjacent dependencies between topics (i.e., fronted objects) and gaps (or verbs), and are allowed to have topics extracted from relative clauses that are known to be structurally complex.

We conducted three self-paced reading experiments. Experiment 1 used simple sentences from Li and Wu (2020) with slight modification, to establish a basic processing profile of Chinese TSVs at the second NP (henceforth NP2) and critically, at the verb. Experiment 2 focused on complex sentences, to further investigate whether an added structural complexity changes Chinese comprehenders’ parsing biases at the NP2 and verb regions. Experiment 3 narrowed down to certain conditions of Experiment 2, with better controlled stimuli, to facilitate direct comparisons between sentences. To foreshadow our results, we find that consistent with experience/surprisal-based theories, Chinese participants generally tended to interpret the sentence-initial topic as Subject, but they had no difficulty integrating the topic NP into the verb, regardless of structural complexity.

2.1 Predictions

2.2 Method

2.3 Results

2.3.2 Reading times

Figure 1 shows the mean and 95 % confidence intervals of the trimmed RTs by regions. Clearly, there is a reversal of RT patterns in the critical regions: at the determiner phrase and the NP2, the TSV condition had a processing disadvantage; but at the verb, the TSV sentences appeared to be read faster than the SVO sentences. Table 1 reports the statistical results of five regions (i.e., from W4 to W8), each including spillover effects from the previous region.

Figure 1:

Experiment 1 word-by-word reading times. Error bars represent standard errors. Critical regions (NP2 & MV) are shown in rectangle boxes (*marks significant effects; n.s., means not significant).

Table 1:

Effects of Word Order by regions in Experiment 1. The dependent variable for each region is the residuals of a linear mixed effect model on log-transformed reading time with extraneous factors including log word frequency, number of strokes, word length and region.

Word	maximal model with random slopes	Contrast	Coef.	SE	t-value
W4 (DetP in TSV condition)	order + (1+order|subj) + (1+order|item)	TSV Pre_logRT	0.018 0.367	0.007 0.019	2.734* 19.715
W5 (NP2 in TSV condition)	order + (1+order|subj) + (1+order|item)	TSV Pre_logRT	−0.007 0.357	0.008 0.019	−0.869 18.374
W6 (NP2+1 in TSV condition)	order + (1+order|subj) + (1+order|item)	TSV Pre_logRT	−0.009 0.310	0.006 0.019	−1.415 16.754
W7 (MV in TSV condition)	order + (1+order|subj) + (1+order|item)	TSV Pre_logRT	0.002 0.358	0.007 0.020	0.339 18.191
W8 (MV+1 in TSV condition)	order + (1+order|subj) + (1+order|item)	TSV Pre_ logRT	−0.003 0.321	0.014 0.021	−0.211 15.524

1. Bold t-values marked with * mean the results of the fixed factor reach statistical significance.

At W4 (i.e., the determiner phrase in the TSV condition), TSVs were read significantly more slowly than SVOs.

At W5 (i.e., NP2 in the TSV condition) and W6 (i.e., NP2+1 in the TSV condition), TSVs were numerically slower than SVOs, without reaching statistical significance.

At W7 (i.e., MV in the TSV condition) and W8 (i.e., MV+1 in the TSV condition), no difference was found between TSVs and SVOs.

2.4 Discussion

With careful statistical analyses, we largely replicated Li and Wu’s (2020) by-region findings. First, at the region immediately preceding the NP2, anticipation of another noun in the TSV structure triggered a significant slowdown in RTs relative to the same position (i.e., W4) in the SVO structure, consistent with the prediction of experience/surprisal-based theories. No effect was found at the NP2, where the difficulty of its preceding word (i.e., spillover effects from W4) caused the reportedly a priori surprisal effect to go away. Note that different from Li and Wu (2020), the reanalysis effect occurred early in our study, possibly for two reasons: (1) A lot more filler trials with the regular SVO order were included in our study than in Li and Wu (2020), potentially heightening our participants’ sensitivity to the TSV structure that was relatively low in frequency. (2) The small sample size in Li and Wu (2020), totaling 32 participants only, made it difficult to detect the anticipatory effect at the noun-denoting determiner region. The early effect in our replication experiment reflects that the nature of human sentence processing is incremental, predictive, and probabilistic (e.g., Demberg 2010).

The second finding is that at the critical verb (i.e., W7), TSVs were faster, though numerically only, than SVOs. This pattern is hard to reconcile with memory-based theories, which predict retrieval/integration costs at the verb in the TSV condition.

The lack of effects at the post-NP2 regions overall, together with the reanalysis effect at the NP2-denoting region in the TSV condition, can be explained by experience/surprisal-based theories. Specifically, as SVO is the dominant word order in Chinese, the TSV structure was initially not the first or preferred parse. But once it was recognized, its remaining part down the stream turned out to be highly accessible. Given Chinese speakers’ extensive experience and considerable familiarity with the topic-comment structures in general, no processing difficulty ensued.

While the results of Experiment 1 are not as predicted by memory-based accounts, it remains unclear whether the long-distance retrieval/integration effects can be detected in more complex structures. This motivated our Experiment 2, where we created a modifying relative clause structure that intervenes between the topicalized NP1 and the verb. This lengthened distance will help us to further investigate whether Chinese comprehenders encounter difficulties at the NP2, which is now the head noun of the prenominal relative clause, and at the verb down the parsing stream. In addition, we also created a new type of object-topicalization to serve as a control condition for direct comparisons of TSVs, by taking advantage of the fact that objects embedded in relative clauses can be topicalized in Chinese.

3.1 Predictions

Given that the primary goal of our research is to evaluate (i) the consequence of the subject-first bias at the NP2 in TSV sentences and (ii) the presence versus absence of the integration cost at the main verb, the predictions outlined below focus on the NP2 (i.e., the head noun of the RC in the two TSV conditions) and the main VP.

3.1.1 Experience/surprisal-based theories

To derive the prediction of experience/surprisal-based theories, we first estimated the structural frequencies of different conditions from spoken corpora. In our own spoken corpus, no TSV-RC structures were found. In Zhang (2017), out of a total of 1,150 topic structures, only 29 topic structures were found to occur in the subordinate clauses (including RCs and object complement clauses). Given that our target structure rarely occurred in corpora, we conducted two gated sentence-completion tests.

Gate-1 Test: In Gate-1 test, sentence fragments were truncated before DE, as in (5a–d). In (5a), the sentence-initial NP1 (zhexiang jixian yundong ‘this extreme sport’) was intended as the fronted argument (i.e., the direct object) of the immediately following VP in which the second word was a transitive verb (ganyu canjia ‘dare to join e_i’). Thus, the fragments could be continued as a topic structure (i.e., “T+V”). In (5c), however, the NP1 cannot serve as the fronted argument of the following VP in which the second word was either an intransitive verb or a noun (ganyu maoxian ‘dare to adventure’). In the two SVO conditions ([5b] and [5d]), the fragments given can be interpreted as canonical VPs.

(5)

Gate 1

a.

TSV-RC

zhexiang jixian yundong ganyu canjia________。 (this-CL extreme sport dare join _______.)

b.

SVO-RC

ganyu canjia zhexiang jixian yundong_______。(dare join this-CL extreme sport __________.)

c.

TSV-MC

zhexiang jixian yundong ganyu maoxian______。(this-CL extreme sport dare adventure _________.)

d.

SVO-MC

ganyu maoxian__________________________。 (dare adventure ______________.)

Gate-2 Test: In Gate-2 test, sentence fragments were truncated up until DE, as in (6a–d). An added nominalizer (or relativizer) DE makes an RC continuation probable, which in turn makes an NP2 (i.e., the head noun of RC) more likely to be anticipated. Thus, we could find out whether Chinese speakers would be more likely to produce a TSV structure in the presence of complex RC sentences in (4a) and (4c), compared to their responses in Gate-1 test.

(6)

Gate 2

a.

TSV-RC

zhexiang jixian yundong ganyu canjia de______。(this-CL extreme sport dare join DE_____.)

b.

SVO-RC

ganyu canjia zhexiang jixian yundong de______。(dare join this-CL extreme sport DE ______.)

c.

TSV-MC

zhexiang jixian yundong ganyu maoxian de_____。(this-CL extreme sport dare adventure DE _____.)

d.

SVO-MC

ganyu maoxian de_________________________。(dare adventure DE ___________.)

For both Gate-1 and Gate-2 tests, 24 sets in four conditions were counterbalanced into four lists using a Latin Square procedure, and then pseudorandomized with 32 filler items in various structures. Those filler fragments were taken from the initial, medial, or final positions of their original forms. We used the online platform wenjuanxing (https://www.wjx.cn) to administer the tests. Two groups of 40 native speakers of Mandarin Chinese participated in the two tests, respectively. None of them took any other experiments. Each participant was paid 30 yuan. The mean ages of participants in two tasks were 22.5 (SD = 2.62) and 22.2 (SD = 2.29), respectively.

A total of 960 sentences were obtained for each test. We eliminated 45 ungrammatical sentences in Gate-1 (TSV-RC: 20; SVO-RC: 4; TSV-MC: 21) and 23 in Gate-2 (TSV-RC: 7; SVO-RC: 2; TSV-MC: 7; SVO-MC: 7) tests, respectively. The remaining 915 grammatical continuations were then coded into two structural types:

1. TC structures, including TSVs with gaps as in (7a) and those without gaps as in (7b);

2. SV structures, including SVs, VOs, and SVOs.^[4]

Each type of structure was further classified into RC or non-RC, depending on whether the fragment was immediately followed by (i.e., was continued as) an RC or not.

(7)

a.

gapped TSV continuation in the SVO-MC condition (Gate 2)
guanzhu	minsheng	de zhongyaoxing i ,	women	meigeren
care-about	people’s life	DE importance	1PL	every-CL-man
dou	yinggai	mingbai e i .
all	should	understand
‘Regarding the importance of caring about people’s life, every one of us should understand.’

b.

non-gapped TC continuation in the SVO-MC condition (Gate 2)
buxi	shejiao	de ren	yiban	xingge
dislike	socializing	DE people	generally	personality
bijiao neilian
rather introvert
‘For people who do not enjoy socializing, their personalities are generally introvert.’

Table 2 shows the number and percentage of each structural type of grammatical continuation for the four types of fragments in Gate-1 and Gate-2 tests. Figure 2 shows the production rates of TC versus SV in terms of RC versus non-RC in grammatical continuations at Gate 1 and Gate 2. Two observations are noteworthy. First, at Gate 1 the production rates of SV sentences were high overall across conditions (mean = 77.38 % or 708/915), in contrast to the low production rate of TC sentences (mean = 22.62 % or 207/915). A generalized linear mixed model with a binomial link function and crossed varying intercepts for subjects and items shows a main effect of Word Order (β = 2.51, SE = 0.20, z = 12.57, p < 0.0001), but no other effects: Participants produced significantly more SVs in the SVO conditions ([5b] and [5d]) than in the TSV conditions ([5a] and [5c]), suggesting that prior to DE, Chinese speakers strongly preferred to interpret the sentence-initial NP1 as the subject, even in the TSV conditions.

Table 2:

Number and percentage of SV versus TC by RC or non-RC in grammatical continuations for four conditions at two gates in the sentence completion test of Experiment 2.

Condition		SV				TC				Total
		RC		Non-RC		RC		Non-RC
		#	%	#	%	#	%	#	%	#	%
Gate 1	TSV-RC	0	0	106	48.18	71	32.27	43	19.55	220	100
	SVO-RC	69	29.24	164	69.49	1	0.44	2	0.84	236	100
	TSV-MC	0	0	134	61.87	76	34.70	9	4.11	219	100
	SVO-MC	28	11.67	207	86.25	5	2.08	0	0	240	100
	SUM	97	10.6	611	66.78	153	16.72	54	5.90	915	100
Gate 2	TSV-RC	63	27.04	16	6.87	135	57.94	19	8.15	233	100
	SVO-RC	193	84.65	37	16.23	8	3.51	0	0	228	100
	TSV-MC	11	4.72	88	37.77	116	49.79	18	7.73	233	100
	SVO-MC	166	71.24	49	21.03	17	7.30	1	0.43	233	100
	SUM	433	46.21	190	20.28	276	29.46	38	4.06	937	100

Figure 2:

The production rates of grammatical TC versus SV continuation in terms of RC versus non-RC at Gate 1 and Gate 2 in the sentence completion test of Experiment 2.

Second, while the production rates of RC were low for both SVs (mean = 10.6 % or 97/915) and TCs (mean = 16.72 % or 153/915) at Gate 1, the presence of DE at Gate 2 led to a drastic increase of RCs for both SV (SVO-RC: 84.65 %; SVO-MC: 71.24 %) and TC continuations (TSV-RC: 57.94 %; TSV-MC: 49.79 %). A generalized linear mixed model shows a main effect of Word Order (β = −2.24, SE = 0.16, z = −14.17, p < 0.0001), a significant interaction between Word Order and Structural Complexity (β = 0.37, SE = 0.12, z = 2.89, p = 0.0039), but no main effect of Structural Complexity. This interaction was mainly reflected by the two TSV conditions where more SV/non-RCs were produced in TSV-MC (6c) (88/233 or 37.77 %) than in TSV-RC (6a) (16/233 or 6.87 %). But the two TSV conditions did not differ in TC/RC productions (135 vs. 116, or 57.94 % vs. 49.79 %), suggesting that upon seeing DE, Chinese participants had no problem expecting an RC and producing a head noun of the RC (i.e., an NP2), even in the non-canonical topic-comment structures. Note that the production rate of RCs is high in two SVO conditions but around the chance level in the two TSV conditions, suggesting that though expected, RC structures occurring in topic sentences were still indeterministic at DE.

From the results of two sentence-completion tests, we derive three predictions of experience/surprisal-based theories. First, given the strong subject-reading of NP1 prior to DE (even in the case of TSV conditions), SVOs ([4b] and [4d]) with a higher frequency of occurrence are predicted to be easier to process than TSVs ([4a] and [4c]), especially in regions prior to the main VP. Second, given that the presence of DE led Chinese speakers to be more deterministic to posit RC structures in SVOs than in TSVs, with no significant differences in RC production between two TSVs, increased processing costs are predicted in TSV conditions relative to SVO conditions, incurred by a reanalysis of NP1 from subject interpretation to topic interpretation, possibly at DE, the head noun of RC (i.e., the NP2), and subsequent spillover regions.

The third prediction is directly related to the critical region of the main VP. Specifically, once the reanalysis involved in TSV sentences is fully recovered down the parsing stream, no parsing differences are predicted between the two TSV conditions ([4a] vs. [4c]) by the time the main VP is seen.

3.2 Method

3.3 Results

3.3.2 RTs

Figure 3 shows the mean of trimmed RTs for each region of the target sentences. As shown in Figure 3, the two TSV conditions patterned alike from the RC-verb (W5) all the way down to the end of the sentence (W11). Compared to the SVO conditions, the TSV conditions showed a consistent processing disadvantage from W5 to W8, spanning the whole RC and the immediate post-HN region. As expected, the TSV structure in conjunction with the RC greatly enhanced the processing difficulties of our TSV conditions. But in the main-clause regions, the RT differences appear to be much reduced between the TSV conditions and the SVO conditions.

Figure 3:

Experiment 2 word-by-word reading times. Error bars represent standard errors. Critical regions (NP2, MV1 & MV2) are shown in rectangle boxes (*marks significant effects; n.s., means not significant).

Table 3 reports the statistical results of five regions of interest, including the final maximal random effects models.

Table 3:

Main effects of Word Order and Complexity and their interaction by region of interest in Experiment 2. The dependent variable for each region is the residuals of a linear mixed effect model on log-transformed reading time with extraneous factors including log word frequency, number of strokes, word length, and region.

Word	Maximal model with random slopes	Contrast	Coef.	SE	t-value
W7 (HN in TSV conditions)	order*clauseType + (1+order*clauseType|subj) + (1+order+clauseType|item) + prev_logRT	Word Order	−0.037	0.006	−6.578*
		Complexity	−0.015	0.006	−2.285*
		Word Order × Complexity	−0.010	0.006	−1.640
		Prev_logRT	0.391	0.019	20.069
W8 (HN+1 in TSV conditions)	order*clauseType + (1+order*clauseType|subj) + (1+order+clauseType|item) + prev_logRT	Word Order	−0.019	0.007	−2.846*
		Complexity	−0.010	0.008	−1.718
		Word Order × Complexity	−0.010	0.006	−1.732
		Prev_logRT	0.426	0.023	18.923
W9 (HN+2 in TSV conditions)	order*clauseType + (1+order*clauseType|subj) + (1+order*clauseType|item) + prev_logRT	Word Order	−0.010	0.007	−1.472
		Complexity	−0.002	0.007	−0.363
		Word Order × Complexity	−0.006	0.006	−0.978
		Prev_logRT	0.418	0.019	21.761
W10 (MC-V1 in TSV conditions)	order*clauseType + (1+order*clauseType|subj) + (1+order*clauseType|item) + prev_logRT	Word Order	0.006	0.006	1.093
		Complexity	−0.002	0.006	−0.394
		Word Order × Complexity	0.004	0.006	0.712
		Prev_logRT	0.413	0.021	19.781
W11 (MC-V2/N in TSV conditions)	order*clauseType + (1+order*clauseType|subj) + 1+order+clauseType|item) + prev_logRT	Word Order	−0.019	0.009	−2.084*
		Complexity	−0.026	0.008	−3.355*
		Word Order × Complexity	−0.015	0.007	−2.186*
		Prev_logRT	0.380	0.029	12.992

1. Bold t-values marked with * mean the results of the fixed factors reach statistical significance.

At W7 (i.e., the HN ren ‘man’ in the two TSV conditions [4a] and [4c]), we found a main effect of Word Order, a main effect of Structural Complexity, and no interaction. TSVs ([4a] and [4c]) were read more slowly than SVOs ([4b] and [4d]). Sentences with topic- and object-NPs occurring in RCs ([4a] and [4b]) were read more slowly than sentences with topic- and object-NPs occurring in MCs ([4c] and [4d]).

At W8 (i.e., HN+1 in the two TSV conditions), we only found a main effect of Word Order. TSVs were read more slowly than SVOs.

At W9 (i.e., HN+2 in the two TSV conditions), we found no main effect of Word Order, no main effect of Structural Complexity, and no interaction.

At W10 (i.e., MC-V1 in the two TSV conditions), again we found no main effect of Word Order, no main effect of Structural Complexity, and no interaction.

At W11 (i.e., MC-V2/N in the two TSV conditions), we found a main effect of Word Order, a main effect of Structural Complexity, and an interaction between Word Order and Structural Complexity. Follow-up tests unpacking this interaction showed that TSVs were read more slowly than SVOs only when the topicalized NP was extracted from the main clauses ([4c] vs. [4d]/canjia ‘join’ vs. yundong ‘sport’: β = −0.06, SE = 0.02, t = −2.70), but not when the topicalized NP was extracted from the RCs ([4a] vs. [4b]/maoxian ‘adventure’ vs. maoxian ‘adventure’: t = −0.33). Furthermore, RCs were read more slowly than MCs only in the canonical SVO sentences ([4b] vs. [4d]/maoxian ‘adventure’ vs. yundong ‘sport’: β = −0.08, SE = 0.02, t = −3.35), but not in the TSV sentences ([4a] vs. [4c]/maoxian ‘adventure’ vs. canjia ‘join’: t = −0.99).

3.4 Discussion

Focusing on topic-NPs that are extracted from either an RC or a main clause, Experiment 2 yielded three findings. First, TSVs were read more slowly than SVOs at the NP2 (i.e., the HN of RC) and in the immediately following (i.e., the first spillover) region, but in the second spillover region, no differences were found between TSVs and SVOs. Second, at the second word of the critical main VP (i.e., MC-V2/N), no differences were found between the two TSV conditions, regardless of whether the topic NP was extracted from the RC (4a) or from the main clause (4c), and no differences were found between the TSV-RC condition (4a) and its counterpart SVO-RC condition (4b). Third, again at the second word of the critical main VP (MC-V2/N), while the TSV-MC condition (with an intervening RC, i.e., [4c]) were read more slowly than its counterpart SVO-MC condition (4d), RTs in the two SVO conditions also significantly differed, with the SVO-RC condition (4b) being read more slowly than the SVO-MC condition (4d). We discuss each finding below.

First, the processing difficulty associated with the TSV sentences (relative to the SVO sentences) at the HN (i.e., NP2) and in the first spillover region (HN+1) are analogous to the significant RT slowdown at NP2 in the simple sentences of Experiment 1. Notice, however, different from Experiment 1, the difficulty with the TSV sentences in Experiment 2 appeared much earlier (beginning at W4), continued for the RC regions and the subsequent region (ending at W9), suggesting that our manipulation of structural complexity was successful, and that the presence of RC – either contains an empty category that bears a co-referential relation with the topic or intervenes between a topicalized NP and the main verb – might be responsible for the inflated RTs.

Importantly, this prolonged processing difficulty could be explained by experience/surprisal-based theories. The extra processing costs were caused by a reanalysis of the sentence-initial NP1 in TSVs due to (i) comprehenders’ less experience (familiarity) with TSVs relative to SVOs and (ii) their initial subject-interpretation bias. Recall that in our spoken corpus investigation, SVO sentences outnumber object-topicalization sentences by 4.2 times (or 196/47). Likewise, as shown by our sentence-completion results (Figure 2), prior to DE Chinese participants not only predominantly produced SV sentences in the SVO conditions, but also consistently produced more SV sentences in the TSV conditions. These production data show that Chinese speakers do have a bias toward subject interpretation of the sentence-initial NP1. Furthermore, there was an added ambiguity in the two TSV conditions where the NP1 functions as the topic of the TSV structure. Specifically, while the NP1 was highly likely to be misinterpreted as the matrix subject (due to the subject-reading bias), the next available RC-verb (i.e., W4) would directly conflict with this misanalysis due to the verb semantics where it could not be plausible for an inanimate NP (‘this extreme sport’) to serve as a sentient agent to perform the action <DARE TO DO SOMETHING>. The presence of the relativizer DE, as well as the following NP2 (i.e., the HN of the RC), further clearly informed our Chinese participants that the NP1 was not the matrix subject. In other words, participants had to realize that no strict subject-predicate relation – but a convenient topic-comment relation instead – could be built between the RC-verb and the sentence-initial NP1, which had to be reanalyzed as a topic originated from the RC (in the condition [4a]) or from the MC after the HN (in the condition [4c]), hence the persistent RT slowdown up until the HN and its spillover region in TSVs. As a side note, our explanation offered here for the RC region is fairly similar to the retrieval bottleneck account for the long-standing difficulty of English object RCs (e.g., Staub et al. 2017).

The second set of findings is the lack of differences between the two TSV conditions at the critical MC-V2/N. This null result is noteworthy in light of the integration-cost metric of the DLT. In the TSV-RC condition (4a), the filler-gap dependency is completed early (i.e., within the RC), with two discourse referents (i.e., RC-V1, RC-V2/N) that intervene between the topic and the post-verb gap. But in the TSV-MC condition (4c), the filler-gap dependency spans across the whole intervening RC, with five discourse referents (i.e., RC-V1, RC-V2/N, HN, MC-V1, MC-V2/N) in between. Therefore, the fronted object-NP should be more difficult to be retrieved/integrated in (4c) than in (4a). Even though the absence of effects does not necessarily mean the effect is absent, the fact that no such difficulties were found at the retrieval/integration site of MC-V2/N imposes challenges to locality effects of memory-based theories.

The third finding regarding the TSV-MC condition (4c) being more difficult to process than its counterpart SVO-MC (4d) at the MC-V2/N might appear to lend support to memory-based theories, which predict processing costs in the region where the memory representation of a filler is retrieved and integrated for completion of filler-gap dependency. This is also consistent with the findings from cross-linguistic corpora that languages tend to minimize dependency length (Futrell et al. 2015; Liu 2008; Liu et al. 2017). Note however, at the MC-V2/N the other canonical SVO-RC condition (4b) was also more difficult to process than the SVO-MC condition (4d), even though these two conditions had exactly the same structure. As the two SVO sentences differ only in the VP types, we speculate that the modifying RC is rendered much shorter in (4d) with an intransitive RC-V2/N than in (4b) with a transitive RC-V2, making it much earlier (and thus easier) for the head noun to be recognized as the immediate constituent of the sentence, namely the matrix subject (Hawkins 1994, 2004). Furthermore, in contrast to (4b) which has a long RC but a short main clause, (4d) begins with a short RC and ends with a heavy main clause, thus fitting the general short-before-long preference in verb-medial languages such as English and Chinese (Arnold et al. 2000; Hawkins 2004; Lu 1993, 2004; Stallings et al. 1998). In short, we attribute the overall processing advantage of the canonical SVO-MC (4d) relative to the TSV-MC (4c) and SVO-RC (4b) conditions to (i) the superiority effect of SVO word order in Chinese, (ii) the earliness/easiness of recognizing the matrix subject, and (iii) a general preference for putting the heavy constituents in the end. But given our primary concern of this paper, we refrain from further discussing the processing difficulties associated with the SVO-RC (4b) versus (4d) and with the TSV-MC (4c) versus (4d).

Taken together, the overall patterns – particularly regarding the effects found at the NP2 (i.e., HN) – are more in line with experience/surprisal-based theories, whereas no clear evidence is found in Experiment 2 for memory-based theories. Importantly, the interaction effect found at the critical MC-V2 could not be taken as solid evidence for the DLT.

4.1 Method

4.2 Results

4.2.2 RTs

Figure 4 shows the mean and 95 % confidence intervals of trimmed RTs for each region of target sentences in Experiment 3. Just as in Experiment 2, the two TSV conditions patterned alike within the RC regions, both appearing to be more difficult to process than the canonical SVO condition throughout the RC and the first post-HN spillover region (W8). But the differences across the three conditions appeared much reduced towards the main clause, with RTs being converged at the MC-V1 ‘dare’.

Figure 4:

Experiment 3 word-by-word reading times. Error bars represent standard errors. Critical regions (NP2 & MVs) are shown in rectangle boxes (*marks significant effects; n.s., means not significant).

Table 4 reports the details of modeling structures and the statistic results of the critical regions from W7 to W10.

Table 4:

Main effects of Word Order and Complexity and their interaction by region of interest in Experiment 3.

Region	Maximal model with random slopes	Contrast	Coef.	SE	t-value
W7 (HN)	logRT ∼ condition + (1+condition|subj) + (1+condition|item)	Intercept(=a)	5.861	0.039	150.58
		condition_b	−0.066	0.020	−3.26*
		condition_c	0.018	0.021	0.86
W8 (HN+1)	logRT ∼ condition + (1+condition|subj) + (1|item)	Intercept(=a)	5.870	0.036	160.93
		condition_b	−0.066	0.019	−3.49*
		condition_c	0.030	0.019	1.53
W9 (HN+2)	logRT ∼ condition + (1+condition| subj) + (1+condition|item)	Intercept(=a)	5.825	0.035	165.88
		condition_b	−0.023	0.021	−1.07
		condition_c	−0.001	0.020	0.03
W10 (MC-V1)	logRT ∼ condition + (1+condition|subj) + (1+condition|item)	Intercept (=a)	5.806	0.035	166.81
		condition_b	−0.008	0.019	−0.43
		condition_c	0.006	0.019	0.31
W11 (MC-V2)	logRT ∼ condition + (1+condition|subj) + (1+condition|item)	Intercept(=a)	5.823	0.034	172.98
		condition_b	−0.015	0.023	−0.67
		condition_c	0.029	0.019	1.50
W12 (MC-V2/N+1)	logRT ∼ condition + (1+condition|subj) + (1|item)	Intercept(=a)	5.854	0.034	171.18
		condition_b	0.032	0.020	1.63
		condition_c	−0.006	0.019	−0.29

1. Bold t-values marked with * mean the results of the fixed factor reach statistical significance.

At W7 (i.e., the HN) and W8 (i.e., HN+1), the SVO-RC condition was read significantly faster than the baseline TSV-RC condition, but no differences were found between the two TSV conditions.

At W9 (i.e., HN+2), no differences were found between the SVO-RC and TSV-RC conditions, and between the two TSV conditions.

At W10 (i.e., MC-V1 ganyu ‘dare’), W11 (i.e., MC-V2/N maoxian ‘adventure’ in [8a/b]/canjia ‘join’ in [8c]), and W12 (i.e., ‘and’), we consistently found no differences between two TSV conditions, nor between the SVO-RC and TSV-RC conditions.

4.3 Discussion

Experiment 3 replicated the results of Experiment 2. Importantly, now that the added continuing clause eliminated the potential sentence-final wrap-up effects in Experiment 2, still no differences were found at the main VP regions across conditions. Clearly, this consistent lack of locality effects is not predicted by the integration metric of the DLT.

The processing disadvantage of the two TSV sentences versus SVO-RC sentences at the HN and its spillover region (HN+1) is, again, not surprising (see detailed explanations in Section 3.4, Experiment 2). But beginning from the second post-HN spillover region (i.e., HN+2) up until the second spillover region of the critical MC-V1, no differences were found across conditions. Note that these RT patterns are essentially the same as the RTs in the corresponding regions (W9, W10) of the first three conditions in Experiment 2. We thus attribute this consistent lack of effects to the TSV structure being fully constructed by the time our participants had completed the structural reanalysis, leading to no difficulty in parsing the TSV and SVO sentences.

In sum, results from Experiments 2 and 3 lead us to two conclusions. First, the TSV structure with an intervening complex RC is more difficult, especially at its medial part S (i.e., the NP2), to process than the canonical SVO, because Chinese comprehenders’ tendency to interpret the NP1 as the subject needs to be curbed in the extended context (owing to the complexity of RCs), leading to a prolonged, if not delayed, reanalysis downstream. As will be discussed in the General Discussion (Section 5.3), this initial subject bias fits a universal parsing heuristic. Second, the lack of integration costs at the topic-related main verbs in TSV sentences indicates that a long-distance dependency does not necessarily incur processing costs in Chinese TSVs, suggesting that object-topicalization may not involve movement of a null operator that binds an empty category it leaves behind (Xu and Langendoen 1985; Xu and Liu 1998; Yang and Wu 2015). Following Xu and Langendoen (1985: 26–27), the empty category in Chinese TSV structure is more likely to be a pronoun co-referential with an antecedent which is the sentence-initial topic, rather than a variable bound by a null operator. We thus suggest that in real-time processing of Chinese TSV structure, completing the dependency between the topic and the verb is essentially driven by semantics or pragmatics (see Section 5.2 for discussion).

5.1 Memory-based theories versus experience/surprisal-based theories

One insight we gained from three experiments is that overall, our results are difficult to reconcile with the memory-based theories of linear locality (Gibson 1998), but are consistent with the predictions of experience/surprisal-based theories (Gennari and MacDonald 2008; Levy 2008).

Memory-based theories of linear locality predict that processing difficulty increases as a function of linear distance between dependent elements. Hence, TSV sentences – with a longer argument-verb dependency – should be more difficult to process than SVOs. But we found an overall lack of processing differences at the critical retrieval/integration site of the main verb across all three experiments. Furthermore, while the argument-verb dependency is much longer in the TSV-MC than in the TSV-RC in our Experiments 2 and 3, yet again, the predicted processing costs were not observed at the main VP. Instead, no differences in RTs were found between the two types of TSV sentences. Accommodation of all these behavioral data by the integration-cost metric of the DLT is difficult. Note that non-null or anti-locality effects were found in verb-final languages such as German (Konieczny 2000; Schwab et al. 2022), Hindi (Vasishth and Lewis 2006), and Japanese (Nakatani and Gibson 2010). Our novel data from Chinese TSV structure thus expands the empirical base for the effects contrary to what locality would predict.

Our data can be accounted for by experience/surprisal-based theories, which predict that processing difficulty is driven by comprehenders’ experience with statistical regularities in linguistic input. First, participants prefer to interpret the sentence-initial noun as the subject, which conforms to the results of our corpus investigation and sentence-completion analyses, showing that while topic-comment structure – of which TSV is one specific subtype – is fairly common, SVOs are the most basic structure in Chinese. We will return to this point in Section 5.3. Second, Chinese readers’ expectations for the target TSV structure vary considerably by just one word, as revealed by gated sentence-completion data: in simple structures, participants produced more TSVs after NP2 (Li and Wu 2020). Similarly, adding a nominalizer DE greatly increased participants’ expectations of TSVs in both TSV-RC and TSV-MC conditions (Experiment 2). Clearly, guided by their ample experience with various topic-comment structures, Chinese comprehenders were able to shift their structural analysis from SVO to TSV fairly quickly upon seeing NP2-denoting modifier or NP2, such that by the time the verb was presented, they no longer had difficulties in relating its meaning with the left-fronted object-NP. Our overall results highlight the relationship of sentence production and comprehension, confirming that parsing preference reflects production biases (Gennari and MacDonald 2009; MacDonald 2013).

While our results from simple structure (Experiment 1) and from two TSVs in complex RC structures (Experiment 2) show no integration costs at the verb as predicted by the integration metric of DLT, this kind of evidence does not invalidate memory-based approaches to sentence processing in general. For instance, the cue-based retrieval model (Lewis and Vasishth 2005) is known to account for both locality and anti-locality effects. In fact, recent development of sentence processing models attempted to unify both the integration component of the memory-based approach and the expectation (or surprisal) component of the experience-based approach (e.g., Demberg 2010), by assuming a lossy (Futrell et al. 2021) or resource-rational (Hahn et al. 2022) model of memory representation. Thus, the lack of integration cost at the verb in our findings might be due to a noisy or non-veridical representation of verb transitivity (see e.g., Frazier and Rayner [1982]; Mitchell [1987]; Van Gompel and Pickering [2001] for evidence that native speakers of English failed to use verb information to avert garden-pathing). We leave it for future research.

5.2 Semantics/pragmatics-based topic-comment processing versus movement-based bound-variables processing

The overall lack of processing cost at main verbs in our experiments begs the question: Why does Chinese TSV processing appear not to rely on linear locality of memory access mechanisms? Here we propose a cognitive-functional account that relates to information organization of topic-comment structure in topic-prominent Chinese. In contrast to syntax-based formation of event structures, topic-comment construction is pragmatics-based, meaning-driven formation that directly maps human knowledge onto a linearly sequenced speech (Comrie 1981; Givón 1979). During natural communication, speakers frequently utter old information (i.e., a Topic) before introducing new information. This way of organizing speech can help listeners instantiate relevant events or schemata during incremental parsing. When new information embodied as a Comment is subsequently given, it can be efficiently processed. Indeed, a growing body of research has shown that sequential structures might be more fundamental than hierarchical structures in language comprehension and production (e.g., psycholinguistic studies: Christiansen and MacDonald [2009]; De Vries et al. [2012]; Gillespie and Pearlmutter [2011]; cognitive neuroscience: Conway and Pisoni [2008]; computational models: Fitz [2010]; Reali and Christiansen [2005], see Frank et al. [2012] for a review).

Furthermore, frequent exposure to formation of speech in terms of topic-comment over time can effectively help Chinese listeners or readers decode, store, and retrieve information during natural communication, probably even evoke anticipatory or “priming” effects upon the processing of the comment (Bransford 1979). Thus, the fronted object-NP in our TSV sentences, once recognized as the discourse topic, may be directly associated with its dependent or comment, including the verb. This process is analogous to relating an empty pronoun (i.e., the gap) to its antecedent (i.e., the sentence-initial topic) for meaning interpretation, rather than establishing a movement-based dependency between bound variables. As shown by our processing data, clearly Chinese comprehenders can complete such semantic dependency directly with ease.

In addition, typologically, pragmatics-based topic-comment structures are more common than syntax-based formation of event structures in languages around the world (Comrie 1981; Givón 1979). In child language development, a pragmatics-centered system typically precedes a syntax-centered system (Givón 1979). An emerging body of EEG research on Chinese sentence processing has also shown that semantic integration can proceed without being first licensed by syntactic structure (Yang et al. 2015; Wang et al. 2009; Zhang et al. 2010, 2013). We therefore contend that when following the “basic path” of human communication and language development, Chinese comprehenders do not necessarily need to construct movement-based bound-variables relation in real-time processing of TSV sentences as English comprehenders typically do, but instead opt to process a pragmatics-based sequential topic structure guided by their prior experiences.

5.3 Initial subject bias: a universal parsing heuristic

Our results also provide strong evidence for a universal parsing preference of subject interpretation. In all three experiments, a significant slowdown was detected at or prior to the NP2 of TSV sentences relative to canonical SVO sentences, suggesting that it is a routine practice for Chinese readers to take the NP1 as the subject initially. Only with availability of other information (e.g., animacy configuration) as the sentence unfolds will their readjustment be made accordingly.

This bias has typological underpinnings and experimental support as well. First, subject has been regarded as a basic grammatical notion in the structure of a language, with SOV and SVO being the primary language types in the world. The distribution of human languages certainly reflects how people tend to use a language in a fixed pattern (MacDonald 2013). Second, recent neurophysiological evidence has also shown that Chinese speakers strongly preferred subject interpretation for simple sentences NVN/NVAdv (Bisang et al. 2013; Wang et al. 2009). Thus, it would come as no surprise that such a strong bias for subject interpretation is also found in our experiments.

In short, the current Chinese topic processing study has provided support for a universal parsing heuristic to interpret the initial NP as the subject.