Null pronoun interpretation probed via thematic role ambiguity: a case in Korean

2.1 Thematic roles and topic status

Sentence (1) is ambiguous between ‘Cengmi must be worried’ and ‘Cengmi must be worrisome’, with Cengmi as an experiencer of worry or the stimulus. The ambiguity arises from the so-called double-nominative construction. Although Korean two-place psych-predicates take a dative experiencer and a nominative stimulus in the canonical structure seen in (2), they can also form a double-nominative construction, as in (3), in which case the word order determines the thematic roles: the experiencer always precedes the stimulus, differentiating (3a) from (3b) in meaning. With one of the two arguments dropped, sentence (1) is ambiguous, and the context clarifies the thematic role of the sole overt NP. Previous studies (Ahn 2017, 2018) reported that the overt NP in sentences like (1) was interpreted as an experiencer when topics shifted but as a stimulus when the same topic continued. That is, a significant dependence between topic status and thematic role was observed.^[1]

Consider (4) and (5). In both (4) and (5), the two entities Cengmi and Yuna appear in the same order in the respective context sentences; however, the thematic roles they get to take in the critical sentences differ; Yuna is the experiencer in (4) but the stimulus in (5). As the experiencer always precedes the stimulus in the double-nominative construction, Yuna will always take the first NP position in (4) but Cengmi in (5), irrespective of its overtness.

Through this combination of the double-nominative construction and an NP drop, segmentally identical sentences as in (4a) and (5b) and in (4b) and (5a) are interpreted differently: Yuna is experiencer in (4a) but stimulus in (5b), and Cengmi is stimulus in (4b) but experiencer in (5a).

Previous studies (Ahn 2017, 2018, 2024) reported that speakers prefer cases where the same overt NPs are not repeated as the first NP across two adjacent sentences: (4a) and (5b) were either judged more acceptable or used more frequently over (4b) and (5a). Ahn (2018) asked participants to create a story using an ambiguous sentence as in (1) at the beginning, middle, and end of the story (see Figure 1), which manipulated the topic status. The study found that participants tended to create a story with the overt NP in (1) as the experiencer when the given construction had to be placed at the beginning, which required a new topic. When the construction had to be placed in the middle or at the end, where the same topic tended to continue, the experiencer reading decreased while the stimulus reading increased. That is, segmentally identical NPs were assigned different thematic roles depending on topic status.

Figure 1:

Illustration of three story-completion conditions in Ahn (2018). (a) Beginning (b) middle (c) end.

A caveat in interpreting the results of the study was that the relationship between topic status and thematic role was not dichotomous nor was it the case that middle and end conditions always resulted in topic continuation. Even in middle and end conditions, topics shifted, in which case the stimulus reading was extremely rare. In general, participants tended to interpret the sole NP of the ambiguous construction (1) as the experiencer, even in the middle and end conditions. But the ratio of the experiencer reading significantly decreased in the middle and end conditions, where topic continuation was more likely.

One might wonder if the ambiguous construction is more commonly used with the overt NP as the experiencer in naturally occurring texts; however, a corpus analysis (Ahn 2024) shows that the overt NP in such a construction is significantly more likely to be a stimulus. The overall experiencer reading preference could be an artifact of a controlled study, where participants were given the critical sentence out of context as in Figure 1. Yet it demonstrates that, out of context, the construction’s default interpretation skews toward the experiencer.

The observed interaction between topic status and thematic role was accounted for in terms of the repeated name penalty (RNP) in the centering theory framework (Gordon et al. 1993; Grosz et al. 1995). The next section briefly introduces the concept and discusses whether it can best account for the phenomenon at hand.

2.2 The repeated name penalty

In the centering theory framework (Gordon and Chan 1995; Gordon and Scearce 1995; Gordon et al. 1993; Grosz et al. 1995), if two consecutive sentences have the same discourse center or discourse topic, the second sentence has the topical entity pronominalized. The topic of (6a) is Bruno, and the same topic is continued in (6b). Via a self-paced reading task, Gordon et al. (1993) showed that reading times increased when the continued topic was repeated as a definite description (Bruno in this case): hence, the repeated name penalty.

Two assumptions are made for the RNP to explain the interaction between topic status and thematic role in (4) and (5). One is that unstressed English pronouns are equivalent to null pronouns in Korean, and the other is that the topic of the sentence is the entity that the sentence is about. With such assumptions, the topic is Yuna in both (4a) and (4b), but Cengmi in (5a) and (5b), irrespective of its overtness. When the topic shifts from Cengmi to Yuna in (4), the new discourse topic should be referred to as a definite description, which explains why (4a) is preferred over (4b). Also, when the same topic (Cengmi) continues in (5), the continued topic should be pronominalized to avoid the repetition of the same NP, which means, in Korean, it should be a null pronoun. This accounts for the preference of (5b) over (5a).

However, the RNP in the centering theory framework focuses mainly on the comprehension of an overt pronoun, and its main argument is based on the assumption that seeing or hearing an overt pronoun provides a cue for listeners and readers to expect the existing discourse topic to continue. In contrast, no overt pronouns can provide such a cue in the Korean construction at issue here. Only when a definite NP is repeated as the topic can the RNP be applied.

The RNP predicts processing difficulty when the same NP is repeated in two consecutive sentences, that is, a main effect of NP repetition. This main effect is also an interaction effect between topic status and thematic role in that NPs are not repeated in the combination between topic shift and experiencer (4a) and that between topic continuation and stimulus (5b) while NPs are repeated in the other cases: (4b) and (5a). Crucially, since the RNP discusses the penalty for the repetition of the topical NP, the effect of repetition should appear when the overt NP is repeated; that is, at or right after the overt NP of the critical sentence. Earlier RNP studies (Gordon and Chan 1995; Gordon and Scearce 1995; Gordon et al. 1993; Yang et al. 1999) and later studies that tested RNP on null pronoun languages (Gelormini-Lezama 2020; Gelormini-Lezama and Almor 2011; Shoji et al. 2017) report summed reading times of several regions. From these studies that used phrase- or sentence-level RTs, it is difficult to tell if the effect of NP repetition is immediate or not. However, Eilers et al. (2019) report time-locked reactions to the (non-)repetition of NPs in German via an eye-tracking-while-reading task; the effects were observed on the anaphor or right after. Therefore, a word-by-word self-paced reading experiment is called for to investigate whether speakers’ sensitivity to NP repetition can be detected in real time.

Another question arises as to whether the RNP can explain the interdependence between topic status and thematic role described above. In (6), both forms (He and Bruno) refer to the same entity and take the same thematic role (the agent of chased). However, the Korean case compares overt NPs referring to different thematic roles. A separate measure is required to probe the interpretation of the overt NP. According to Ferreira and Yang (2019), comprehension questions are not simple tools to ensure that participants pay attention to the content of the test items. They can indicate whether participants misinterpret the sentences and whether the misinterpretations are revised or linger. The current study attempts to probe which thematic role is assigned to the overt NP in the critical sentence by asking who performed an action to stimulate the psychological state.

A comprehension question as in (7) can probe whether a reader can identify the stimulus in the critical sentence of (4) and (5). Measuring the accuracy of the answers to such comprehension questions will shed light on the disambiguation of the ambiguity in (1).

2.3 Research questions

Three key observations have been highlighted so far that motivate the current study. First, two-place psych-predicates with double-nominative constructions introduce thematic role ambiguity between an experiencer and a stimulus when one of the required arguments is dropped. Prior work suggests that context – particularly whether a discourse topic is shifting or continuing – interacts with this ambiguity. Second, the RNP offers one explanation by focusing on the cost of repeating a topical NP, but it is not obvious that it generalizes straightforwardly to Korean, where discourse can hinge on null elements. Third, the overt NP of the ambiguous construction (1) tends to be interpreted as the experiencer outside the context.

Given these previous findings, the present study directly examines how speakers resolve the thematic role ambiguity of psych-predicate constructions by topic status. By collecting word-by-word reading times and comprehension data, we aim to (i) test whether RNP effects emerge when the same NP is repeated in adjacent sentences and (ii) determine whether thematic roles are properly assigned to an overt NP, which will indicate the thematic role of the missing element. If the RNP alone can explain the relationship between topic status and thematic role in (4) and (5), we expect the effect of repetition (as in the form of interaction between topic status and thematic role) at or right after the overt NP region. Furthermore, if the experiencer reading bias observed in previous studies was an experimental artifact, the effect of thematic role should not be observed in the current study, in which the critical sentence is always contextualized.

3.1 Participants

Eighty native Korean speakers (64 female) were recruited from a major university community in Seoul for monetary compensation. Participation was limited to those who had not used languages other than Korean with their parents or caregivers and had not lived outside Korea for more than six consecutive months before age 7.

3.2 Stimuli

Experimental items were designed to have a context sentence and a critical sentence as in (4) and (5). Each context sentence introduced two referents of the same gender to prevent any gender biases. Each item was counter-balanced in a 2 × 2 Latin-square design so no participant would see the same item through different conditions. After each critical item was a comprehension question asking who stimulated the psychological state. A complete list of critical items is given in the Appendix. Forty fillers, without psych-predicates, had two sentences as in the experimental items and were followed by a comprehension question asking about the main idea or details of the sentence contents.

3.3 Procedure

After a screening survey, participants were invited into a Zoom session and were asked to turn their cameras and microphones on to ensure they were participating in the experiment in a quiet space with no disruptions. The experiment was conducted via the Penn Controller for IBEX platform (Zehr and Schwarz 2018). Since the platform downloads all experimental stimuli into a user’s local temporary folder before the experiment begins, no delays or disturbances were expected due to varying internet connection. Each item was presented word by word as participants pressed the space bar, and the comprehension questions appeared after the context and critical sentence screen was cleared. The entire session of instructions, consent form signing, practice trials, and experiment trials took approximately 45 min.

3.4 Analysis

Raw reading times (RTs) per character were first transformed to logarithms to normalize skewed data; then, residuals were calculated in a linear mixed-effects model that factored in word length, word position, and trial order, all of which were centered and scaled, with participant intercepts as random effects.^[2] Log RT residuals that exceeded three standard deviations below or above the mean were filtered out, which resulted in 2.04 % of data loss. All statistical analyses were conducted in R version 4.4.2 (R Core Team 2024) using lme4 version 1.1.31 (Bates et al. 2015) and lmerTest version 3.1.3 (Kuznetsova et al. 2017). In linear mixed-effects modeling for RT data analyses, fixed factors were centered; thematic roles (experiencer as −0.5 vs. stimulus as 0.5) and topic status (continue −0.5 vs. shift 0.5). Only participant and item intercepts were included for random effects. The maximal random-effects structure resulted in singularity issues, different combinations of slopes were compared, and more complex models did not significantly differ from the intercept-only model (Barr et al. 2013; Matuschek et al. 2017). Since RTs in self-paced reading are heavily affected by the RTs of previous words, log RT residuals of two immediately preceding words were factored in (Hofmeister 2011).^[3] Responses to comprehension questions were analyzed using logistic mixed-effects regression with topic status and thematic role as fixed factors and participant and item for random effects.^[4]

4.1 Reading times

Figure 2 shows the residuals of the log-transformed RT data in four regions: the overt NP, the psych-predicate, and two additional regions after the psych-predicate. The black and gray lines indicate the experiencer and stimulus readings of the overt NP, respectively, and the filled and unfilled circles indicate the shift and continuation of the topics. The solid and dashed lines indicate the results of combining the two factors. If the overt NP is the experiencer in the topic shift (+SHIFT) condition (4a) or the stimulus in the topic continuation (−SHIFT) condition (5b), the NP is not repeated, indicated by the solid lines. On the other hand, if the overt NP is the stimulus when the topics shift (+SHIFT) (4b) or the experiencer when the topic continues (−SHIFT) (5a), the NP is repeated, indicated by the dashed lines. The error bars indicate 95 % confidence intervals. The subject region with the overt NP of thematic role ambiguity is the region right after the last word of the context sentence. The next region is the critical region where the psych-predicate is given, which, in turn, is followed by spillover and post-spillover regions.

Figure 2:

Log RT residuals by region.

The RNP predicted an effect of overt NP repetition at or right after the overt NP, which should have appeared in the form of interaction between topic status and thematic role, but it was not observed in the overt NP region nor in the psych-predicate region. Instead, an effect of topic status was observed in the overt NP region. Log-transformed RT residuals showed a significant increase in the topic shift condition (β = 0.042, SE = 0.014, t = 2.938, p = 0.003). The significance remained the same with or without the previous word analysis in all RT analyses (see Table 1 for complete statistics).

An interaction effect of topic status and thematic role was observed in the spillover region, the region right after the psych-predicate (β = 0.106, SE = 0.027, t = 3.973, p < 0.001).^[5] Table 2 summarizes fixed effects coefficients and relevant statistics for log RT residuals in the spillover region. A pairwise comparison analysis was performed, and a significant difference was observed between experiencer NPs and stimulus NPs in the topic shift condition, that is, between (4a) and (4b) (β = −0.053, SE = 0.019, t = −2.759, p = 0.012). In the topic continuation condition, experiencer NPs (5a) were read significantly longer than stimulus NPs (5b) (β = 0.053, SE = 0.019, t = 2.773, p = 0.012). Experiencer NPs were read significantly faster in the topic shift condition (4a) than in the topic continuation condition (5a) (β = 0.070, SE = 0.019, t = 3.631, p = 0.002). Stimulus NPs were read marginally faster when the topic continued (5b) than when it shifted (4b) (β = −0.037, SE = 0.019, t = −1.906, p = 0.085).

4.2 Comprehension accuracy

The comprehension accuracy data is summarized in Table 3. Comprehenders showed a higher accuracy when the overt NP was an experiencer than when it was a stimulus. Also, the decrease in the accuracy ratio from the experiencer reading to the stimulus reading was greater in the topic shift condition than in the topic continuation condition.

Figure 3 visualizes the accuracy ratio by topic status and thematic role, and Table 4 summarizes the fixed effect coefficients and relevant statistics of a logistic mixed-effects model. A significant main effect of thematic role (β = −1.612, SE = 0.169, t = −9.530, p < 0.0001) and a significant interaction effect of topic status and thematic role (β = −0.704, SE = 0.324, t = −2.175, p = 0.030) were observed. Also, a pairwise comparison analysis was conducted and significant differences were found in all pairs of conditions (p < 0.001) except between topic shift and continuation when the overt NP was experiencer. That is, the difference between the two highest bars in Figure 3 was insignificant (β = 0.806, SE = 0.222, z = −0.782, p = 0.434).

Figure 3:

Comprehension question accuracy by topic status and thematic role.

4.3 Experiencer bias

Despite contextualization, comprehension accuracy data indicated a significant bias toward experiencer reading. To probe whether the experiencer bias also influenced RTs, the RT difference between the psych-predicate and spillover regions was analyzed. The overt NP can be assigned a thematic role only when the psych-predicate is presented. If participants are biased to interpret the overt NP as the experiencer in the critical region, the interpretation will suit the context only in (4a), in which an experiencer NP was intended in the topic shift condition. In the three other conditions, experiencer reading will not make sense, and RTs will increase.

Figure 4 shows that the RT difference between the psych-predicate and spillover regions in the topic shift/experiencer condition (4a) nears zero, while the three other conditions show significant differences. The RT difference between the psych-predicate and spillover regions was modeled with topic status and thematic role as fixed effects and participants and items as random effects, and the model resulted in a main effect of topic status (β = 0.052, SE = 0.018, t = 2.819, p = 0.005) and an interaction effect of topic status and thematic role (β = −0.131, SE = 0.036, t = −3.654, p = 0.0003). Pairwise comparisons also showed significant differences between the topic shift/experiencer condition and the three other conditions (p < 0.05). Table 5 summarizes the fixed effects coefficients of RT differences between the psych-predicate and spillover regions.

Figure 4:

RT difference between the critical and spillover region.