In this section, we present existing literature suggesting that, depending on the contexts, the de re reading of noun phrases embedded in intensional domains can be less acceptable or more acceptable than a de dicto interpretation.

First, we see claims that argue for a preference for the de re reading of the de re/de dicto ambiguity from works in Default Semantics (Capone, 2011; Jaszczolt, 1999, 2005, 2015, a.o.). In this framework, where a main claim is that utterance meaning is jointly determined by its compositionality, the intention of interlocutors, and their cognitive biases in communication, Jaszczolt (1997) argues that the de re reading of definite noun phrases is the default and the most salient one, because the primary objective of communication is understanding the speaker’s intention by securing the referent of the speaker’s utterance in the conversational context. Since the de re reading highlights the referential property of noun phrases and is able to select objects in the broader conversational context, it should stand out during interpretation. This default de re theory not only predicts the availability of de re interpretations but also predicts them to be even more accessible relative to de dicto during communication.

In a very different domain, namely legal studies, Anderson (2013) reports that the de re/de dicto ambiguity has been overlooked in the interpretation of legal statutes – the emphasis on de re readings of statutes and the lack of attention to the de dicto reading have even led to puzzling judicial results.2 For example, in the famous Enron Scandal in 2001, Enron’s auditor Arthur Anderson anticipated litigation and urged its employees to destroy related financial documents. Their action stopped on the day when the U.S. Securities and Exchange Commission subpoenaed records. At that time, two federal obstruction statutes applied to document destruction. The more general statute makes it a federal offense to “corruptly…endeavor to influence, obstruct, or impede the due administration of justice.”3 The more specific one prohibits “knowingly…corruptly persuad[ing] another person…with the intent to…destroy an object…[or]…impair the object’s…availability for use in an official proceeding.”4 The intentional words in these two statutes, endeavor and intent, grant each statute a de re and a de dicto interpretation, as shown in (4) and (5).

Because there was an overwhelming reliance among the judges on the de re interpretation of the two statutes and there was no sufficient evidence that suggests Arthur Anderson’s destruction was under the knowledge of a specific pending proceeding, the defendant was not charged. What is worth noting is that the ruling could have been different had the de dicto reading been picked up. Anderson relies on children’s difficulties in acquiring Theory of Mind (Apperly et al., 2010; Baron-Cohen et al., 1985; Wellman, 1992, a.o.) to back up her observation of the preference for de re over de dicto interpretations. That is, the observation that children tend to focus on reality and the broader context and find it difficult to reason about others’ minds helps explain why (even) adults have a preference for de re that emphasizes what is going on from the actual world and from the global context, compared with de dicto which emphasizes mental status. This study in law cites different sources from the Default Semantics framework, but arrives at similar conclusions on the empirical accessibility of the de re/de dicto ambiguity.

Finally, evidence for a dispreference for de re comes from experimental findings reported in Zhang and Davidson (2021). They designed an acceptability task, exemplified in (6). The evaluating context featured a protagonist Julie who falsely associated the authorship of the target poem with Nicole, but in reality, and from the speaker’s perspective, the poem was written by Elizabeth. According to a similar scenario in Romoli and Sudo (2009),5 the belief report with the noun phrase Nicole’s poem should be interpreted de dicto and the one with Elizabeth’s poem should be de re.

In the experiment of Zhang and Davidson (2021), speakers of English read four stories similar to (6). Two stories had a follow-up test sentence with a de re reading, and the other two stories had a sentence with a de dicto reading. Participants used a continuous slider bar to give their judgment. Transforming the judgment data into the sides of agreement and disagreement, the results show that around 20% of participants disagreed with both de re sentences, about 40% disagreed with one of the de re sentences, and the remaining 40% agreed with both de re sentences. In contrast, almost all of the participants showed strong agreement with the de dicto sentences. What is more noteworthy is that a sizable proportion of the disagreement concerning de re aggregated on the very edge of the slider bar. In contrast, the majority of the de dicto judgments aggregated on the edge of the agreement side. This empirical finding, especially the peculiar bimodal distribution of de re acceptability, is surprising in light of other work, such as Romoli and Sudo (2009), where both readings are predicted to be generally acceptable.

Thus, we see several different sources in the literature that seem to report greater ease or difficulty, respectively, in accessing de re readings in comparison to de dicto readings of noun phrases. This seems like an area ripe for more careful experimental consideration, and, especially, further investigation of the role that context and pragmatics may play in these judgments.

Given the reported disparities in the existing literature between de re and de dicto interpretations, more carefully controlled experimental research seems prudent; in this section we further motivate an experimental take on this topic.

For one thing, from the perspective of research methodology, the experimental investigation of de re/de dicto provides an instance that enriches the discussion on “the nature of empirical evidence in research on meaning” (Tonhauser & Matthewson, 2015, p. 1). To yield stable replicable and transparent data for theoretical development, Tonhauser and Matthewson (2015) argue that one needs to provide not only the linguistic expression under investigation, but also the context in which the expression is uttered, a response by a native speaker to a task involving the linguistic expression in the context, and information about the native speakers that provide the response. Gibson and Fedorenko (2010) also argue for using multiple items with controlled experiments to eliminate confounding factors like specific lexical properties in a single sentence and idiosyncrasies from the contexts (aspects of experimentation that are valuable even in very small scale studies, as discussed in Davidson (2020)). Existing research in semantics and pragmatics has already uncovered valuable insights by studying the influence of contexts, experimental paradigm, and/or response options on linguistic judgments, and has illustrated some of the advantages of becoming more “experimental” (see, e.g. Schwarz et al. (2007) and Jasbi et al. (2019) for scalar implicature and the influence of contexts as well as response options; see Jasbi, Bermudez, and Davidson (2023) and Jasbi, Bermudez, Zhang, et al. (2023) for logical connectives and the effect of experimental paradigm on cross-linguistic findings). Our research is another case study that advocates for rigorous experimental practice in the study of meaning.

Second, empirically investigating the de re/de dicto reading acceptability provides a case study of how formal theories of meaning that involve logic and mathematical techniques align with psycholinguistic findings that delineate the psychological representation of meaning. For example, it would be interesting to explore whether there is an analogy between (i) the semantic representation of the de re/de dicto ambiguity with the formal scope technique, compared to its empirical acceptability, and (ii) the formal quantificational scope and the related psycholinguistic processing mechanisms in classic scope interactions (see Brasoveanu & Dotlačil, 2019, for a review). We know that the inverse scope reading is sometimes hard to obtain, as in (7) (e.g. Anderson, 2004; Tunstall, 1998). What is more interesting is whether a wider existential scope in de re that is opposite to its surface position also leads to difficulty in interpretation or processing.

Last but not least, the de re/de dicto ambiguity in belief reports is one of many phenomena that can be used to investigate the cognitive relation between people and propositions (see Nelson, 2023, for a review). It tries to answer the question of how one’s belief is encoded in language, and in turn, how the ambiguity of language leads to varied and nuanced interpretations of one’s belief content. It has attracted interdisciplinary discussion from philosophy (e.g. Cohen et al., 2021; Epstein et al., 2023; Lederman, 2022; Richard, 1990), cognitive science (e.g. Apperly & Robinson, 2003; Robinson & Apperly, 2001), and even artificial intelligence (e.g. Wiebe & Rapaport, 1986; Wu et al., 2023). Understanding what contexts facilitate or impede a certain reading of the belief report could shed light on the research agenda that aims to understand the relation between language and mind. More practically, this line of research could also complement existing research about false belief tasks and Theory of Mind (e.g. Baron-Cohen et al., 1985; Wimmer & Perner, 1983) where the focus has been on the predicate of the belief content (e.g. Sally believes that the marble is in the box.) rather than the referential properties of noun phrases (e.g. the marble) inside belief reports.

In this section, we set up a finer-grained categorization of (i) the de re permitting contexts and (ii) the de re/de dicto terminology to better operationalize our experimental investigation. The aim is to collect accurate linguistic judgments for a specific linguistic expression with a clarified meaning in a well-controlled context.

First, we subdivide the de re permitting contexts into two types: the ignorance context and the misapprehension context, adopting the terminology introduced by Sudo (2014). The misapprehension context in Sudo (2014) is shown in (8) where the attitude holder is wrong about the identity of the referent (Sudo does not give an ignorance context example).

In general, the misapprehension context applies to cases where the belief holder is wrong about the identity of a referent and assigns to it a noun phrase that is true in the belief holder’s mind but false in reality. The de re context in the poetry competition example in (6) features such a misapprehension context. Judge Julie believes the poem was written by Nicole and assigns the noun phrase Nicole’s poem to it. But in reality and the broader story context, the poem was written by Elizabeth (the latter noun phrase Elizabeth’s poem is read de re). On the other hand, the ignorance context applies to cases where the belief holder is ignorant of the association between the de re noun phrase and the referred object. The context in the abstract reviewing example in (6) that permits the de re reading can be categorized as the ignorance context, since the belief holder John is not aware that the abstract was written by the addressee.

Consistent with most formal semantics literature on the de re/de dicto ambiguity, Sudo (2014) does not predict or discuss differences between these two contexts in the acceptability of the de re interpretation of definite noun phrases. Yet in Zhang and Davidson (2021), the judgment data shows that the misapprehension context prioritizes de dicto and disfavors de re. One potential hypothesis is that the misapprehension context highlights the contrast between the belief holder’s mental state and the story’s broader context. When the target sentence Julie believes that Nicole’s poem_{de dicto} will win the competition starts with Julie believes…, it highlights the belief holder’s mental state. Since the de dicto term is consistent with the content of the belief holder’s mind and, thus, is consistent with the preamble, it is prioritized over the de re noun phrase. Previous research on children’s interpretation of belief reports shows that highlighting the belief state of the protagonist in the context increases the likelihood that children attend to the belief content (Lewis et al., 2017). It is thus interesting to see whether the same mechanism is playing a role in the disproportionate judgment pattern in de re/de dicto noun phrases.

Secondly, we adopt a three-way taxonomy of the de re/de dicto ambiguity, as referential de re, referential de dicto, and attributive de dicto. This is motivated not only by the fact that ontologically two types of contexts emerge which admit the de dicto reading of definite noun phrases, but also that, from an experimental perspective, we could have better control over what context corresponds to what specific readings of the sentence. Based on previous discussion of the referential and attributive properties of definite noun phrases (Donnellan, 1966; Fodor, 1970; Jaszczolt, 1997), we introduce this tripartite taxonomy using the poetry competition example, repeated here as (9).6

We can achieve a precise and finer-grained understanding of how the distinction of de re/de dicto in definite noun phrases can be mapped to specific readings in specific contexts by integrating the two context types and the three readings of intensional definite noun phrases. The ignorance context only permits a referential de re reading of the corresponding definite noun phrase. The misapprehension context (theoretically) permits a referential de re reading of one definite noun phrase and a referential de dicto reading of another (competing) definite noun phrase. The following experiments aim to test whether the division of context types and terminologies can help address the bimodal distribution of acceptability judgments of de re reported in prior work by Zhang and Davidson (2021).

This section introduces the outline of three experiments. All three experiments adopted an acceptability task with adult native speakers of English.

Experiment 1 replicated Zhang and Davidson (2021), and found that when the context featured misapprehension and allowed a referential de re reading for one noun phrase and a referential de dicto reading for another, the referential de re reading showed a bimodal distribution of acceptability judgments.

Experiment 2 put the critical sentence from Experiment 1 under contexts that supported only an attributive de dicto reading (the control context) and contexts that supported only a referential de re reading (the ignorance context). There, we found no statistical difference in acceptability ratings between the two readings – both received high agreement. This shows that the referential de re interpretations of belief reports are acceptable in the ignorance contexts.

Experiment 3 tested the hypothesis that the misapprehension context, i.e. the co-existence of a referential de re definite noun phrase and a referential de dicto definite noun phrase in the same context, would make the former reading less acceptable. By juxtaposing the misapprehension context from Experiment 1 and the ignorance context from Experiment 2, we found evidence supporting this hypothesis.

In sum, we show that while the de re reading of definite noun phrases is undoubtedly allowed by the grammar, its acceptability is vulnerable to contextual factors relating to competition between multiple possible referential expressions. When a competing referential de dicto noun phrase appears in the same context, in other words, when the belief holder associates a wrong term with the target object, the de re reading with a different term becomes much less acceptable. We discuss the potential mechanisms and implications in Section 5.

We only analyzed the judgments from participants who passed the practice trials and whose judgments of the fillers were correct more than 75% of the time, so 51 out of 60 participants (85%) contributed their data to our final analysis.9

Figure 1 shows that in the referential de dicto condition, the majority of judgments goes to the “highly agree” side. This contrasts with the referential de re condition, where around 25% of the judgments are “highly disagree”, with the majority of choices still going to “highly agree”.

We fit the judgment data into Bayesian multilevel cumulative ordinal models, using the brms package (Bürkner, 2017, 2018) in R. The five-point judgments were the dependent variable, with non-equidistant intervals between levels on the Likert scale. They were coded from 1 to 5, where 1 indicated “highly disagree” and 5 indicated “highly agree”. The two critical condition levels were entered as a dummy-coded fixed effect (reference level = referential de dicto). The story was also entered as a fixed effect, and the interaction between the two fixed effects was also included.10 Random intercepts and slopes for the full fixed effects for the subject were included as random effects to obtain the maximal random effects required for mixed-effects models (Barr et al., 2013). The prior distributions for all the intercepts and coefficients of fixed effects were fitted to a normal distribution with the mean as 0 and the standard deviation as 2 (i.e. Normal (0,2)); the prior for the correlation matrices was set to be LKJ(2).11 The variances for the correlation matrices were set as the default in R. The priors mildly restricted the possible coefficient for each parameter, but still allowed reasonably large variance. The model had four sampling chains, each with 4000 iterations. The first 2000 samples were taken as a warmup. An R̂ close to 1.0 marks the convergence of the sampling chain to the underlying posterior distribution of the target predictor (Gelman & Rubin, 1992). The parameter setup also followed previous acceptability rating tasks in psycholinguistics (e.g. Paape et al., 2020; Zhang et al., 2023).

All R̂s for the sampling chains for all fixed effects were 1.0, indicating successful convergence. We used the package emmeans (Lenth et al., 2018) to evaluate the main effect exerted by the de re/de dicto manipulation and the judgment distinction in each story setting. Here we use β to refer to the coefficient estimate and HPD, i.e. highest posterior density, to refer to the shortest interval with the highest density in the posterior distribution of the target coefficient (Box & Tiao, 2011).

Overall, the referential de dicto condition receives more agreement than the referential de re (β = 1.66, HPD = [0.16, 3.12]). Interestingly, the agreement distribution of the two conditions varied by the story: in story A, there does not seem to be a difference (β = 1.40, HPD = [–0.16, 2.99]); there also does not seem to be a difference in story B (β = –1.30, HPD = [–3.84, 0.95]); in both story C and story D, the referential de dicto reading has a higher agreement rating than the referential de re reading (C: β = 3.48, HPD = [1.33, 5.98]; D: β = 3.10, HPD = [0.65, 5.68]). The by-item distribution can also be seen in Figure 2.

When it comes to individual differences, we see that for the referential de dicto condition, more than 50% of the participants chose “highly agree” for both trials; nobody chose “highly disagree” and only one participant chose “somewhat disagree” once. In comparison, only 15 out of 51 participants (29.4%) chose “highly agree” twice for the referential de re condition; 5 participants (9.8%) chose “highly disagree” twice. It is clear that participants showed more disagreement and chose more intermediate options for the referential de re readings.

In this experiment, we made use of contexts that should support, through two possible means of description, both the referential de dicto interpretation and the referential de re interpretation of belief reports and varied the target sentence to gauge the judgment distribution of the two readings. We found that while the referential de dicto reading was overwhelmingly agreed, the referential de re reading led to a bimodal pattern of judgments. This successfully replicated Zhang and Davidson (2021), but uncovered judgment patterns that are new with respect to other literature on this topic, such as Sudo (2014).

The by-item investigation shows that while all four stories received almost identical rating patterns for the referential de dicto reading, they exhibited different patterns for the referential de re reading. Specifically, in story B, the proportions for “highly disagree” and “somewhat disagree” were the lowest by comparison, replicating the findings in Zhang and Davidson (2021). We speculate that the unique judgment pattern for the de re sentence in story B, i.e. Mrs. Jackson believes that Grace’s gift was sent by Mike, might be related to the information structure of the passive complement clause (see Appendix Table 1 for the complete experimental details). Since passives could (i) highlight the relative newness of the information in the by phrase – it was Mike, not someone else, who sent the gift, and/or (ii) emphasize that the subject is affected by the action denoted by the verb – the gift was sent, not received (Ambridge et al., 2016; Pullum, 2014), this passive construction could potentially modulate the information flow so that readers might focus on verifying information in the predicate and ignore the subject in the belief report. Nevertheless, story D also featured a passive structure in the complement clause, i.e. Tracy believes that Alice’s spare apron needs to be washed, but there is still a significant proportion of disagreement on de re. We speculate that the by phrase in story B could play a role here, and we leave to future work the further investigation of the effect of passives as well as the information structure of the complement clause on the judgments of the referential de re reading.

The investigation of individual differences shows that while more than half of the participants had no problem accessing the referential de dicto condition, only 30% accessed the referential de re condition plus 10% going for the opposite truth-value judgment.

In sum, the bimodal distribution of the referential de re reading was replicated and observed to be systematic. Future studies are needed to disentangle the effects of linguistic information structure and individual differences on de re judgments.

We only analyzed the judgments from participants who passed the practice trials and whose judgments of the fillers were correct more than 75% of the time. As a result, 60 (90.9%) participants contributed to the analysis.

Figure 3 shows that in the attributive de dicto condition, more than 75% of the judgments fall within the “highly agree” category and the distribution is strongly skewed towards the agreement edge. In the referential de re condition, more than half of the judgments aggregate to the “highly agree” category, with a similar skewness pattern. By visual comparison, the proportion of judgments from “highly disagree” to “somewhat agree” in the de re condition is slightly larger, indicating that the de re interpretation might be less acceptable, although, in general, both readings are overwhelmingly acceptable.

We fit the judgment data into Bayesian multilevel cumulative ordinal models. The condition, the story, and their interaction were entered as fixed effects; random intercepts and random slopes for the full fixed effects structure for the subjects were entered as random effects. The prior setting, as well as all the other parameters, were the same as in Experiment 1.

The Bayesian model shows that all R̂s for the sampling chains for all fixed effects were 1.0, indicating successful convergence. There was no difference between the attributive de dicto condition and the referential de re in their agreement distribution (β = 0.825, HPD = [–0.18, 1.76]) and only in story A was there a difference in judgments of the two conditions (β = 1.31, HPD = [0.069, 2.63]).

Experiment 2 found no statistical difference between the referential de re reading and the attributive de dicto reading of the definite noun phrase in belief reports. This is consistent with prior literature and the general agreement in the field about the de re/de dicto ambiguity of noun phrases in belief reports. The comparison between Experiments 1 and 2 reflects the fact that the unacceptability of the de re reading of definite noun phrases is conditional on the context. More specifically, we know that the de re context in Experiment 1 featured the misapprehension of the belief holder, while the de re context in Experiment 2 featured the ignorance of the belief holder with respect to the de re expression.

54 out of 60 participants (90%) contributed to the crucial analysis, after the same screening procedure. Aligning with the prediction, Figure 4 shows that the bimodal distribution only appeared in the referential (de dicto + de re) condition; in the referential de re condition, the majority of the judgments accrued on the edge of agreement. This finding was also supported by Bayesian multilevel cumulative ordinal models. Here, both the dummy coded condition (reference level = (de dicto + de re)) and the story (reference level = a) as well as their interaction were entered as fixed effects; random intercepts and random slopes for the full fixed effects structure for the subject were entered as random effects. The priors and all the meta parameters were set the same as in previous experiments. All R̂ for the sampling chains for all fixed effects were 1.0, indicating successful convergence. The result shows that, overall, the de re only condition elicited more agreement than the (de dicto + de re) condition (β = 1.21, HPD = [0.15, 2.71]).

Figure 5 exhibits the by-story judgment pattern between the two conditions. We see clearly that within the context designed to support both the referential de dicto and referential de re reading, there are larger proportions of disagreement on the target sentence, compared to the de re only condition. The statistical analysis shows that in story A and story C, in the (de dicto + de re) condition, there was marginally more disagreement than in the de re only condition (story A: β = 1.06, HPD = [–0.11, 2.54]; story C: β = 1.29, HPD = [–0.19, 3.32]). In story B, there wasn’t a significant difference between the two conditions (β = 0.22, HPD = [–1.79, 2.60]). In story D, there was a statistical significance between the de re only ratings and the (de dicto + de re) ratings (β = 2.32, HPD = [0.58, 4.70]). The peculiarity of story B persisted here.

Similar to Experiment 1, we also observe a difference in the percentage of participants choosing edge judgment labels between the two conditions. In the referential de re condition, 18 out of 54 participants (33%) chose “highly agree” twice and only 1 participant chose “highly disagree” twice. In the referential (de re + de dicto) condition, only 14 (26%) stayed with the “highly agree” option twice, but 9 out of 54 participants (16.7%) went to “highly disagree” twice. This confirms that participants’ judgment behavior changed between conditions.

Experiment 3 further supports our hypothesis. The de re readings of definite noun phrases are acceptable in contexts that uniquely support a de re reading with an ignorance scenario. The acceptability of de re is affected in contexts which also permit a referential de dicto reading of a competing noun phrase with a misapprehension scenario. The item-wise differences persist in this experiment, which further supports the claim that the de re bimodal distribution arises in the face of this particular contextual manipulation.

Furthermore, there seem to be more disagreements on de re under the ignorance context in Experiment 3 than in Experiment 2 (Figure 4 vs. Figure 3).13 We speculate that a priming effect (Bock, 1986) could play a role here: while participants chose “disagreement” for the de re interpretation in the misapprehension context (i.e. the referential (de re + de dicto) context in Figure 4), they were more tempted to do so with de re in the ignorance context (i.e. the referential de re context in Figure 4). Since priming effects have been found not only in syntactic processing (e.g. Tooley & Traxler, 2010), but also in semantic and pragmatic reasoning (Bott & Chemla, 2016; Raffray & Pickering, 2010; Rees & Bott, 2018), it would be interesting to see whether and how they can be applied to the de re/de dicto context in further controlled studies.