1.1 Perspective-sensitivity with anaphors

Perspective-sensitive anaphors constitute the first ingredient relevant to this paper. Some languages (e.g. Ewe, Donno Sɔ) have specific logophoric pronouns that refer to the ‘subject of consciousness’, the person whose speech, thoughts or feelings are being reported (e.g. Hagège 1974; Clements 1975; Sells 1987; Culy 1994; Huang 2000; Pearson 2015). Although English does not have a distinct class of logophors, it is widely agreed that in some contexts, certain anaphoric forms – in particular reflexives – are not governed by the syntactic principles of classic Binding Theory and exhibit perspective sensitivity (e.g. Cantrall 1974; Kuno 1987; Zribi-Hertz 1989; Reinhart & Reuland 1993; Fludernik 1993; Buering 2005; Rooryck & Vanden Wyngaerd 2011; see also Ross 1970, inter alia). Although some researchers use the term ‘logophor’ for these English anaphors (e.g. Kuno 1987; Reinhart & Reuland 1993; Charnavel 2020 and others), I mostly use the term ‘perspective-sensitive.’ This is because the grammatical distribution and behavior of perspective-sensitive anaphors in English is not the same as logophoric pronouns in languages like Ewe.

The perspective-sensitivity observed in English relates to the well-known observation that the complementary distribution of pronouns and reflexives (e.g. Chomsky 1981) breaks down in certain contexts. For example, in possessorless picture-NPs (ex.2) – often viewed as a ‘poster-child’ for lack of complementarity – both a pronoun and a reflexive are acceptable when coindexed with the subject (e.g. Reinhart & Reuland 1993; Stirling & Huddleston 2002; Tenny 2003):

The existence of various syntactic configurations where both pronouns and reflexives are acceptable is well-established by theoretical, functionalist and corpus-based research (e.g. Lees & Klima 1963; Chomsky 1965; Jackendoff 1972; Cantrall 1974; Lebeaux 1984; Chomsky 1986; Williams 1987; Kuno 1987; Zribi-Hertz 1989; Parker, Riley & Meyer 1990; Pollard & Sag 1992; Chomsky & Lasnik 1993; Reinhart & Reuland 1993; Baker 1995; Brinton 1995; König & Siemund 2000; Stirling & Huddleston 2002; Tenny 2003). Researchers have noted that in many of these contexts, perspectival factors play a role – in particular, that a reflexive signals that the perspective being conveyed is that of the reflexive’s antecedent (rather than the first-person speaker or narrator) – i.e., the reflexive refers to the perspectival center.²

While perspectival effects are observed in various contexts, the present paper focuses on picture-NPs. This construction has the advantage of being extensively researched (e.g. Jackendoff 1972; Chomsky 1986; Williams 1987; Pollard & Sag 1992; Reinhart & Reuland 1993; Stirling & Huddleston 2002; Tenny 2003).³ The second reason for focusing on picture-NPs is that there exists prior psycholinguistic work on this construction (Kaiser et al. 2009) that provides a basis for the current study. This earlier work provides experimental evidence for divergent perspectival behaviors of pronouns and reflexives in picture-NPs (without subjective predicates). The third reason is that picture-NPs allow subjective predicates to co-occur with pronouns and reflexives in the same nominal structure (e.g. funny photograph of her/herself), i.e., in close proximity.

Let us review prior claims on the perspective-sensitivity of pronouns and reflexives in picture-NPs.⁴ As noted by authors like Kuno (1987), factors like point-of-view can influence whether a given referent can antecede a reflexive. Consider (3a,b):

Kuno notes that if the antecedent is the source of information (Mary in (3a)), it can trigger use of a reflexive in a picture-NP. However, if Mary is the perceiver of information (recipient), a reflexive sounds worse (3b). These patterns lead to the hypothesis that reflexives in picture-NPs prefer antecedents that are sources-of-information. I follow Sells (1987) and define the ‘source’ as the intentional agent of the communication. The ‘source-of-information’ hypothesis for reflexives in picture-NPs echoes the behavior of specialized logophoric pronouns that refer to the source of reported speech or thought (e.g. Clements 1975; Sells 1987).

Strikingly, prior research shows that pronouns in picture-NPs are also sensitive to perspectival information. First, consider (3c,d), from Reinhart & Reuland (1993) (see also Jackendoff 1972; Chomsky 1986, i.a.). In (3c), the pronoun’s antecedent is a perceiver of information, whereas in (3d) the pronoun refers to the source. Crucially, (3c) is reported to sound better than (3d).

Tenny (2003) calls pronouns in this context short-distance pronouns (SDPs), and notes that “verbs that provide a sentient, perceiving antecedent are especially conducive to SDPs” (Tenny 2003:14). She further states that “SDPs in representational contexts […..] are especially felicitous with perceiving subjects.” These kinds of patterns lead to the hypothesis that pronouns in picture-NPs prefer antecedents that are perceivers-of-information – the opposite of reflexives’ source preference.

The idea that perceivers also play a privileged role has a typological precedent: Some languages (e.g. Mapun, Angas) have specialized logophoric pronouns referring to the addressee, the person perceiving the information (e.g. Frajzyngier 1993). This suggests that, in addition to sources having a special status as perspectival centers, perceivers are also privileged in this regard.

In sum, both reflexives and pronouns in picture-NPs have been claimed to be sensitive to perspectival information – in opposing directions that correspond to the two perspectives inherent in communicative information-transfer events: (i) someone who conveys the information (source) and (ii) someone who receives it (perceiver): Reflexives are sensitive to the source, the individual whose speech/thoughts are being expressed, and pronouns to the perceiver, the one receiving the information.

1.2 Psycholinguistic experiments on perspective-sensitive anaphors in picture-NPs

To assess the judgements reported in the theoretical literature and to gain insights into real-time processing, Kaiser et al. (2009) conducted experiments to test the hypotheses that reflexives in picture-NPs (without PPTs) prefer sources of information while pronouns prefer perceivers. In this earlier work, we used sentences like (4), where the verb (told vs. heard from) was used to manipulate whether the subject or object is the source or perceiver. The sentences did not contain adjectives. We used both offline questionnaires and visual-world eye-tracking, which provides a fine-grained and time-sensitive measure. The results confirm that reflexives and pronouns in picture-NPs are sensitive to the source and perceiver-of-information respectively, and that this can be detected early during real-time processing.

We also found that pronouns and reflexives differ in how sensitive they are to structural information (subject/object) vs. the source/perceiver manipulation: Reflexives in picture-NPs are guided by a strong structural subject preference and a weaker but (still significant) source-of-information preference, whereas pronouns are guided by a structural object preference and a strong perceiver-of-information preference.⁵ Similar patterns arise in Dutch and German (Kaiser & Runner 2008; see Kaiser & Do 2012 for further data on English).

These results confirm intuitions discussed in prior work. Although English picture-NP anaphors are not directly comparable to logophoric pronouns in languages like Ewe or long-distance reflexives in languages like Japanese, these studies confirm they are sensitive to the source/perceiver distinction. These experiments also provide a more gradient picture of the relative impact of structural and perspectival information, by showing that the effects of grammatical role and source/perceiver status on pronouns and reflexives are not categorical, but nevertheless have a robust effect on reference resolution.

1.3 On the relation between perspective-sensitive anaphoric expressions and subjective expressions

Research on perspective-sensitive anaphors has also yielded claims that interpretation of these anaphors is linked to interpretation of other perspectival expressions (e.g. Sells 1987; Kuno 1987; Charnavel 2020; 2021). The core observation is that multiple perspectival expressions can exhibit what I call perspectival uniformity – i.e., be interpreted relative to the same perspectival center. ⁶ In particular, many have noted that whoever is the antecedent of the perspective-sensitive anaphor is also the attitude-holder of other perspectival expressions (broadly construed) in the same sentence.⁷ Here I focus solely on contexts with perspective-sensitive anaphors. (For research on the behavior of multiple non-anaphoric perspectival expressions, see e.g. Bylinina & McCready & Sudo 2014; 2015; see also Kneer & Vicente & Zeman 2017 for a different view. See Anand & Korotkova 2021 for an analysis of sentences with multiple PPTs, especially in attitude reports, see also Saebø 2009; Pearson 2013 for related points.)

In his influential 1987 paper, Sells discusses Japanese examples like (5) and points out (using different terminology) that the attitude-holder of the adjective beloved – what he calls an ‘expressive term’⁸ – is interpreted as the same referent as the antecedent of the reflexive zibun ‘self.’ These examples provide evidence of perspectival uniformity. (I follow Sells in translating zibun as ‘him’ here.)

Sells notes that (5b) also allows a reading where Taroo is the antecedent of the reflexive, and that in this case, beloved reflects Taroo’s perspective. This further suggests that zibun and beloved exhibit perspectival uniformity. He also points out that “English shows the same properties as Japanese: beloved links with whichever referent is understood as the antecedent of the pronoun” (p.465).

Additional examples illustrating perspectivally-uniform behavior between anaphors and subjective expressions, from Japanese, Korean, Swedish and English (Oshima 2004; Giorgi 2006; Park 2018; Hellberg 1980; Strahan 2001; Gast 2004) are in Supplementary File 1. Research on multiple languages suggests that perspective-sensitive reflexives and subjective expressions exhibit perspectival uniformity, i.e., are interpreted relative to the same perspectival center.

Evidence specifically for picture-NPs and predicates of personal taste – the ingredients that this paper focuses on – comes from Charnavel’s (2020) work on French. She notes that in (6), the attitude-holder of the PPT (e.g. horrible) is whoever the emphatic reflexive in the picture-NP (elle-même/lui-même ‘herself/himself’) refers to. Charnavel also discusses adjectives like étrange ‘strange’ and ignoble ‘despicable, horrible’ in related structures such as ‘the media’s horrible remarks about herself,’ where she notes that the referent of herself is whoever is the attitude-holder of horrible (see her ex.109).

2.1 Possible outcomes

As regards the interpretation of reflexives and pronouns, the expectation is that Experiments 1, 2 and 3 will replicate prior work showing that (a) reflexives in picture-NPs show a preference for the subject, modulated by a weaker preference for the source-of-information, while (b) pronouns show a preference for the object as well as a strong preference for the perceiver-of-information (Kaiser et al. 2009).

What about the question of who is interpreted as the attitude-holder (judge) of PPTs in sentences like (1a,b)? It seems reasonable to posit that the individual who is the source of information is the most plausible attitude-holder. The source can be inferred to possess relevant knowledge/information/experience of the photograph that allows them to judge it to be funny, frightening etc.¹⁴ Indeed, a source preference has been suggested by Kuno (1987) and Sells (1987). According to Sells, this is because what he calls “speaker-evaluative” expressions are anchored to the individual who is the intentional agent of the communication (p.462–462), and in structures like (1), expressing communicative events with tell/hear, that is the source-of-information. The idea that the source tends to be interpreted as the attitude-holder of a PPT has not been experimentally tested. Indeed, the first-person speaker could also be construed as the attitude-holder (e.g. Lasersohn 2005).

What about the relation between the antecedent of a perspective-sensitive anaphor and the attitude-holder of a PPT? In the structures tested in Experiment 1, the PPT modifies the noun photograph, and the pronouns and reflexives are inside a prepositional phrase modifying that noun (see also (6) from Charnavel 2020). For now, I make the simplifying assumption that the PPT and the anaphor are inside the same perspectival domain. I mostly focus on the following four outcomes (though these are not the only possibilities):

Option 1: Strong version of perspectival uniformity. If PPTs and picture-NP pronouns and reflexives in structures like (9) are subject to perspectival uniformity, (i) the attitude-holder of the PPT and (ii) the antecedent of the pronoun or reflexive should coincide. This is what’s expected if a high, general perspective-shifting operator binds all perspective-sensitive elements in its scope to the same perspectival center. However, in light of prior work showing complementary source/perceiver biases for reflexives and pronouns, this kind of strong, across-the-board approach already faces challenges. This leads to the second prediction:

Option 2: Form-specific perspectival uniformity. It could be that perspectival uniformity applies selectively to some perspective-sensitive anaphors but not others. In particular, if the source is preferred as a PPT’s attitude-holder, it might be that perspectival uniformity holds between reflexives (also sensitive to sources) and PPTs, but not between pronouns (sensitive to perceivers) and PPTs. Such form-specific perspectival uniformity predicts that the antecedent of reflexives in picture-NPs and the attitude-holder of PPTs will be categorically yoked together, whereas the antecedent of pronouns and the attitude-holder of PPTs diverge. In operator-based accounts, this pattern could arise if the perspective-shifting operator binds reflexives and PPT in its scope, but not pronouns.

Option 3: Overlapping preferences create an appearance of partial perspectival uniformity. Another possibility is that PPT attitude-holder identification and the interpretation of pronouns and reflexives are not intrinsically linked, but that these processes are guided by potentially overlapping constraints and thus may show similarities in certain contexts. This prediction is based on Kaiser et al. (2009), who propose that the interpretation of pronouns and reflexives is guided by multiple weighted constraints – specifically, that reflexives are guided by a heavily-weighted subjecthood constraint and a weaker source constraint, whereas pronoun interpretation is guided by fairly evenly weighted perceiver and object constraints. Under this view, if PPT attitude-holder identification is an independent process guided by a source constraint, then in certain contexts we may see what looks like a trend towards perspectival uniformity but is actually only a coincidence of the relevant factors happening to overlap. For example, if reflexives prefer subjects and PPTs prefer sources, an illusion of perspectival uniformity will emerge where the subject is the source (with told) – a pattern that is predicted to vanish when the subject is the perceiver (with heard).

Option 4: No relation. A logical possibility is that the interpretation of PPTs and perspective-sensitive pronouns/reflexives are not only distinct from each other, but also not guided by any related constraints. In this case, no relation is expected to hold between (a) the antecedents of pronouns and reflexives and (b) the attitude-holders of PPTs.

These possible outcomes are summarized below:

Overview of predictions regarding perspectival uniformity

3.2 Experiment 1 Results

Statistical analysis In all experiments reported here, data was analyzed using R (R Core Team 2013). To analyze participants’ answers to the ‘who shown’ questions I fit logistic mixed effects regression models (glmer, using lme4 1.1-20, Bates et al. 2015) to the data. Subject responses were coded as 1, object responses as 0. Models contained pronoun type (pronoun vs. reflexive), verb (told vs. heard) and their interaction as fixed effects. As random effects, all models had intercepts for subjects and items whenever possible,¹⁷ as well as by-subject and by-item random slopes for the fixed effects when justified by model comparison. (For each model, I started with the maximal random effect structure for subjects and items, and used model comparison to identify the maximal random effect structure justified by the design and supported by the data. Only random effects that contributed significantly to the model (p < 0.05) were included, Baayen et al. 2008.) All experiments are analyzed in the same way.

To analyze answers to the ‘whose opinion’ questions, glmer models were fitted for each possible answer choice: one model for subject-opinion responses, where subject opinion responses are coded as 1, and all other responses are coded as 0; another model for object-opinion responses, where object-opinion responses are coded as 1 and all other responses are coded as 0, and a third model for narrator-opinion responses, where narrator-opinion responses are coded as 1 and all other responses as 0. ‘Condition’ was entered as a fixed effect into the model. Random effects were determined as described above. The emmeans package (emmeans 1.5.0, Lenth 2018) was used to obtain Bonferroni-corrected pairwise comparisons. ‘Whose opinion’ data for all three experiments reported in this paper is analyzed the same way. (In Experiments 2 and 3, in addition to analyzing subject-opinion, object-opinion and narrator-opinion choices, the combination responses possible in those experiments, e.g. subject+narrator opinion, were also analyzed.)

3.2.1 Reference resolution

Participants’ responses to the ‘who shown’ question that probes interpretation of pronouns and reflexives are in Figures 2a and 2b. Figure 2a shows the proportion of trials where participants selected the preceding subject as the antecedent. Figure 2b shows the proportion of trials where participants selected the preceding object as the antecedent. (The corresponding bars in Figure 2a and 2b sum to 1). As the figures show, reflexives elicit mostly subject responses, whereas pronouns are more split between subjects and objects. (The no-anaphor condition is not shown.) Furthermore, the two verbs clearly do not pattern alike: There is a higher rate of subject interpretations with reflexives than with pronouns (β = 1.722, SE = 0.155, z = –11.109, p < 0.0001) and with heard than with told (β = 0.427, SE = 0. 0.151, z = 2.821, p < 0.005), but these effects are modulated by a significant interaction (β = 1.636, SE = 0.304, z = 5.383, p < 0.0001). See Supplementary File 3 for statistical models. More specifically, pronouns are more sensitive to the verb manipulation than reflexives. Indeed, planned comparisons show that pronouns elicit more subject interpretations with heard than told (β = 1.7897, SE = 0.558, z = 3.206, p < 0.002) – exhibiting the predicted perceiver preference. Reflexives exhibit a numerical source preference, in line with prior work, but this effect is only marginally significant (β = –0.368, SE = 0.223, z = –1.652, p = 0.0985). In light of Kaiser et al.’s finding that pronouns prefer perceivers more strongly than reflexives prefer sources, the absence of a strong verb effect with reflexives in the current offline task is not unexpected.

Figure 2a

Experiment 1: Proportion of trials where participants interpreted the anaphor as referring to the subject (Error bars show +/– 1 SE). Throughout the paper, SEs were calculated according to Cousineau (2005) and corrected according to Morey (2008).

Figure 2b

Experiment 1: Proportion of trials where participants interpreted the anaphor as referring to the object (This data is the inverse of the subject response data shown in Figure 2a, error bars show +/– 1 SE).

Overall, the ‘who shown’ data replicates prior work. In sentences like Nora {told/heard from} Kimberly about the frightening photograph {of herself/her}, participants tend to interpret (i) a reflexive as referring to the subject, especially when the subject is the source-of-information, i.e., when the sentence contains the verb told, and (ii) participants tend to interpret a pronoun as referring to whichever argument is the perceiver of information (the subject of heard and object of told).

3.2.2 Attitude-holder identification

Participants’ responses to the ‘whose opinion’ question, which probes attitude-holder identification, are in Figure 3. All three anaphor-type conditions (pronoun, reflexive, no-anaphor) elicit significantly more subject-opinion responses with told (subject=source) than with heard. (See Table 1 for statistics). Conversely, all three anaphor-type conditions elicit significantly more object-opinion responses with heard (where the object is the source) than with told (see Table 2).

Figure 3

Experiment 1: Proportion of trials where, for each condition, participants responded that the PPT is the opinion of the subject, object or the narrator (attitude-holder identification). (Error bars: +/–1 SE).

Table 1

Experiment 1: “Subject’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests, shading indicates p < .05 here and elsewhere).

contrast	estimate	SE	z	p value
Told_none vs. heard_none	4.985	0.339	14.721	<.0001
Told_pro vs. heard_pro	4.497	40.316	14.217	<.0001
Told_ref vs heard_ref	4.412	0.331	13.312	<.0001
Heard_none vs. heard_pro	–0.379	0.284	–1.334	1.0000
Heard_none vs heard_ref	–1.003	0.271	–3.703	0.0019
Heard_pro vs heard_ref	–0.625	0.255	–2.449	0.1289
Told_none vs. told_pro	0.109	0.303	0.358	1.0000
Told_none vs. told_ref	–0.430	0.330	–1.303	1.0000
Told_pro vs told_ref	–0.539	0.327	–1.650	0.8909

Table 2

Experiment 1: “Object’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
Told_none vs. heard_none	–6.350	0.550	–11.538	<.0001
Told_pro vs. heard_pro	–5.441	0.435	–12.505	<.0001
Told_ref vs heard_ref	5.322	0.459	–11.586	<.0001
Heard_none vs. heard_pro	0.124	0.249	0.498	1.0000
Heard_none vs heard_ref	0.502	0.245	2.050	0.3635
Heard_pro vs heard_ref	0.378	0.242	1.560	1.0000
Told_none vs. told_pro	–0.785	0.581	–1.350	1.0000
Told_none vs. told_ref	–0.526	0.603	–0.874	1.0000
Told_pro vs told_ref	0.259	0.510	0.507	1.0000

The proportion of narrator-opinion responses does not differ depending on the verb (see Table 3) and is quite low. This may seem surprising, since PPTs are often regarded as having a default bias for the speaker (here the narrator) to be the attitude-holder, but prior work on Free Indirect Discourse and related phenomena (Kaiser 2015) indicates that in narrative contexts, it is easy to shift from the speaker (narrator) to a character as the perspectival center.

Table 3

Experiment 1: “Narrator’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
Told_none vs. heard_none	–0.995	0.408	–2.437	0.1334
Told_pro vs. heard_pro	–0.834	0.420	–1.984	0.4249
Told_ref vs heard_ref	–1.146	0.476	–2.410	0.1434
Heard_none vs. heard_pro	0.444	0.391	1.135	1.0000
Heard_none vs heard_ref	0.985	0.407	2.420	0.1397
Heard_pro vs heard_ref	0.541	0.411	1.317	1.0000
Told_none vs. told_pro	0.282	0.429	0.656	1.0000
Told_none vs. told_ref	1.136	0.475	2.392	0.1507
Told_pro vs told_ref	0.854	00.482	1.772	0.6874

In sum, there is a preference to interpret the source as the attitude-holder of the PPT, which is independent of (i) whether the sentence contains an anaphor or not, and (ii) whether that anaphor is a pronoun or a reflexive. PPT interpretation is guided by a strong source preference.

Furthermore, as is apparent in Figure 3, this source preference is stronger when the source is the subject (with told) than when the source is the object (with heard). This difference is significant in all three anaphor-type conditions (no anaphor, reflexive, pronoun; Table 4).¹⁸ This suggests that in addition to a strong source preference, the attitude-holder identification process is guided by a weaker subject preference. This fits with the well-known finding that subjects are more prominent/more topical than objects (e.g. Chafe 1994; Crawley et al. 1990, inter alia).

Table 4

Experiment 1: “Source opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 3 tests).

contrast	estimate	SE	z	p value
Told_none vs. heard_none	1.726	0.299	5.769	<.0001
Told_pro vs. heard_pro	1.693	0.292	5.796	<.0001
Told_ref vs heard_ref	2.688	0.326	8.237	<.0001

Overall, these results show that the identification of PPT attitude-holders is guided by a strong source constraint and a weaker subjecthood constraint.

3.2.3. Combining the data types

So far, the reference resolution data (‘who shown’ questions, Figures 2a, b) corroborates earlier results showing that reflexives have a strong subject bias modulated by a source bias, while pronouns have a strong object bias and a strong perceiver bias. But the attitude-holder identification data (‘whose opinion’ questions, Figure 3) shows no effect of anaphoric form, and instead identification of PPT attitude-holders is guided by a strong source bias and a weaker subject bias.

Put together, these data challenge claims about strong perspectival uniformity. To see this more clearly, let’s visually combine the two data types. Figures 4a and 4b combine the data from Figures 2a, 2b and 3, to show how choice of antecedent for the anaphoric expression relates to choice of PPT judge on each trial (with the no-anaphor condition omitted). In Figure 4, the height of each bar shows the proportion of trials where a certain anaphor was interpreted as referring to the preceding subject (Figure 4a) or the preceding object (Figure 4b). Thus, the heights of the bars in Figures 4a, b are the same as the heights of the bars in Figures 2a, b. For ease of exposition, the proportion of subject responses and object responses are plotted separately (like Figures 2a, b), and sum to 1. Crucially, in Figures 4a and 4b, the segments within each bar show what proportion of responses were associated with the PPT being interpreted as the subject’s, object’s or the narrator’s opinion. Thus, the segments inside each bar convey the information from Figure 3, now relativized to the antecedent-choice proportions.

Figure 4a

Experiment 1: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the subject?

Figure 4b

Experiment 1: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the object?

According to the strong version of perspectival uniformity, (a) all subject-antecedent interpretations (Figure 4a) should also be subject-opinion PPT interpretations (all subject-antecedent choice bars should only contain black) and (b) all object-antecedent interpretations (Figure 4b) should be object-opinion PPT interpretations (all object-antecedent choice bars should only contain white). However, this is clearly not the case for subject-antecedent or object-antecedent interpretations, and it is not the case for reflexives or pronouns. Instead, Figures 4a and 4b confirm data patterns already discussed, namely a preference to interpret the source of information as the attitude-holder: Regardless of anaphoric form, told conditions elicit high proportions of subject opinion responses (black shading; the subject is the source) and heard conditions elicit high proportions of object opinion responses (white shading; the object is the source).

Could this data be taken as evidence for form-specific perspectival uniformity – specifically, that only reflexives and PPTs pattern together, but that pronouns and PPTs diverge? The answer is no. Even if one only looks at the reflexive conditions, there are clear signs of reflexives and PPTs diverging. For example, in the heard condition, most trials where participants said the reflexive refers to the subject are still interpreted as having an object attitude-holder (third bar in Figure 4a). Furthermore, even on the trials when participants interpreted the reflexive as referring to the subject in the told condition (the source; fourth bar in Figure 4a), there is no perfect correlation between antecedent choice and attitude-holder identification, as participants sometimes report that the object (perceiver) or the narrator is the attitude-holder of the PPT.

3.3 Experiment 1 Discussion

The antecedent-choice patterns for pronouns and reflexives in Experiment 1 largely replicate earlier findings (Kaiser et al. 2009) showing that pronouns and reflexives in picture-NPs are sensitive to the distinction between source and perceiver, but in different directions and not equally strongly. The attitude-holder identification data in Experiment 1 shows a strong source bias and a weaker subject bias. The source bias is in line with suggestions in prior work (e.g. Sells 1987).

Crucially, this preference to interpret the source of information as the attitude-holder obtains regardless of whether the sentence contains a pronoun or a reflexive, and regardless of whether that pronoun or reflexive is interpreted as referring to the subject or object. There is no evidence for (i) a strong perspectival uniformity constraint that would yoke together the attitude-holder of PPTs and the antecedent of perspective-sensitive pronouns and/or reflexives, nor (ii) for a form-specific perspectival uniformity constraint that only applies to reflexives.

As a whole, Experiment 1 provides no clear evidence that the interpretation of PPTs and anaphors in contexts such as ex.(10) exhibit perspectival uniformity. Instead, the data seem to be more compatible with the multiple-constraints, form-specific account discussed in Kaiser et al. (2009). Under this view, anaphor resolution and attitude-holder identification are guided by multiple weighted constraints. Reflexives are guided by a strong subject preference and a weaker source preference, pronouns by a perceiver preference and an object preference, and PPT attitude-holder interpretation is guided by a strong source preference and a weak subject preference. Thus, the preferences of reflexives and PPTs happen to resemble each other in the right configuration (when the subject is the source, with told) – creating what might look like a preference for form-specific perspectival uniformity. But even this is revealed as an illusion when we consider configurations where the preferences of reflexives and PPTs come apart, i.e., where the object is the source (with heard from): it’s clear that, like pronouns, reflexives in picture-NPs are not inherently linked to PPTs’ attitude-holders.

In sum, these results indicate that interpretation of picture-NP anaphors and identification of the attitude-holders of PPTs are not yoked together, although they may show some overlap in contexts where the constraints guiding these processes happen to partially converge

4.2 Predictions

For antecedent choice data and attitude-holder identification data, the predictions are the same as Experiment 1. The key question is whether, on trials where the anaphor is interpreted as referring to the perceiver, is the perceiver now also selected as an attitude-holder (perhaps in addition to the source referent)? Data along these lines would provide stronger evidence for perspectival uniformity than suggested by the results of Experiment 1.

4.3 Experiment 2 results

4.3.1 Reference resolution

Figures 5a and 5b show participants’ responses to the ‘who shown’ questions. There are more subject interpretations with reflexives than pronouns (β = –2.618, SE = 0.217, z = –12.075, p < 0.0001), as well as a significant interaction between anaphor and verb (β = 3.43, SE = 0.431, z = 7.955, p < 0.00001), with pronouns showing greater sensitivity to the verb manipulation than reflexives. There is no main effect of verb (β = –0.076, SE = 0.396, z = –0.192, p = 0.848). As is visible in Figures 5a and 5b – and echoing Experiment 1 – pronouns are more sensitive to the verb manipulation than reflexives. Indeed, planned comparisons reveal that pronouns elicit more subject interpretations with heard than told (β = 1.47, SE = 0.211, z = 6.95, p < 0.0001), replicating Kaiser et al. (2009)’s findings as well as the results of Experiment 1. Reflexives exhibit a numerical source preference, although it does not reach significance (β = –1.007, SE = 0.722, z = –1.393, p = 0.164). Nevertheless, this reflexive pattern goes in exactly the same direction that was observed in earlier work and in Experiment 1. As a whole, the referential patterns in Experiment 2 largely replicate prior work.

Figure 5a

Experiment 2: Proportion of trials where participants interpreted the anaphor as referring to the subject. (Error bars show +/– 1 SE.)

Figure 5b

Experiment 2: Proportion of trials where participants interpreted the anaphor as referring to the object (This data is the inverse of the subject response data in Figure 5a). (Error bars: +/– 1 SE).

4.3.2 Attitude-holder identification

Participants’ responses to the ‘whose opinion’ question are in Figure 6. As in Experiment 1, all three anaphor-type conditions (pronoun, reflexive, and no-anaphor) elicit (i) more subject-only opinion responses with told than heard and (ii) more object-only opinion responses with heard than told (Tables 5 and 6). Furthermore, although combined “subject + narrator” responses and “object + narrator” responses are relatively infrequent (never higher than 20% in any condition), they echo the same pattern as “subject-only” and “object-only” responses: there are more “subject + narrator” responses with told than heard, and more “object + narrator” responses with heard than told, with all three anaphoric form options (no anaphor, pronoun, reflexive). This is in Tables 8 and 9. “Narrator-only” opinion responses are infrequent (never above 10% in any condition) and, as in Experiment 1, are not influenced by the verb manipulation (see Table 7). Thus, narrator responses on their own exhibit no sensitivity to the source/perceiver distinction, in line with Experiment 1, but participants are willing to consider shared perspectives (albeit not very frequently) such that the narrator shares the perspective of the source-of-information referent. This could be taken as evidence that the narrator’s perspective is a neutral default that can ‘come along’ when a character’s perspective is being assumed.

Figure 6

Experiment 2: Proportion of trials where, for each condition, participants responded that the PPT is the opinion of the subject, the object or the narrator (attitude-holder identification). (Error bars: +/– 1 SE).

Table 5

Experiment 2: “Subject’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	4.000	0.304	13.153	<.0001
told_pro vs. heard_pro	3.727	0.284	13.128	<.0001
told_ref vs heard_ref	3.699	0.285	12.977	<.0001
heard_none vs. heard_pro	–0.487	0.296	–1.646	0.8978
heard_none vs heard_ref	–0.821	0.287	–2.861	0.0380
heard_pro vs heard_ref	–0.334	0.264	–1.265	1.0000
told_none vs. told_pro	–0.213	0.247	–0.863	1.0000
told_none vs. told_ref	–0.520	0.255	–2.035	0.3763
told_pro vs told_ref	–0.307	0.258	–1.190	1.0000

Table 6

Experiment 2: “Object’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–4.254	0.520	–8.188	<.0001
told_pro vs. heard_pro	–3.444	0.404	–8.534	<.0001
told_ref vs heard_ref	–5.173	1.010	–5.122	<.0001
heard_none vs. heard_pro	0.238	0.178	1.336	1.0000
heard_none vs heard_ref	0.480	0.180	2.670	0.0682
heard_pro vs heard_ref	0.242	00.180	1.344	1.0000
told_none vs. told_pro	–0.572	0.633	–0.903	1.0000
told_none vs. told_ref	1.398	1.121	1.247	1.0000
told_pro vs told_ref	1.970	1.073	1.837	0.5962

Table 7

Experiment 2: “Narrator’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–1.269	0.588	–2.160	0.2769
told_pro vs. heard_pro	–1.714	0.700	–2.448	0.1292
told_ref vs heard_ref	-0.911	0.688	–1.323	1.0000
heard_none vs. heard_pro	0.551	0.530	1.041	1.0000
heard_none vs heard_ref	1.069	0.570	1.876	0.5464
heard_pro vs heard_ref	0.518	0.592	0.875	1.0000
told_none vs. told_pro	0.995	0.722	1.378	1.0000
told_none vs. told_ref	0.711	0.697	1.020	1.0000
told_pro vs told_ref	–0.285	0.756	–0.376	1.0000

Table 8

Experiment 2: “subject and narrator opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	3.019	0.472	6.395	<.0001
told_pro vs. heard_pro	2.237	0.440	5.083	<.0001
told_ref vs heard_ref	2.165	0.442	4.894	<.0001
heard_none vs. heard_pro	–0.2801	0.511	–0.548	1.0000
heard_none vs heard_ref	–0.293	0.515	–0.569	1.0000
heard_pro vs heard_ref	–0.013	0.494	–0.026	1.0000
told_none vs. told_pro	0.502	0.325	1.545	1.0000
told_none vs. told_ref	0.561	0.328	1.708	0.7882
told_pro vs told_ref	0.059	0.334	0.175	1.0000

Table 9

Experiment 2: “object and narrator opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–4.596	1.063	–4.322	0.0001
told_pro vs. heard_pro	–3.588	0.787	–4.561	<.0001
told_ref vs heard_ref	–4.899	1.066	–4.595	<.0001
heard_none vs. heard_pro	0.286	0.339	0.843	1.0000
heard_none vs heard_ref	–0.306	0.325	–0.940	1.0000
heard_pro vs heard_ref	–0.591	0.334	–1.770	0.6910
told_none vs. told_pro	–0.722	1.252	–0.577	1.0000
told_none vs. told_ref	–0.003	1.440	–0.002	1.0000
told_pro vs told_ref	0.72	1.253	0.574	1.0000

The results also show that combined “subject + object” opinions are quite rare (always below 10%) but, interestingly, are more frequent with heard than with told, in all three anaphoric form conditions (Table 10). In other words, regardless of whether the sentence contains no anaphor, a pronoun or a reflexive, participants are more likely (although it’s still quite rare) to judge the PPT as reflecting the opinion of both the subject and the object when the verb is heard than told. All three candidates (subject, object, narrator) were rarely selected; the data is too sparse to analyze statistically.

Table 10

Experiment 2: “subject and object opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–1.86	0.532	–3.502	0.0042
told_pro vs. heard_pro	–1.61	0.429	–3.763	0.0015
told_ref vs heard_ref	–2.34	0.523	–4.484	0.0001
heard_none vs. heard_pro	–0.534	0.348	–1.536	1.0000
heard_none vs heard_ref	–0.481	0.349	–1.377	1.0000
heard_pro vs heard_ref	0.0533	0.326	0.163	1.0000
told_none vs. told_pro	–0.784	0.574	–1.365	1.0000
told_none vs. told_ref	–0.0000292	0.633	0.000	1.0000
told_pro vs told_ref	0.784	0.574	1.365	1.0000

Overall, the ‘whose opinion’ results replicate Experiment 1, namely that the single source of information (the subject with told, the object with heard) tends to be interpreted as the attitude-holder of the subjective predicate, even if multiple options could be chosen. This preference for the source is present in the no-anaphor condition, the pronoun condition and the reflexive condition (Tables 5 and 6).

As in Experiment 1, there is evidence of subjecthood effects on attitude-holder interpretation: The source preference is again stronger when the source is the subject (with told) than when it is the object (with heard), in in all three anaphor type conditions (no anaphor, reflexive, pronoun), see Table 11.²⁰

Table 11

Experiment 2: “Source opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 3 tests).

contrast	estimate	SE	z	p value
Told_none vs. heard_none	1.199	0.242	4.946	<.0001
Told_pro vs. heard_pro	1.833	0.250	7.322	<.0001
Told_ref vs heard_ref	2.558	0.264	9.685	<.0001

Overall, the results again show that PPT attitude-holder identification is guided by a strong source preference and a weaker subject preference.

4.3.3 Combining the data types

To assess the relationship between perspective-sensitive anaphora and attitude-holder identification more directly, let’s again visually combine the two data types analyzed above. Figures 7a and 7b replot the same data from Figures 5a, 5b and 6. Akin to Experiment 1, there is a dissociation between reference resolution and attitude-holder identification. (As a reminder, perspectival uniformity predicts all subject-antecedent interpretations will be subject-opinion PPT interpretations – all bars should be fully black in 7a – and all object-antecedent interpretations will be object-opinion PPT interpretations – all bars should be fully white in 7b.) This is not the case in either Figure 7a or 7b. Instead, what emerges is the same kind of dissociation between reference resolution and attitude-holder identification as in Experiment 1.

Figure 7a

Experiment 2: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the subject?

Figure 7b

Experiment 2: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the object?

The dissociation is especially clear on those trials where the anaphor refers to the perceiver (subject of heard, first and third bars in Figure 7a, object of told, second and fourth bars in Figure 7b). On these trials, there is no sign of a corresponding bias to interpret the perceiver (subject of heard, object of told) as the judge of the PPT. Instead, there is again a preference to interpret the source as the attitude-holder. Even on trials where the anaphor refers to the source (subject of told, second and fourth bars in Figure 7a, object of heard, first and third bars in Figure 7b), there is no evidence for strong perspectival uniformity.

But now let’s consider if there is any evidence from participants’ multiple-opinion responses to support the prediction that, on trials where the anaphor is interpreted as referring to the perceiver (subject of heard, first and third bars in Figure 7a; object of told, second and fourth bars in Figure 7b), the perceiver is selected as an additional attitude-holder, in addition to the source of information. In other words, is there an interaction between verb type and antecedent choice? Are there more “subject + object” responses when the anaphor refers to the perceiver (subject of heard, first and third bars in Figure 7a; object of told, second and fourth bars in Figure 7b) than when it refers to the source (subject of told, second and fourth bars in Figure 7a, object of heard, first and third bars in Figure 7b)? The answer is no. Although there are more “subject + object” opinion responses in Figure 7a with heard than with told, this is not the case in Figure 7b. In fact, as above, “subject + object” opinions are overall more frequent with heard than with told, in all three anaphoric form conditions.

Thus, there is no evidence of an interaction, no evidence of participants preferentially selecting the perceiver as an additional attitude-holder alongside the source when the anaphor refers to the perceiver.

4.4 Experiment 2 discussion

The results of Experiment 2 largely replicate Experiment 1. The reference resolution data show that reflexives have a strong subject preference and a numerical source preference, while pronouns have a strong object preference and a perceiver preference. The attitude-holder identification data reveal that participants still mostly opt for only a single attitude-holder: Averaging across conditions, participants chose only one attitude-holder on 73% of all trials. The results again show that PPTs have a strong source preference and a weak subject preference.

I also checked whether, on trials where the anaphor is interpreted as referring to the perceiver, participants selected the perceiver as a secondary attitude-holder, in addition to the source. The results show that this is not the case.

As a whole, these results fit with the multiple-constraints approach proposed for Experiment 1, where anaphor resolution and attitude-holder identification are guided by multiple constraints, which happen to partially resemble each other in the case of reflexives (weak source preference, strong subject preference) and PPTs (strong source preference, weak subject preference) but not with pronouns (perceiver preference and object preference).

In sum, Experiment 2 replicates the results of Experiment 1: both support the ‘partially overlapping constraints’ situation sketched out in Section 2.1. This shows that Experiment 1’s conclusions about a lack of perspectival uniformity between attitude-holder identification and interpretation of perspective-sensitive anaphora is not an artifact of the single-answer format.

One might wonder about a follow-up where participants can indicate multiple potential antecedents for anaphors. There are at least two reasons why this is unlikely to provide evidence for perspectival uniformity. First, given that offering participants the possibility of selecting multiple PPT attitude-holders did not yield support for perspectival uniformity, it is logically unlikely that a multiple-antecedent study would do so: If perspectival uniformity holds, it should have been detected in Experiment 2. Second, singular anaphors in English can be ambiguous but cannot refer to multiple referents at once – unlike plural them/themselves, and presumably unlike PPTs that could (at least intuitively) reflect shared opinions. If two antecedents are competing for selection, the expectation is that when the data is aggregated, one would see at least weak reflexes of this competition. However, this was not the case. Thus, the results suggest that a ‘multiple-antecedent’ follow-up study would be unlikely to provide clear evidence for perspectival uniformity.

5.2 Predictions

In terms of attitude-holder identification, the prediction is again that the source is the preferred attitude-holder; there is no reason to expect a change in syntactic configuration to change this.

In terms of reference resolution, if presence of an intervening tensed clause boundary has no effect, the ‘who shown’ data for Experiment 3 is predicted to pattern like the results from Kaiser et al (2009) and Experiments 1–2. However, if truly perspective-sensitive anaphors only occur when they are in a separate finite clause from their antecedents, Experiment 3 will show how such elements behave in response to the verb-based source/perceiver manipulation. Relatedly, if the reason for the lack of perspectival uniformity in Experiments 1–2 was that the anaphors being tested were not sufficiently perspective-sensitive, evidence for perspectival uniformity may arise in Experiment 3.

5.3 Experiment 3 results

5.3.1 Reference resolution

In participants’ answers to the ‘who shown’ question (Figures 8a and 8b), the proportion of subject choices reveal main effects of anaphor type (β = –0.822, SE = 0.15, z = –5.494, p < 0.0001), verb (β = –1.772, SE = 0.151, z = –11.727, p < 0.0001) and a verb x anaphor interaction (β = –1.78949, SE = 0.299, z = –5.980, p < 0.0001). The interaction reflects the fact that pronouns are again more sensitive to the verb manipulation than reflexives – echoing Experiments 1–2. Planned comparisons show that pronouns again exhibit a significant perceiver preference: more subject choices with heard than with told (β = –2.648, SE = 0.23, z = –11.501, p < 0.0001). The source preference for reflexives that was previously found by Kaiser et al. and that was marginal in Experiment 1, is now significant: more subject choices with told than with heard (β = –1.42, SE = 0.633, z = –2.244, p < 0.03).

Figure 8a

Experiment 3: Proportion of trials where participants interpreted the anaphor as referring to the subject. (Error bars: +/– 1 SE).

Figure 8b

Experiment 3: Proportion of trials where participants interpreted the anaphor as referring to the object (This data is the inverse of the subject response data in Figure 8a). (Error bars: +/– 1 SE).

5.3.2 Attitude-holder identification

Participants’ responses to the ‘whose opinion’ question are in Figure 9. Echoing Experiments 1–2, all three anaphor-type conditions (pronoun, reflexive, and no-anaphor) elicit (i) more subject opinion responses with told than with heard and (ii) more object opinion responses with heard than with told (see Tables 12 and 13). As is clear from Figure 9, the rate of narrator opinion responses is very low and shows no sign of being influenced by the verb manipulation (no statistical analyses could be conducted due to scarcity of data). The rates of ‘multiple attitude-holder’ opinion responses are also low.

Figure 9

Experiment 3: Proportion of trials where, for each condition, participants responded that the PPT is the opinion of the subject, the object or the narrator (attitude-holder identification).(Error bars: +/– 1 SE).

Table 12

Experiment 3: “Subject’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	8.083	0.588	13.746	<.0001
told_pro vs. heard_pro	7.929	0.579	13.691	<.0001
told_ref vs heard_ref	7.081	70.538	13.152	<.0001
heard_none vs. heard_pro	0.220	0.471	0.468	1.0000
heard_none vs heard_ref	–1.220	0.394	–3.097	0.0176
heard_pro vs heard_ref	–1.440	0.417	–3.453	0.0050
told_none vs. told_pro	0.374	0.435	0.860	1.0000
told_none vs. told_ref	–0.218	0.468	–0.466	1.0000
told_pro vs told_ref	–0.592	0.451	–1.314	1.0000

Table 13

Experiment 3: “Object’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 9 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–8.659	0.777	–11.146	<.0001
told_pro vs. heard_pro	–7.432	0.595	–12.493	<.0001
told_ref vs heard_ref	–6.547	0.549	–11.919	<.0001
heard_none vs. heard_pro	0.230	0.317	0.724	1.0000
heard_none vs heard_ref	1.059	0.300	3.531	0.0037
heard_pro vs heard_ref	0.829	0.290	2.862	0.0379
told_none vs. told_pro	–0.996	0.764	–1.304	1.0000
told_none vs. told_ref	–1.053	0.760	–1.386	1.0000
told_pro vs told_ref	–0.057	0.605	–0.094	1.0000

Overall, then, Experiment 3 replicates the first two experiments: there is a strong preference to interpret the source of information (the subject with told, the object with heard) as the attitude-holder of the PPT. Thus, this preference extends to the cross-clausal context.

As in the two earlier studies, attitude-holder identification is also guided by a subject preference: there is a stronger source preference in source = subject conditions (with told) than source = object conditions (with heard), regardless of anaphor type (Table 14).²³

Table 14

Experiment 3: “Source’s opinion” responses, pairwise comparisons (P-value adjustment: bonferroni method for 3 tests).

contrast	estimate	SE	z	p value
told_none vs. heard_none	–8.659	0.777	–11.146	<.0001
told_pro vs. heard_pro	–7.432	0.595	–12.493	<.0001
told_ref vs heard_ref	–6.547	0.549	–11.919	<.0001

5.3.3 Combining the data types

Figures 10a and 10b visually combine the two data types analyzed above, to show how choice of antecedent relates to PPT attitude-holder identification. The results replicate the outcomes of the first two experiments.

Figure 10a

Experiment 3: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the subject?

Figure 10b

Experiment 3: Whose opinion does the PPT reflect on trials where participants said that the anaphor refers to the object?

5.4 Experiment 3 discussion

In essence, Experiment 3 replicates Experiments 1–2: ‘Who shown’ responses indicate that reflexives have a subject preference modulated by a source preference, while pronouns have an object preference and a perceiver preference. The fact that the referential biases are similar to Experiments 1–2, and also to prior work by Kaiser et al. (2009) – all of which used a structure with only one tensed clause – shows that the lack of perspectival uniformity cannot be attributed to reflexives not being ‘sufficiently’ perspective-sensitive in structures like (10).

Granted, reflexives’ sensitivity to the source manipulation is stronger in Experiment 3 than in Experiments 1–2. However, the fact that Experiments 1–2 show a numerical and/or marginal source effect in the exact same direction, and given that Kaiser et al. (2009) obtained significant evidence for a source preference in structures like ex.(10) in several experiments, there is no basis to conclude that the interpretative processes in contexts like ex.(10) (one tensed clause) and ex.(11) (two tensed clauses) are qualitatively different.

In Experiment 3, responses to the ‘whose opinion’ questions also replicate Experiments 1–2: Identification of the attitude-holder of the PPT has a strong source preference and a weaker subject preference. Thus, the results fit with the two preceding studies.

As a whole, Experiment 3 is line with Experiments 1–2 and provides no evidence that anaphora in a two-clause structure exhibit perspectival uniformity with PPTs. The results again support the idea that identification of PPTs’ attitude-holders and interpretation of pronouns and reflexives are not linked in an intrinsic way, but that these processes are guided by partially overlapping constraints.

6.1 Overview of the results within a multiple-constraints framework

In a nutshell, the results show that attitude-holder identification with PPTs and reference resolution with picture-NP anaphors, although sensitive to some of the same factors, do not exhibit inherent perspectival uniformity.

As regards anaphor resolution, the results corroborate earlier findings that reflexives and pronouns in picture-NPs show divergent and asymmetric sensitivities to source and perceiver (and grammatical role) and further show that these patterns extend across finite clauses. Interpretation of reflexives is guided by a strong subject preference and a weaker source preference, whereas interpretation of pronouns is guided by an object preference and a perceiver preference. The patterns are not categorical, and instead point to a system with multiple interacting constraints. Thus, these results replicate and extend the findings of Kaiser et al. (2009) who propose, building on the form-specific account from Kaiser & Trueswell (2008), that pronouns and reflexives are guided by multiple weighted constraints (see also Kaiser 2003; Brown-Schmidt & Byron & Tanenhaus 2005 on form-specific effects in reference resolution; Cunnings & Sturt 2014; 2018 for relevant discussion; Ariel 1990; Arnold 1998 for earlier work on a multiple-constraints approach to reference resolution; McRae & Matsuki 2013 for a general overview of constraint-based approaches to sentence processing).

When it comes to PPTs: all three experiments reveal a strong preference to interpret the source of information as the attitude-holder of the PPT, regardless of whether an anaphor (pronoun or reflexive) is present, in a way that is also modulated by subjecthood.²⁴ Thus, identifying the attitude-holder of PPTs is guided by a strong source preference and a weaker subject preference. This is the first experimental evidence for a strong source preference for PPTs, and corroborates suggestions by Kuno (1987) and Sells (1987).

The existence of these constraints – subject preference, object preference, perceiver preference, source preference – is independently motivated by existing work, including prior research on Binding Theory, pronoun resolution, PPT semantics, and source-oriented vs. addressee-oriented logophoric pronouns.

Crucially, the present results yield no evidence for a categorical perspectival-uniformity principle that yokes together the attitude-holder of PPTs and the antecedent of perspective-sensitive pronouns or reflexives, nor for a more restricted form-specific constraint that only applies to reflexives (or to pronouns, for that matter). But it would be inaccurate to conclude that these processes are entirely unrelated.

Instead, the data point to a situation where reflexives and PPTs are sensitive (but to different degrees) to some of the same constraints (source, subject status), while pronouns are sensitive to complementary constraints (perceiver, object status). Consequently, although we see perspectival uniformity tendencies with PPTs and reflexives in some contexts, a closer look at other configurations shows that these patterns are dissociable and not intrinsically yoked together.

In particular, in contexts where the subject is the source (with told, ex.12a), reference resolution of reflexives and attitude-holder identification of PPTs seem to show a tendency towards perspectival uniformity: Lisa is the preferred antecedent of the reflexive (the subject) and the preferred judge of the PPT (the source). Nevertheless, even here the correspondence is not categorical. This suggests that the notion of source across these two phenomena (reference resolution, attitude-holder identification) is independently processed.

Furthermore, when subjecthood and source status are dissociated (with heard, ex.12b), there is no perspectival uniformity: Kate is the preferred PPT attitude-holder whereas Lisa is the preferred antecedent of the reflexive. Once subjecthood and source status are disentangled, it becomes clear that reflexives and PPTs do not exhibit systematic perspectival uniformity. This point is even clearer with pronouns (12c,d) which prefer the perceiver (subject or object) while the PPT is associated with the source.

6.2 Challenges for uniform perspective-shifting accounts

Although these results fit well within the form-specific multiple-constraints architecture proposed in prior psycholinguistic work, they seem to challenge approaches where perspective shifting is accomplished categorically by a single high-level operator. This kind of approach is very schematically illustrated in (13a). (Following Charnavel’s (2020) work on French, the representations in (13) use a covert logophoric pronoun pro_log – generated by each perspective-shifting operator – that refers to the perspectival center, ensuring that perspective-sensitive elements in the scope of the operator are anchored to the relevant perspectival center.) In this configuration, one would presumably expect all relevant perspective-sensitive elements to be categorically interpreted relative to the same perspectival center, which is not what the current experiments found for English.

What about positing two smaller perspectival domains, as in (13b)? Once we consider what the perspectival center associated with the second OP_log would be, it becomes less clear how this approach could capture the asymmetric behavior of pronouns and reflexives in English picture-NPs: Not only are they sensitive to different properties of the antecedent (subject vs. object, source vs. perceiver), but also to different extents (pronouns are more sensitive to perceiver status than reflexives to source status). Thus, if one wants to keep all three pieces (PPTs, reflexives and pronouns) in the class of perspective-sensitive elements, it seems necessary to acknowledge that, at least in English, they are not equally sensitive to a uniform conceptualization of ‘perspectival center’.²⁵ On this kind of account, even if one were to allow configurations with multiple perspectival domains, it would presumably also be necessary to, first, allow different kinds of perspective-sensitive operators or different kinds of covert logophoric pronouns, in order to derive both differences in level of sensitivity as well as the complementary-but-not-mirror-image behavior of pronouns and reflexives, and second, to constrain the system to prevent overgeneration.

6.3 Broader implications

As a whole, the results reported here regarding (a) reference resolution of picture-NP anaphors and (b) attitude-holder identification of PPTs seem to fit better with a gradient, multiple-constraints approach along the lines of what has been proposed for reference resolution in prior work. This kind of approach can acknowledge the (partial) similarities between these two phenomena in terms of the constraints at play (e.g. sensitivity to source-of-information), without ignoring their differences. Of course, open questions remain for this approach as well, including questions of overgeneration (i.e., the need to constrain the possible set of constraints). In addition, further work using other methods (e.g. visual-world eye-tracking) is welcome. The present work aims to provide an initial foundation that can be built on with other methods.

Importantly, the finding that the relationship between picture-NP anaphors and PPTs in English is not subject to a categorical principle requiring all perspectival elements to be anchored to the same perspectival center does not mean that other languages could not have grammaticalized such a pattern. Indeed, it is often the case that something that is grammaticalized in one language crops up as a tendency in another language (see e.g. Heine & Narrog 2011). Thus, experimental work on other languages is an important future direction, as is work on other kinds of perspectival elements in English. The present results do not mean that perspectival uniformity effects could not occur elsewhere.

Nevertheless, the observed lack of perspectival uniformity exhibited by picture-NP anaphors on the one hand, and PPTs on the other hand, provides insights into the division of labor that holds between syntactic and non-syntactic factors. Whereas resolution of English reflexives is highly sensitive to a syntactic constraint (subjecthood) and pronoun resolution also shows clear sensitivity to a syntactic constraint (objecthood), PPT attitude-holder identification is guided mostly by source status, which is probably best viewed as a semantic/pragmatic constraint. Although reflexives and pronouns in picture-NPs are perspective-sensitive (as shown by multiple experiments), these perspectival effects are at the mercy of the stronger syntactic constraints. Indeed, the observation that syntactic factors are more fundamental to the interpretation of anaphoric expressions and semantic/pragmatic factors are more central for the interpretation of adjectives seems eminently reasonable, given everything else that is known about the grammatical properties of anaphors vs. adjectives.

This observation brings us back to the importance of crosslinguistic work, as it raises questions about perspective-sensitive anaphora in other languages. Japanese zibun (self), for example, is known to exhibit extensive perspective-sensitivity and, though it is a reflexive anaphor, is subject to somewhat different syntactic constraints than English reflexives (e.g. Sells 1987). We should be careful to not lump all reflexive elements together – to not make predictions purely based on an expression’s linguistic class – when thinking about perspectival uniformity effects and perspective-shifting. Further crosslinguistic work is needed to clarify these issues.