1 Referent accessibility in ambiguous pronoun resolution

Two interlocutors that want to establish a discourse model of which they both have a similar understanding must engage in cooperative referent management. The speaker introduces new referents into the discourse by means of proper names, demonstratives, or definite or indefinite noun phrases (NPs). These discourse referents can be rementioned using anaphoric expressions such as pronouns or definite NPs. The choice of an anaphoric expression depends on the accessibility of the antecedent. More prominent antecedents tend to be picked up by attenuated expressions, such as personal pronouns. Upon hearing a pronoun, a listener will therefore try to resolve it to the most prominent referent.

In our study, we focus on sentences with a subject and a second argument, typically a direct object as in (1), and a second sentence that contains a subject pronoun that is ambiguous between the subject (Peter) and the object (a waiter) of the first sentence.

We assume that the degree to which a subsequent ambiguous pronoun is resolved to one of the antecedents mirrors the prominence of that referent. We define prominence as a relation between two or more discourse referents (see von Heusinger & Schumacher 2019). Accessibility, on the other hand, is a static property of one discourse referent with respect to its potential anaphoric expressions in the sense of Ariel (1990) and Gundel et al. (1993), reflected by the lexical size of an anaphoric expression. In our experiment with two competing referents and one ambiguous pronoun, both notions are closely correlated, as the pronoun is linked to the more prominent or more accessible antecedent (Crawley & Stevenson 1990; Stevenson et al. 1994; Bosch et al. 2007; Bouma & Hopp 2007; Kaiser & Trueswell 2008; Arnold 2010, Kaiser 2011; Kurczek et al. 2013; Rohde & Kehler 2014, von Heusinger and Schumacher 2019).

There are several factors that determine the prominence of referents and, as a consequence, the processing and resolution of ambiguous pronouns (e.g. Chafe 1976; Hobbs 1979; Givón 1983; Ariel 1990, Gundel et al. 1993, Almor 1999, Kehler et al. 2008). Two of the main factors guiding the disambiguation of pronouns are grammatical role and the position of the argument in the sentence. Often these two factors align, as subjects and arguments that are first in the linear order of referents in a sentence are the most accessible (Givón 1983; Gernsbacher & Hargreaves 1988; Gordon et al. 1993; Gundel et al. 1993; Portele & Bader 2016). According to these factors, the pronoun in (1) would preferably refer to the subject Peter, which is also the first argument in that sentence. In addition to these more structural properties, the accessibility of referents is also influenced by semantic factors, including thematic role (Stevenson et al. 1994; Ferretti et al. 2009; Fukumura & van Gompel 2010; Schumacher et al. 2015), topichood (Kaiser 2011), focus (Kaiser 2011), aspect (Kehler et al. 2008; Kehler & Rohde 2013), verb semantics (Stevenson et al. 1994; Arnold 2001; Kehler et al. 2008; Fukumura & van Gompel 2010; Hartshorne & Snedeker 2013), and prosodic prominence (Baumann & Schumacher 2020).

Two additional factors that might influence accessibility but that are rarely investigated are informativity¹ and information status. In a series of production and comprehension studies, Karimi, Fukumura, Ferreira and Pickering (2014), Karimi and Ferreira (2016) and Karimi, Diaz, and Ferreira (2019) investigated the effect of informativity, operationalized as the length of noun phrases, on accessibility. Karimi and Ferreira (2016) conducted a visual-world eye-tracking experiment, in which participants listened to two-sentence auditory stimuli, see Table 1, while being presented with a visual display that contained pictures of two human referents and one unrelated inanimate object. The first sentence of the auditory input had two human referents (NP1 and NP2); the second sentence started with an ambiguous pronoun. The informativity of the NP1 and the NP2 from sentence 1 was manipulated to create three conditions, see Table 1 (a)–(c).

Fixation times were measured in three windows starting at the onset of the pronoun (W1: 0–1000 ms, W2: 1000–2000 ms, W3: 2000–3000 ms), and the fixation time of NP1 was compared to that of NP2. In the baseline short-short condition, there were significantly more looks to NP1 only in W3. In the long-short condition, there were significantly more looks to (long) NP1 in W2 and W3. Finally, in the short-long condition, there were significantly more looks to (short) NP1 in W1, but more looks to (long) NP2 in W3. Karimi and Ferreira (2016) suggest that the higher proportion of looks to NP1 in W1 in the short-long condition might be caused by the complexity of the sentences: People confronted with such a complex sentence might initially rely on structural factors, i.e. on referring to the subject or the first referent. Overall, however, their experiment shows that informativity contributes to the prominence and thus to a higher accessibility of discourse referents. This result is in line with findings from Karimi et al. (2014) and Karimi et al. (2019): More informative, i.e., longer and more complex, noun phrases appear more accessible to pronouns.

The second factor investigated in this paper is information status. We operationalize the familiarity of noun phrases in three categories, namely given, inferred, and brand-new, cf. (Prince 1981b: 237). This three-way distinction, as opposed to a two-way distinction between given and new, helps accommodate the difference between coreferential and bridged definites (a car … the car vs. a car … the steering wheel) on the one hand and inferred indefinites and brand-new indefinites (a bar … a waiter vs. an exhibition hall… a waiter) on the other hand. Strube & Hahn (1999) expand Centering Theory (Grosz et al. 1995) with this 3-way information status categorization in such a way that given referents outrank inferred ones and inferred ones outrank brand-new ones. Using an English and German corpus, they show that their enriched Centering algorithm performs slightly better than the original algorithm.

In a series of eye-tracking experiments, Brocher and von Heusinger (2018) and von Heusinger and Brocher (2019) investigated the effect of information status on the accessibility of indefinite noun phrases in object position in transitive sentences, comparing inferred and brand-new indefinites.² Von Heusinger and Brocher (2019) compared the accessibility of regular indefinite objects to the accessibility of indefinite demonstrative objects in a visual-world experiment very similar to the design of Karimi and Ferreira (2016).³ Auditory input consisted of German short stories of three sentences each, see Table 2. The first sentence provided an anchor expression (bar) that activated a frame. The second sentence was a transitive sentence with a proper name in subject position and the critical referent (a regular indefinite or an indefinite demonstrative noun phrase) as second argument. The third sentence contained an ambiguous pronoun. The first sentence manipulated the information status of the critical item by using an anchor expression that either evoked a frame such that the critical item was part of that frame and thus pre-activated that item (inferred: bar → waiter), or a frame that was unrelated to the critical item and thus did not pre-activate that item (brand-new: exhibition hall↛ waiter).⁴

While participants listened to the short stories, they were presented with a visual display on a screen with pictures for the two arguments from Sentence 2 and two pictures of unrelated inanimate objects. Looks towards the picture of the first argument (NP1: Peter) were compared to looks to the picture of the second argument (NP2: waiter) from 100 ms before the onset of the ambiguous pronoun to 1,400 ms after pronoun onset, divided in time frames of 300 ms each. There was no difference between inferred and brand-new indefinite demonstratives, but there was a significant effect of information status for regular indefinites in the 1,100 to 1,400 ms bin, with more looks to the inferred indefinites than to the brand-new indefinites. In other words, indefinite demonstratives (this waiter) show the same level of accessibility whether brand-new (exhibition hall … this waiter) or inferred (bar … this waiter). But regular indefinites are more accessible when they are inferred than when they are not inferred. This result supports the assumption that information status contributes to the accessibility of discourse referents introduced by definite and indefinite noun phrases (see also Strube & Hahn 1999; Brocher et al. 2016).

The two experiments reported above suggest that informativity and information status each contribute to a higher accessibility of referents. However, these two factors may not do so additively. Brocher and von Heusinger (2018) investigated the accessibility of modified (i.e., more informative) definite and indefinite NPs in inferred and brand-new conditions in a visual-world paradigm similar to that of von Heusinger and Brocher (2019). The NPs were modified by post-nominal locative prepositional phrases (PPs) such as am Fenster (‘next to the window’), see Table 3.

Unlike in von Heusinger and Brocher (2019), there was no significant positive difference in the proportion of looks toward the object between the brand-new indefinite and the inferred indefinite condition. Numerically, the effect was even reversed.

In this paper, we further investigate influences on the accessibility of indefinite noun phrases building on these experimental insights. Since the main difference between the Brocher and von Heusinger (2018) and von Heusinger and Brocher (2019) experiments is the modification on the NPs, it appears that informativity interacts with information status when it comes to the accessibility of referents introduced by indefinite NPs. The current experiment tests this hypothesized interaction.

2 Current Experiment

We conducted a visual world eye-tracking study, comparing the accessibility, as measured by the probability of looks toward a referent upon hearing an ambiguous pronoun, of unmodified brand-new indefinites, unmodified inferred indefinites, informative brand-new indefinites, and informative inferred indefinites. This experiment was different from Brocher and von Heusinger (2018) as it did not manipulate the determiner (regular indefinite article vs. indefinite demonstrative), but used regular indefinites in all items and manipulated information status and informativity. If informativity and information status interact, we expect informative brand-new indefinites to be more accessible than less informative brand-new indefinites, but either a reversal of the informativity effect for inferred indefinites or no difference in accessibility between the two inferred indefinite conditions. If the effect of the two factors on the accessibility of referents is additive, we expect informative indefinites to be more accessible than unmodified indefinites, and inferred indefinites to be more accessible than brand-new referents.

2.1 Method

2.1.1 Participants

Thirty-four native speakers of German participated in the experiment for either course credit or monetary compensation (EUR 8/hour). All participants had normal or corrected-to-normal vision.

2.1.2 Materials

We constructed 40 German short stories consisting of three sentences each, see Table 4. The first sentence contained context information that did or did not provide an anchor expression of a frame related to the target noun in object position of sentence 2. The second sentence introduced two human referents, one in subject position and one in direct object position. The object referent was either modified by a post-modifying PP, such as am Eingang (at the entrance) or unmodified.⁵ The PPs that were included in the respective conditions were always descriptions of locations.⁶ The locations were not represented in the accompanying visual display, which only showed the referents in isolation (see Figure 1). The final sentence of the auditory input always started with a temporal subordinate clause,⁷ followed by a main sentence including an ambiguous pronoun that referred to either the subject or the object of the second sentence. The pronoun was always the German masculine singular er (‘he’) and ambiguous between two male referents, because the feminine singular sie (‘she’ or ‘they’) is also compatible with a plural interpretation, which would then no longer be ambiguous between two singular antecedents.

Table 4

Sample stimulus (English translations in italics).

Sentence number	Manipulation	Sentence
1	Information status: inferred	In der Bar war es mal wieder rappelvoll. The bar was again very crowded.
	Information status: brand-new	In der Messehalle war es mal wieder rappelvoll. The exhibition hall was again very crowded.
2	Informativity: modified	Peter rief einen Kellner am Eingang. Peter called a waiter at the entrance.
	Informativity: unmodified	Peter rief einen Kellner. Peter called a waiter.
3	–	Als ein Glas vom Tisch fiel, drehte er sich verwundert um und schüttelte den Kopf. When a glass fell off the table, he turned around and shook his head.

In addition to the 40 experimental short stories we also included 80 filler stories in the experiment, which were similar in length and structure to the experimental items. To make them as similar as possible to the experimental stories, the fillers also only introduced male characters. None of the filler items involved any referential ambiguity.

All materials were recorded by a trained female speaker and were pre-processed before the experiment. All pre-processing procedures were conducted in Praat (Boersma 2001). For all experimental stories, the three sentences (context sentence, referent sentence, ambiguous pronoun sentence) were first separated and then re-arranged to keep everything except the manipulated sentence identical between conditions (e.g., all four conditions of an item had the same final sentence, both stories in the brand-new condition used the same recording of the first sentence, etc.).

Figure 1

Example display. The referent in subject position (see Table 4), Peter, is presented in the upper right corner, the critical referent in object position, waiter, in the lower left corner.

Each auditory stimulus had a corresponding visual display with four pictures: one for the referent in subject position, one for the referent in object position, and two unrelated filler pictures showing inanimate objects. Figure 1 shows the display for the item in Table 4. We partly selected our pictures from a database by Duñabeitia et al. (2018). The remaining pictures for test items, including all pictures of the human referents, were provided by a professional illustrator. The four pictures were arranged in the same way for each condition of an item. Between items, subject and object pictures appeared equally often in each corner of the screen.

All stories were distributed across four presentation lists. No item appeared more than once per list, and all conditions appeared equally often between lists.

2.1.3 Procedure

Participants sat in front of an EyeLink 1000 configured with a tower mount at a distance of approximately 65 cm from the screen. We recorded participants’ eye movements at a sampling rate of 250 Hz. At the beginning of a trial, we presented participants with the picture of the subject referent that was about to appear in the following story. This picture was accompanied by a phrase that introduced the referent, such as Hello, my name is Peter. This was done because, unlike for the object referents, subject referents could not be depicted in an identifiable way (e.g., there is no picture of ‘Peter’). After familiarization, participants clicked their mouse to start the next screen, which showed the entire visual display, see Figure 1.

After 500 ms, the audio started. Participants were asked to carefully follow the stories to be able to correctly answer a comprehension question in case one appeared. 1,000 ms after the end of the auditory stimulus, participants saw either a comprehension question or the question Weiter? (‘Continue?’).⁸ In case of a comprehension question, participants pressed the right mouse button to respond ‘yes’ and the left button to respond ‘no.’ 2,000 ms after the response, the next trial was initiated.

2.1.4 Data analysis

For pre-processing of the data and inferential statistics, we used R (version 3.5.0) with the lme4 (version 1.1.19) and the tidyverse package (version 1.2.1). We analyzed eye fixations on the object and subject pictures for a total of 1,500 ms, from 100 ms before until 1,400 ms after onset of the ambiguous pronoun. We divided the 1500 ms into five windows of 300 ms each. For each time window and picture individually (i.e. object and subject picture), we fitted a linear mixed effects model with the proportion of looks to the target picture as a dependent variable and information status (inferred vs. brand-new) and informativity (modified vs. unmodified) as predictors.

All models contained between-participants and between-items intercepts as well as random slopes for information status, informativity, and their interaction. We assume absolute t-values of 2 or higher as indicating statistical robustness (Baayen 2008).

2.2 Results

Accuracy of responses to the comprehension questions was 90.1%. 12.9% of the data were not included in the analyses, either because they reflected blinks or because no picture on the screen had been fixated.⁹

The time course of looks to the picture of the object referent is plotted in Figure 2. The plot shows that, as early as the onset of the pronoun, referents of unmodified NPs received overall more looks when they were inferred than when they were brand-new. This part of the results is in line with the findings from von Heusinger and Brocher (2019). For modified, i.e., more informative, NPs, the opposite pattern emerges somewhat later in the time course: Referents of brand-new noun phrases were fixated more than referents of inferred ones. This pattern is in line with the results of Brocher and von Heusinger (2018). Table 5 summarizes the results of the inferential statistics.

Figure 2

Time course of looks (in %) to the picture of the referent in object position; Zero marks the onset of the ambiguous pronoun; Without PP = less informative object NPs; With PP = informative object NPs; Brand-new = object NPs with brand-new concept; Inferred = object NPs with inferred concept.

The output of the regression model that included the fixation times for the window 200 ms – 500 ms¹⁰ post pronoun onset revealed a significant Information status x Informativity interaction. No other effects reached statistical reliability.¹¹ To help interpret the interaction effect, we plotted the model estimated marginal means (obtained using the emmeans package, version 1.8.5), see Figure 3. As the plot shows, the effect of Informativity is different for brand-new referents than for inferred referents: while a PP makes inferred referents less accessible, it makes brand-new referents more accessible. This effect is in line with the main hypothesis of our study.

Table 5

Model output for looks to the picture of the object referent.

Time	Main Effect/Interaction	b	SE	t
–100 – 200	Intercept	0.440	0.031	14.00
	Information status	–0.016	0.026	–0.60
	Informativity	0.012	0.031	0.40
	Information status × Informativity	–0.079	0.050	–1.59
200 – 500	Intercept	0.428	0.027	16.12
	Information status	–0.016	0.026	–0.63
	Informativity	0.006	0.031	0.19
	Information status × Informativity	–0.111	0.051	–2.17
500 – 800	Intercept	0.421	0.027	15.87
	Information status	0.014	0.031	0.47
	Informativity	–0.015	0.027	–0.55
	Information status × Informativity	–0.060	0.054	–1.10
800 – 1100	Intercept	0.394	0.026	14.88
	Information status	0.004	0.030	0.13
	Informativity	–0.027	0.028	–0.95
	Information status × Informativity	0.004	0.048	0.09
1100 – 1400	Intercept	0.372	0.024	15.27
	Information status	–0.016	0.033	–0.49
	Informativity	–0.033	0.026	–1.29
	Information status × Informativity	0.048	0.047	1.02

• Notes. Time in milliseconds relative to the onset of the ambiguous pronoun. Information status = information status of the object noun phrase (inferred or brand-new); Informativity = Informativity of the object noun phrase (modified by a PP or not). Significant effects are highlighted in bold.

Figure 3

Model estimated marginal means of the proportion of looks to the object picture in the 200–500 ms time bin, by Information status (brand-new versus inferred) and Informativity (with PP versus without PP).

3 General Discussion

Our results are in line with studies that show that informativity and information status affect the accessibility of referents. However, our study is the first to show that these two factors interact: Modification makes inferred indefinites less accessible, while it makes brand-new indefinites more accessible. Informativity boosting the accessibility of brand-new referents is in line with the results of von Heusinger & Brocher (2019) for indefinites, and of Karimi et al. for definites. However, the finding that informativity has a negative effect on the accessibility of inferred referents is more surprising, although a first hint of this effect was suggested by the experiment reported in Brocher & von Heusinger (2018) for indefinites. As this result is not predicted by any theoretical account, but seems robust, we are in need of an explanation of why informativity impacts inferred indefinite NPs differently than brand-new ones. We discuss one potential explanation for this interaction, namely that the type or role reading triggered by inferred indefinites does not match the locative PP modification, which requires a token or individual reading.

Before we elaborate on this explanation, we have to briefly discuss two assumptions that have so far not been discussed in this paper: First, we assume that while individual referents are typically introduced by pronouns, demonstratives and definite NPs, the semantic contribution of indefinites is a predicate or a concept describing a set of elements with the same property (Heim & Kratzer 1998). Additional contextual and general pragmatic inferences allow type shifts from predicates to individual referents (tokens) or to roles (types) (Partee 1987). We assume that contextual information can trigger one or the other type shift. Second, we assume that in our test design the ambiguous pronoun in the third sentence refers to an individual referent, as required by the predicate in that sentence (e.g. turned around). We assume that the pronoun establishes the anaphoric relation to the indefinite direct object by two interacting processes: First, the referent of the pronoun must be an instance of the concept introduced by the indefinite, and second the pronoun refers to the same individual referent as the indefinite in the antecedent sentence. Both processes support each other and create the prominence level of the direct object that can compete with the prominence level of the subject.

A potential explanation could thus be developed along the following lines: we start with the assumption that indefinites introduce a predicate that describes a set of potential referents (say different waiters). During the process of the incremental interpretation this predicate can be shifted to an individual (or token) or to a role (or type) or the indefinite can be underspecified and both types are available. In the case of brand-new indefinites (exhibition hall … a waiter [PP]), the indefinite is shifted to a token interpretation that can be picked up by the anaphoric pronoun in the next sentence without any type mismatch. The additional modification with a locative PP provides more information and makes the referent more prominent. This would account for the observation that for brand-new indefinites modification (more informativity) raises accessibility. The situation is different for inferred indefinites (bar … a waiter [PP]). The frame of a bar triggers a role (or type) reading of the indefinite, i.e., a waiter refers to a stereotypical referent fulfilling the assumed properties of a waiter in a bar.¹² In the unmodified case, the personal pronoun in the third sentence refers then to an instance of the role or type, and not yet to a specific individual referent. Since the locative PP clashes with the frame reading, the PP triggers a shift to a token reading. For both brand-new indefinites and inferred indefinites, the eventual interpretation is thus a token reading, but the frame reading is retained longer for inferred indefinites. Since the locative PP triggers the token reading, it does not have the prominence-boosting effect that it has on already established referents (token readings), resulting in an overall lower prominence level.

We thus postulate that the interaction of informativity and information status is caused by the very underspecified semantics of indefinites. This would predict that for other NPs, such as definite descriptions (the waiter) or demonstratives (this waiter), such an interaction cannot be found, a prediction that is supported by Brocher and von Heusinger’s (2018) results: In their study, inferred informative definites are more accessible than brand-new informative ones.

To sum up, we corroborated studies by Brocher and von Heusinger (2018) and von Heusinger and Brocher (2019) that find that informativity and information status both influence the accessibility of discourse referents, and we provided novel empirical evidence that these features interact when it comes to the accessibility of indefinite NPs. Future research should work toward uncovering the exact mechanism behind this interaction, for instance by using another type of modification, or modification that is compatible versus incompatible with a role reading (e.g., the mathematics teacher vs. the big teacher); this will contribute to a better understanding of the discourse-level function of indefinite NPs, the processes governing pronoun resolution, and, in general, discourse processing.

Data Availability

All experimental items as well as our data and data analysis script can be accessed at https://doi.org/10.17605/OSF.IO/46REU.

Ethics and consent

The experiment was approved by the Ethics committee of the German Linguistic Society before data collection commenced.

Funding information

The research for this paper was funded by the German Research Foundation (DFG) as part of the SFB 1252 “Prominence in Language” (Project-ID 281511265) in the project C04 “Conceptual and referential activation in discourse” at the University of Cologne, Department of German Language and Literature I, Linguistics.

Competing interests

The authors have no competing interests to declare.