Glossa Psycholinguistics

2.1 Uniqueness and familiarity requirements as triggers of bridging

There are, broadly speaking, two families of theories of definites: uniqueness theories and familiarity theories. Uniqueness theories hold that definite NPs⁴ imply, or require for their felicity, that there is a unique (salient) satisfier of the descriptive content of the definite in some contextually relevant situation. (Henceforth, we’ll say that the intended referent is situationally unique.) The uniqueness analysis is extended to plural definites, too. In the case of plurals, the unique satisfier must be a collective entity, the mereological sum of all entities satisfying the descriptive content (Link, 1983; Sharvy, 1980). For example, sentence (3) below is predicted to be felicitous just in case there is a unique (salient) maximal set of doors in the relevant situation.

Moving to the second family of views, familiarity theories of definites characterize definites as carrying a familiarity presupposition or implication (Haviland & Clark, 1974), or as anaphoric (Heim, 1982; Roberts, 2003). Distinguishing uniqueness from familiarity is not always straightforward, as contextual uniqueness is often taken to be adequate to establish familiarity (see Coppock, 2022). Going in the other direction, Roberts (2003) attempts to derive the observations supporting uniqueness claims from the familiarity requirement. In contrast, Schwarz (2009) argues that definites are ambiguous between a familiarity-requiring meaning and a uniqueness-requiring meaning.

Both familiarity and uniqueness perspectives on definites have been used to argue for definiteness as a driver of bridging. Clark (1975) was first to discuss bridging, and did so in the context of his work on the Given-New Contract (Clark, 1973; Clark & Haviland, 1977). Clark is a familiarity theorist. He argues that definites signal that the intended referent is Given, which, for him, means that it can be identified with something already in the memory of the addressee, or something which the addressee can compute on the basis of their existing information. On this view, definites which lack an antecedent, such as the door in (1a), present a problem for the addressee: the speaker has referred to some door with a definite, signaling that it is Given, but the addressee has no relevant door stored in memory. The addressee solves this problem by making the bridging inference that the house mentioned has a door, and that that door is the intended referent of the definite. (Cf. Clark’s (1975) discussion of “indirect reference by association”.)

It should be noted from the outset that this account contains the implicit assumption that hearers prefer to resolve the apparent violation of constraints on definites by bridging, rather than by simply assuming that the referent is some door or other that they are supposed to already know about. Hearers sometimes do, in fact, interpret novel definites this way; the occurrence of the house in the first sentence of (1a/b) is of this sort, and presents no problem for interpretation (at least in the current context). The hearer understands that they are supposed to envision a situation containing a unique house. This kind of solution to the problem is called accommodation (Lewis, 1979, a.o.). If simple accommodation were a default or preferred strategy to deal with apparent violations of definiteness constraints, with bridging as a less-preferred option, then bridging would be expected to occur more rarely, and when it does occur, some explanation would be needed for why the interpreter has deviated from the preferred strategy. So we can assume that those who propose definiteness as a trigger of bridging assume that bridging is a kind of default solution, or at least preferred over simple accommodation, given an available anchor.⁵ The idea that bridging is preferred over simple accommodation is, indeed, argued for by Heim (1982, pp. 239–240), for whom this presumed preference serves an explanatory role. She considers examples such as the following:

Heim observes that this sentence can only naturally be interpreted as ‘John wrote to the author of the book that he read,’ and not simply as ‘John wrote to some author or other that is already familiar’. She points out that this demonstrates that “accommodation in response to definites is not normally a matter of just adding the minimal amount of information that would restore felicity”, but rather that a definite lacking a contextual antecedent “has to be linked by crossreferences to some already-present [discourse referent]” (p. 240).⁶ Heim states the case a little too strongly here, as clearly there are interpretable cases of novel definites where bridging is not possible, as in (2) above; but Heim’s example (4) is equally well explained by assuming that bridging is the default solution to an otherwise infelicitous definite, with accommodation of an unlinked discourse referent a kind of last resort.⁷

We follow the literature in assuming that whatever strategy is presumed to be involved in rendering the definite felicitous in its context, the semantics of the definite remains the same. In the theories under discussion, accommodation and bridging are interpretational strategies triggered by the proposed semantics. The theoretical literature offers purely formal models of these different options. Following antecedents in the psycholinguistic literature, we presume these interpretative strategies to have cognitive correlates, at least with respect to mental representations of discourse content.

We have seen, then, that familiarity theorists see bridged interpretations of definites as arising as a solution to an apparent violation of the familiarity requirement. Bridging of indefinite NPs, which lack any familiarity requirement (and, indeed, are sometimes taken as markers of novelty), obviously cannot be explained in the same terms. Hence, on this view, the occurrence of bridging-like interpretations of indefinites requires a different explanation than bridging of definites. It cannot be taken to be triggered by definiteness. We emphasize that what is at issue here is what triggers the bridged reading, and not the (cognitive) procedure involved in establishing a bridged reading, which could be identical in both cases.

Like familiarity theorists, uniqueness theorists — those who posit that definites require situational uniqueness or unique identifiability of the referent — have also offered accounts of definiteness as a trigger of bridging. While Roberts (2003) is a familiarity theorist, she discusses the example in (5) in terms of situational uniqueness (which, in her account, is required to meet the retrievability condition on an antecedent). Imagine the following, uttered in a context where no dashboard has previously been mentioned:

She points out that hearers “generally know that … there is more than one dashboard in the world,” so in the absence of any situationally unique dashboard, the definite is potentially infelicitous. However, as hearers also know that there is generally only one dashboard per car, if the NP is interpreted (roughly) as “dashboard of the just-mentioned car,” uniqueness is satisfied, and the felicity of the NP is guaranteed. Specifically, Roberts proposes that bridging involves reinterpreting dashboard as a relational noun, with (the discourse referent of) the just-mentioned car providing the second relatum.⁸ For plurals, a parallel argument can be given. Consider:

Utterance of this sentence requires that there be a unique salient set of tires in the context. Here, reasoning might go as follows: there is potentially a unique maximal set of tires, namely, the set of all tires in the world, but the speaker could not plausibly intend to refer to these. The context does not make salient any other set of tires, but if tires is reinterpreted as “tires of the just-mentioned car”, then once again, a single set of tires becomes salient, namely, those of the car. Hence, tires is given a bridged interpretation, satisfying (plural) uniqueness.

A similar account is given in Kehler (2015) of the example below:

Kehler explains the bridging as follows: “On a plausible analysis of such cases, the hearer accommodates (Lewis, 1979) the existence of the referent… The hearer, confronted with a referring expression that requires a uniquely identifiable referent, infers the existence of courthouse steps to meet that constraint.” (Kehler, 2015, p. 634).

Thus, both familiarity theories and uniqueness theories of definites provide explanations of the discourse conditions under which bridging will occur, namely, in order to satisfy the otherwise unmet felicity conditions on the use of a definite. In both cases, the explanation does not extend to bridging of indefinites, so despite the surface similarity of bridging of definites and indefinites, different explanations are required.

2.2 Extending bridging to indefinites: The one-phenomenon view of bridging

The observation that indefinites can bridge (in our sense) goes back at least to Hawkins (1978, pp. 173ff.), who treats them as parallel to the associative anaphoric uses of definites. Hawkins’ examples include these:

Hawkins proposes that associative links involving indefinites are subject to the same (complex and difficult to specify) pragmatic constraints that govern associative anaphora with definites, but observes a contrast in the relation to uniqueness: the felicitous use of an associative indefinite requires that the listener can assume that the intended referent (singular or plural) is not situationally unique. Hence, (8) is felicitous with the singular a page, because books typically have more than one page, and also with the plural some pages, as books typically have more than two pages (hence, the speaker can use some pages to describe a plural subset of the total set of pages).

Hawkins observes that indefinites with a preceding “trigger” (in our terminology, an anchor) can often be understood either as associated with the trigger or not. (See also the discussion in Kehler (2015, p. 635) regarding his example (34).) In the following example, Hawkins claims, it’s unclear whether the tires that Fred sold came from the car he bought, or not:

Hawkins holds, though, that there are cases where an indefinite is obligatorily interpreted associatively, namely, where the context satisfies the non-uniqueness requirements that he assumes and where the “pragmatics of the remainder of the sentence” force this reading, as intuitively is the case in (8) and (9) above. Hawkins’ discussion, thus, suggests a commonality between the bridged, or associative, readings of definites and indefinites. Both stem, somehow, from the associations arising from the mention of the anchor noun.

To our knowledge, the only established formal model of bridging which recognizes a commonality between bridging of definite and indefinite NPs is that of Asher and Lascarides (1998). Asher and Lascarides argue that bridging is a consequence of, and subserves, the construction of coherent discourse structure, an idea which originates with Hobbs (1979). On their view, constructing discourse structure requires the establishment of rhetorical relations between each Elementary Discourse Unit (typically, but not always, a clause; see Asher & Lascarides, 2003). They propose that bridging relations are constructed where they are necessary to support a plausible coherence relation or, alternatively, where the relation between the entities is implied by the most plausible coherence relation available. Let’s illustrate this with the basic examples from the beginning of the article, repeated here:

On the Asher and Lascarides account, the interpreter of the sequences in (1) is in search of a plausible rhetorical relation between the two segments. For both examples, the most plausible relation would be Elaboration or Background. But in order for the second segment to elaborate on the eventuality described in the first, the door under discussion must be a door related to the eventuality of Yasmin approaching the house. Given world knowledge that houses typically have doors, it is straightforward for the listener to infer that the door in question belongs to the house introduced in the first segment. As houses can have a unique (or uniquely salient) external door, use of the definite is felicitous and plausibly leads the hearer to further infer unique instantiation. As houses also can have multiple external doors, use of the indefinite (which can, in turn, be seen as avoidance of the more restricted definite) implies that uniqueness is not satisfied, so the hearer may (but need not) envision multiple doors in the situation. Thus, on the Asher and Lascarides view, definiteness has a clear role to play in interpretation overall, but is not the feature responsible for triggering bridging. Indeed, we should emphasize here that rejection of the claim that definiteness is the primary trigger of bridging does not entail a rejection of the claims surveyed in the previous section about the function of, or constraints imposed by, definites (or indefinites).

The approach advocated by Asher and Lascarides, as well as Hawkins’ earlier discussion, demonstrates that a theory can fully represent the role of definiteness (and indefiniteness) in supporting and constraining bridging without taking definiteness to be the trigger of bridging. Suppose that interpreters encountering a novel noun with a plausible anchor already in the discourse typically build the relation to the anchor into their initial interpretation, irrespective of definiteness. But interpreters also seek to satisfy any requirements created by the use of a definite (or indefinite), such as a requirement for situational uniqueness (or non-uniqueness). On this account, bridging driven by an observation of plausible relations between entities facilitates satisfaction of the requirements of a definite, by making salient a uniqueness-satisfying relation. Thus, definiteness may reinforce or strengthen the likelihood of a bridged reading without being its immediate trigger.

This kind of view is consistent with conclusions drawn by Schumacher (2009) from an ERP study of the interpretation of definites and indefinites. In her study, she investigated the ERP signatures associated with the interpretation of definite and indefinite NPs in three conditions: repetition of a previously used noun (Given), a new noun highly related to prior context (Inferred), and an unrelated noun (New). Schumacher explored both early (N400) and later (P600) ERP effects. With respect to the N400, Schumacher found no significant effect of definiteness; on this basis, she argues that in early interpretation, hearers attempt to integrate new NPs into existing representations regardless of definiteness. It is only with respect to the P600 that any effect of definiteness occurs. Schumacher argues that “the linking attempts occur for both definite and indefinite entities and are governed by the fit of the head noun with the information provided by the context. This suggests that the underlying processes are guided by coherence constraints and that a strict correspondence between definiteness marking and integration processes cannot be maintained” (p. 100).⁹ Thus, Schumacher’s data suggest that the search for an antecedent in context is not motivated by the presence of a definite. Consequently, her conclusions can be read as supporting a one-phenomenon view of bridging, a view according to which the bridging of both definites and indefinites is triggered in the same way – by general expectations of, or preferences for, links between prior discourse and new information.

The experiments we report here similarly explore the role of definiteness in bridging, but using a different set of methodologies that explore later stages in interpretation. A key observation underlying our experimental methodology is that there are cases where definiteness and more general considerations of coherence push in different directions. These cases are particularly useful for exploring whether definiteness plays an independent role in triggering bridging. Consider, then, the examples in (11), where the subject NPs the/a doctor are not highly related to any entities mentioned in the context given:

Regardless of the identity of the doctor, the second sentence here is interpretable as cohering with the first via the Elaboration relation. Establishing this relation doesn’t require assuming any particular relation between the doctor and anything mentioned in the first sentence. As far as coherence establishment is concerned, there is no more reason to posit a bridge in the definite case than in the indefinite case. With respect to interpretation of the definite, bridging is not obligatory for felicity; the interpreter can simply assume that the definite the doctor refers to some assumed-familiar referent, as with the stable in the initial sentence. But, as also noted earlier, theories of bridging as definiteness-driven seem predicated on the assumption that bridging (accommodation of a between-entity relation) is preferred over simple accommodation of a new referent as a strategy for meeting the requirements of definites. The two-phenomena view of bridging would, therefore, seem to predict an increased tendency to attempt to bridge doctor to the content of the previous sentence in the definite case, over the indefinite case. It is this prediction that will be tested in the experiment described in Section 4, where we test for an independent effect of definiteness that persists even in passages that lack a highly related entity to serve as an anchor.¹⁰

4.1 Participants

For the main task, seventy-two participants who had IP addresses in the United States or Canada were recruited via Amazon Mechanical Turk. All participants were paid $10 for a task estimated to take under an hour. We only used data from participants who answered “no” to a question in a background questionnaire that asked whether any language besides English was spoken at home before the age of 6, in order to restrict ourselves to analysis of data from monolingual native-speaker participants. The data for the analysis comes from 55 monolingual English-speaking participants.

4.2 Materials

The materials followed the dialogue structure shown in (15), with the content of the first turn determined from an elicitation task with a separate set of participants.

For the elicitation task, we recruited a separate set of 72 English-speaking participants, none of whom participated in the main task and all of whom were paid $13 for a task estimated to take an hour. We used the data from 54 monolingual English-speaking participants. Participants accessed a web-based interface where they wrote story continuations for prompts like those in Table 1.

For each item set in the elicitation task, we manipulated Definiteness and Relatedness: each of the 40 target nouns (waiter in Table 1) appeared in either a definite or indefinite NP and was preceded by a context sentence that mentioned either a high-related anchor (restaurant) or contained only low-related nouns (here, news). The materials were constructed so that each noun and each context sentence appeared in all four conditions. In other words, not only did the target noun waiter appear in all 4 conditions, as shown in Table 1, but the high-related context sentence in Table 1 also appeared as a low-related context sentence for another noun (Nigel and I went out last night to that new restaurant. The report…), and vice versa for the other context sentence (I was really shocked by something I saw yesterday on the news. The report…). This counterbalancing ensures that in the main task, any evidence for bridging in the high-related condition could not be attributed to an independent effect from the context sentence. For example, it is possible that some context sentences have content that is more likely to be re-used in a dialogue continuation; maybe going out to a restaurant is inherently more interesting to talk about than seeing something in the news, or vice versa. See Appendix A for further information about the elicitation task, with the full set of elicitation materials available in the Supplementary Materials.¹³

To construct the main task materials, we sampled the set of continuations produced by participants in the elicitation task. All 40 target NPs from the elicitation task were represented, and we selected 4 continuations written in the high-related condition and 4 written in the low-related condition. These included both singular and plural target nouns (26 singular, 14 plural). We independently manipulated the definiteness of the target NP to yield a set like those shown in Tables 2 and 3 for each target NP.

Using elicited continuations, as we did here, avoids experimenter bias in the readability of the high- versus low-related conditions. The manipulation of definiteness means that roughly half the time, the Experiment 1 passage deviates from the elicitation participant’s original version (as elicitation participants were responding either to a definite or an indefinite prompt). However, we note that there were over a dozen cases where the elicitation data included the same continuation in both the indefinite and definite conditions, suggesting that these continuations are amenable to a definiteness manipulation (e.g., Hilda created a nice arrangement of fruit. [The, A] banana was the centerpiece). We included a number of those cases in the Experiment 1 materials, yielding a distribution across the continuations selected from the elicitation task such that half the continuations were from underlying indefinite prompts, and half were from underlyingly definite prompts, with 5 produced for both indefinite and definite prompts.

For the choice of continuations, we randomly selected NP continuations from the elicitation dataset and then eliminated selections and made new selections according to the following criteria: no continuations with modifiers (e.g., The waiter who brought our food was rude), no continuations with non-referential NPs (e.g., A rug would make the room look better), no continuations with additional common ground assumptions (e.g., beforehand in The lion had attacked some people beforehand), no continuations with idiomatic expressions that require a definite (e.g., The clouds are few and far between), no continuations that sounded awkward or confusing to a native speaker (e.g., The page was given in the mail for coupon deals at the grocery store), and no continuations for which the definiteness manipulation produced incoherence (e.g., Some leaves are brown and some are green, which becomes contradictory if the first NP is definite as in The leaves are brown and some are green). Replacements were selected to balance the overall number of continuations that had been originally definite/indefinite in the elicitation task; the materials for any given target NP included at least one continuation that had been elicited from a definite prompt, and at least one from an indefinite prompt. We fixed small typos (e.g., it’s → its, wont → won’t). The only other alteration we made to the elicited continuations was to use the/some as the definite/indefinite alternation for NPs with the target noun fish (instead of the/a); this was because almost all of the continuations with this target noun with the definite in the elicitation task had been assigned a plural interpretation. (See the Supplementary Materials for the full set of materials).

Eight lists were constructed, such that participants saw each of the 40 target NPs once in a single passage (e.g., waiter in one of the 8 variants in Table 2). Each list included a particular context sentence only once (i.e., a participant who saw Nigel and I went out last night to that new restaurant with the target NP the waiter did not see that context sentence with its low-related NP the report). Each list contained an equal number of high-related and low-related items and an equal number of definites and indefinites. In addition to the 40 target items, each list contained 40 filler passages which were similar to the target dialogues: Speaker A’s first turn consisted of two sentences; Speaker B asked a clarification question (e.g., What did you say? Why is school closed?, or Sorry, someone was talking to me. What was on the test?), and the participant was asked to fill in Speaker A’s reply.

4.3 Procedure

Participants accessed the experiment via a web-based interface linked from the Mechanical Turk environment. For all experiments reported here, participants gave informed consent before proceeding to the task. Each item was displayed on a separate page. Participants were instructed to fill in the final utterance in the dialogue. (See the Supplementary Materials for full detail.)

4.4 Results

For the 2200 responses collected, we used a string similarity metric as a proxy for participants’ assignment of a bridged interpretation. The participant responses in (16) and (17), shown underlined, illustrate the contrast we are aiming to capture. In (16), the participant is clearly linking the newly mentioned waiter to the context provided by the initial sentence; this is shown by the re-use in the response of multiple words from the context sentence. In (17), in contrast, the participant does not re-use words or content from the context sentence, but instead gives a rather under-informative response, simply repeating the content that is asserted in the continuation sentence. We take this to indicate that the participant has assigned a non-bridged interpretation to the NP – plausibly, the participant doesn’t have any clear idea of what lion is being talked about.

For the string overlap measure, we compute a score for the context~response sentence overlap as a proxy for bridging. We also compute the overlap between the participant’s response and the continuation sentence, with the hypothesis that in unbridged cases like (17), the participant may resort to reasserting information from the continuation sentence in order to answer the clarification question. For any pair of sentences, the similarity measure was computed by treating each sentence as a “bag of words” represented as a vector in a multi-dimensional lexical space. We then evaluated the distance between those two vectors. Pre-processing of the sentence text was conducted using the Python Natural Language Toolkit (nltk; Bird et al., 2009). Specifically, the bag of words for a given sentence consisted of all the words in that sentence after we excluded punctuation and stop words (e.g., and, that, was; nltk.corpus’ stopwords.words(“english”)).

We present two analyses – one with stemmed words (e.g., cats/cat both appear as cat; nltk PorterStemmer) and one with lemmatized words (e.g., buy/buys/buying/bought all appear as buy; nltk WordNetLemmatizer). The distance between two sentences was measured as the cosine of the angle between their respective word vectors (Manning & Schütze, 1999). With this measure, two sentences that consist of an identical set of words have value 1, and two sentences that have no overlapping words have value 0.¹⁴ Appendix B includes a sample of participant responses and the associated similarity scores that were computed for the context~response comparison and the continuation~response comparison.

Table 4 and Figure 1 show the mean similarity scores between the context sentence and the participant response. There are higher scores for high-related NPs than low-related NPs, and little difference by Definiteness. We construct a linear mixed-effects regression to predict the similarity score with fixed effects of Relatedness and Definiteness and their interaction (coded as –.5/+.5 low/high and –.5/+.5 indefinite/definite). The model contained random effects for participants and NPs, with random intercepts and slopes. The converging model for stemmed similarity contained only by-participant random slopes for Relatedness and Definiteness and a by-NP random slope of Relatedness (see Barr et al., 2013, for model convergence approaches). For lemmatized similarity, the converging model contained full by-participant random-effects structure, but only a by-NP random slope of Relatedness. The results show a significant main effect of Relatedness (stemmed: B = 0.147, SE = 0.02, t = 7.08, p < 0.001; lemmatized: B = 0.141, SE = 0.02, t = 6.688, p < 0.001). There is no main effect of Definiteness nor a Relatedness × Definiteness interaction (p’s > 0.3).

Figure 1: Context~response similarity (bridging proxy) as a function of Relatedness and Definiteness on two types of similarity scores (stemmed on left and lemmatized on right); error bars show standard error over by-participant means.

Table 5 and Figure 2 show the similarity scores between the continuation sentence and the participant response. This measure doesn’t index bridging and, instead, may capture ways in which participants seek to link an entity to the context when a bridged interpretation is not favored. The results show the highest similarity scores in the low-related indefinite condition. The converging model for stemmed similarity contained only by-participant random slopes for Relatedness and the Relatedness × Definiteness interaction and a by-NP random slope of Relatedness. For lemmatized similarity, the converging model contained only by-participant and by-NP random slopes for Relatedness. The results show a significant main effect of Relatedness, whereby the continuation~response similarity is higher for the low-related condition than the high-related condition (stemmed: B = –0.077, SE = 0.016, t = –4.810, p < 0.001; lemmatized: B = –0.077, SE = 0.016, t = –4.942, p < 0.001), and a less reliable effect of Definiteness, whereby definites have lower scores than indefinites (stemmed: B = –0.0181, SE = 0.009, t = –2.001, p < 0.05; lemmatized: B = –0.0154, SE = 0.09, t = –1.723, p = 0.09), which together are driven by a significant Relatedness × Definiteness interaction (stemmed: B = 0.052, SE = 0.019, t = 2.693, p < 0.001; lemmatized: B = 0.054, SE = 0.019, t = 2.894, p < 0.005).

Figure 2: Continuation~response similarity as a function of Relatedness and Definiteness on two types of similarity scores (stemmed on left and lemmatized on right); error bars show standard error over by-participant means.

In reviewing our materials, it is apparent that our items are somewhat heterogeneous. We, therefore, undertook several post hoc analyses (see the Supplementary Materials) to test the stability of our findings across different subsets of the data, focusing on whether an independent effect of definiteness is apparent under a different treatment of our data. We compare singular vs. plural target nouns and relational vs. non-relational target nouns,¹⁵ and we eliminate additional cases with non-referential nouns. Lastly, we report a uniqueness rating study (see Appendix C) that we conducted to test that our items satisfied our intended constraint that all context~noun pairings allow an interpretation in which the noun (or the entity it corresponds to) could plausibly be either uniquely or non-uniquely instantiated in the situation given by the context. The only problem stimuli we identified in the rating study were two cases where the likelihood of uniqueness was low. However, in both of these cases, the target noun is robustly relational (target noun leg, given chair in context, and target noun page, given book in context) and clearly allows for use of the definite, despite contextual non-uniqueness. We conclude that our stimuli do satisfy the intended constraint.¹⁶ The upshot of these post-hoc analyses is that the findings from the main analysis are robust. The entity relatedness effect on bridging (as measured via context~response similarity) is consistently significant, and, crucially, no evidence of an independent effect of definiteness emerges.

4.5 Discussion

As we argued earlier, an independent effect of definiteness on bridging would support two-phenomena approaches, which take there to be a specific definiteness-driven phenomenon of bridging distinct from coherence-driven bridging. In contrast, the absence of such an effect supports a one-phenomenon approach like that of Asher and Lascarides (1998), which takes there to be a single, broad phenomenon which encompasses bridging of both definite and indefinite NPs. Experiment 1 shows no independent effect of definiteness, supporting the one-phenomenon view. The experiment suggests that in interpretation, recognition that a newly mentioned entity is highly related to some entity already in the discourse model will lead to a bridged interpretation of the new noun, regardless of definiteness marking. It suggests further that bridging is available as a way of satisfying the uniqueness/familiarity requirements of a definite only in the specific case where bridging is already supported by contextual factors.

Besides the findings related to context~response similarity, we saw that the continuation~response similarity scores show a marked increase for low-related NPs, and particularly for indefinites in that condition. Some cases of high continuation~response similarity may be coming from participants who have not bridged the target noun. Faced with the “Which X?” question, a non-bridging participant has two choices: to invent properties of an imagined brand-new referent, or to use material from the continuation sentence. The observation that particularly high rates of continuation~response similarity occur with low-related indefinites may reflect the fact that no feature in the context sentence inclines the interpreter towards a bridged interpretation. And, in fact, the effect of indefiniteness here is quite weak. When we compare the effect of Relatedness on bridging from the context~response similarity analysis with the Relatedness × Definiteness interaction in the continuation~response similarity analysis, the former represents a difference of at least 0.14 in the 0 to 1 scores (see Table 4), whereas the latter represents a difference of at most 0.04 (def vs. indef in the low-related condition, see Table 5). We take this to suggest that the extent to which bridged interpretations are sensitive to relatedness (with no influence from definiteness) far outweighs the extent to which definiteness influences participants’ strategies in potential cases of non-bridging in the low-related condition.

Before concluding Experiment 1, we turn to a possible alternative explanation of context~response similarity. It is possible that high string overlap between the participant’s response and the context sentence occurs in the high-related condition even in the absence of bridging, simply because of the high relatedness of the target noun and its potential anchor. For example, high overlap could be driven by responses like the (non-attested) Speaker A response in (18). Here, the word restaurant is used, not because the context provides it as the anchor for a bridge, but because restaurants are common places to find waiters.

In order to rule out this alternative explanation for the high similarity scores between the context sentence and the response, we conducted an additional check, with the aim of ascertaining, for example, how often participants use restaurant in a response about a waiter, even if the context sentence doesn’t contain the word restaurant. For each target noun, we checked the frequency of occurrence of the high-related anchor noun (e.g., restaurant for waiter, classroom for desks, arrangement of fruit for banana, room for window, wedding for guests) in responses in both the high-related condition, in which the high-related noun occurred in the context sentence, and in the low-related condition, which did not contain the high-related noun in the context sentence. Table 6 reports the by-condition averages of the rate of mention of the related word for each target NP.

While participants do occasionally use the related word in their responses even when it has not occurred in the context sentence (i.e., in the low-related condition, where the use rates are 0.09 and 0.08), they do so at a much higher rate when the related word is present as a candidate anchor in the context sentence (i.e., in the high-related condition, where the use rates are 0.70 and 0.68). An analysis of the binary outcome of the mention of the related word shows only a main effect of Relatedness (B = 4.844, SE = 0.506, z = 9.581, p < 0.001) and no effect of Definiteness nor an interaction (p’s > 0.4). Therefore, our assumption is that the mention of the related noun, which contributes to the high context~response similarity in the high-related condition, likely reflects bridging, rather than coincidental use of a noun from the context sentence.

The proportions in Table 6 also address another open question about our results. High context~response similarity can, in principle, be achieved without actually mentioning a contextually given high-related noun. For example, returning to our restaurant/waiter prompt, both of the (constructed, not attested) responses shown in (19) have high similarity to the context sentence, yet only the (19a) response utilizes the high-related noun from the context sentence.

If responses to definites and indefinites systematically varied in this way, we would have clear evidence that definiteness is affecting the way in which participants bridge. However, Table 6 shows that definiteness does not affect the rate of use of the high-related noun. So, whether the target NP is definite or indefinite, a high-related noun in the context sentence is equally likely to be used in constructing the response.

The analyses we present here represent a subset of our analyses; we also conducted several other exploratory analyses (e.g., computing scores for a bag of words to which no stemming/lemmatizing was applied or in which we eliminated the target noun). Across these analyses, the effect of Relatedness was robust, whereas the effect of Definiteness was not. We chose to report the current set of analyses, because we believe they represent the best treatment of the data for identifying similarity (stemming, lemmatizing) and because we wanted to be conservative and allow for the possibility of an independent effect of definiteness, if it is present, to be recognized.

In summary, then, this experiment suggests that entity relatedness is a significant predictor of a bridged interpretation, independent of definiteness (as shown in the context~response similarity scores); in particular, in the low-related condition, we do not see an increased rate of bridging of definites over indefinites, as predicted by the two-phenomena view. We, therefore, take our results to support the one-phenomenon view of bridging.

5.1 Participants

In order to create a dataset of 100 monolingual English-speaking participants, we recruited and paid 126 participants via Prolific ($2.50 for a task that was estimated to take 10 minutes). All participants indicated they were monolingual English-speaking US nationals, but 24 subsequently mentioned growing up with a non-English language at home when we asked about their language background. Those 24 were removed, as were a further 2 participants with low accuracy on the comprehension questions. With these exclusions, our dataset consists of 100 monolingual English participants’ reading times.

5.2 Materials

The target items consisted of 40 passages that followed the structure of (20–21). As in Experiment 1, items included both singular and plural target nouns (28 singular, 12 plural). We varied the Relatedness and Definiteness of the target noun, as in the sample item sets in Table 7. Each item consists of a context sentence followed by a continuation sentence. The continuation sentence begins with a determiner-noun sequence followed by a restrictive relative clause. The determiner is either the or a. The following noun either denotes an entity that is highly related to an anchor entity introduced by the context sentence (window/living room) or one that is unrelated to the context sentence (knife/living room). As shown in the sample items, in the critical items, the determiner-noun sequence was always presented without the RC, to allow for an initial reading of this sequence as a complete NP.

Table 7 shows the chunking we used in the self-paced reading paradigm. The context sentence was presented as a single chunk and the continuation sentence had chunks for the determiner-noun sequence, the start of the RC, the bridge-relevant content of the RC, and one or more spillover regions.

The target region (in bold) consists of the portion of the RC that identifies the referent as a brand new item, unrelated to the context sentence. If a participant had already inferred a bridge to the context sentence (e.g., assumed, in the first example, that the window being mentioned was a window in the living room), then this target region would require them to revise that interpretation. For a given item, the context sentence, target, and spillover regions were the same for all conditions, and the only words that varied were the sentence-initial determiner and noun in the continuation sentence.

An additional 24 two-sentence passages were used as fillers. Of these, 8 followed the structure of the target items (context sentence, continuation-initial Det-Noun, RC), but they were set up to ensure that participants couldn’t learn over the course of the experiment that potential NP bridges are always broken. In these fillers, the bridge-compatible noun at the start of the continuation sentence was always followed by a bridge-compatible RC. We also varied the chunking, to help prevent participants from associating the chunking of the target items with a determiner-noun sequence and then a bridge-breaking RC. In half of the 8 fillers with RCs, the context sentence was presented in smaller chunks (to avoid Det-Noun-RC sequences only appearing after a single-chunk context sentence); in the other half, the continuation contained a Det-Noun-RC single chunk (to avoid a pattern where a single-chunk context sentence would always be followed by the separate Det-Noun and bridge-breaking RC). A further 16 fillers contained no RC. For the chunking, half of the no-RC fillers contained a single-chunk context sentence (to ensure that participants could not learn that single-chunk context sentences and sentence-initial Det-Nouns were always followed by an RC); the other half provided additional variability in the chunking patterns, by presenting a multi-chunk context sentence and a single-chunk continuation. See the Supplementary Materials for the full set of items.

5.3 Procedure

From Prolific, participants were directed to a website hosted by IbexFarm (Drummond, 2013) for the moving-window self-paced reading experiment. Passages initially appeared on the screen as a series of horizontal lines, where the line length corresponded to the length of the regions. Participants revealed each subsequent region of the passage by pressing the space bar on their keyboard. Passages were presented non-cumulatively, so that each newly revealed region was the only visible region on the screen.

After a quarter of the items, participants saw verification statements which they responded to by clicking on “TRUE” or “FALSE” with their cursor. They received feedback for incorrect answers only.

5.4 Results

Figure 3 and Table 8 show the means of the raw reading times starting at the sentence-initial Det-Noun in the continuation sentence and proceeding through the RC, the RC_break, and the Spillover. Of note are the slow reading times at the bridge-breaking RC region for passages with a high-related definite NP. Contra the prediction of the two-phenomena view, definites and indefinites show no apparent difference in the low-related condition.

Figure 3: Experiment 2 reading times, calculated as by-participant means with standard error (ms) by condition, from Det-Noun region to the Spillover region.

For the analysis, we use linear mixed-effects regression models (LMER; Baayen et al., 2008), using the lme4 package in R (Bates et al., 2015; R Development Core Team, 2017). Traditional analysis of reading time data involves a series of separate analyses, one for each region. However, such an approach raises the possibility of a Type I error, given the non-independence of reading times at different positions in the same sentence and the use of multiple comparisons. Instead, we first build a single large model that contains fixed factors for Relatedness, Definiteness, and also Region. For interactions that reach significance in this large model, we conduct follow-up analyses. This latter approach limits the number of region-specific analyses by only targeting those whose interaction reached significance in the omnibus analysis. This follows recent work on multi-window analyses (see Grüter et al., 2018, for eye-tracking data, and Rohde et al., 2021, for self-paced reading). We use maximal random-effects structure as permitted by the data (see Barr et al., 2013).

Relatedness and Definiteness are coded as in Experiment 1 (–.5/+.5 for low/high and indefinite/definite), and Region uses the sentence-initial Det-Noun sequence as the reference level. Note that this coding means that an interaction between a particular region and one or more of the manipulated factors signals that the behavior of those factors at that region differs from their behavior at the Det-Noun. For example, the prediction that high relatedness supports bridging would correspond to an interaction between Relatedness and the RC_break region: At the Det-Noun region, high relatedness is expected to yield faster reading times (words that are semantically related to the preceding context are typically read faster), but if high relatedness supports a bridged interpretation, then at the RC_break region, high relatedness is expected to yield slower reading times when a posited bridge is broken.

Table 9 shows the model output for Experiment 2. The main effect of Relatedness indicates that high-related entities yield faster reading times at region’s reference level, i.e., at the sentence-initial Det-Noun. The three main effects of region (Region_RC, Region_RC_break, Region_Spillover) indicate that the reading times in these regions differ significantly from that at the sentence-initial NP, independent of condition (faster at Region_RC, slower at Region_RC_break, and faster in Region_Spillover). As noted above, the role of Region is most relevant to our research question when Region interacts with one or more of the manipulated factors.

The 2-way Relatedness × Region_RC_break interaction is likely driven by the 3-way interaction at that region, which we describe in the paragraph below. In addition, we see a Relatedness × Region_Spillover interaction, whereby high relatedness yields reading times at the Spillover region that differ from those at the Det-Noun region. For this 2-way interaction, we conducted follow-up analyses. At the reference level Det-Noun region, we already know that high relatedness yields faster reading times than low relatedness (see Table 9: beta = –31.61, SE = 15.82, t = –2.00, p < 0.05), whereas at the Spillover region, this difference is neutralized and the high-related condition instead yields marginally slower reading times than the low-related condition (beta = 20.33, SE = 11.60, t = 1.75, p = 0.09).

The 3-way Relatedness × Definiteness × Region_RC_break interaction can be understood to capture the slow mean reading time that is apparent in Figure 3 in the high-related definite condition in the RC_break region. For this 3-way interaction, we conducted follow-up analyses. At the reference level Det-Noun region, we already know that the Relatedness × Definiteness interaction is not significant (Table 9: beta = –21.36, SE = 31.66, t = –0.68, p = 0.50), whereas at the RC_break region, the interaction is significant (beta = 85.12, SE = 41.11, t = 2.07, p < 0.05). A further follow-up of this interaction confirms that it is the high-related definite condition that yields the slowest reading time. Definites are slower than indefinites in the high-related condition (beta = 67.91, SE = 22.61, t = 3.004, p < 0.005), but there is no effect of Definiteness in the low-related condition (beta = –0.90, SE = 18.25, t = –0.05, p = 0.96).

As in Experiment 1, we conducted a series of post hoc analyses (see the Supplementary Materials) on the singular vs. plural subsets of the data and on the relational vs. non-relational subsets. The upshot is that some analyses show the same 3-way Relatedness × Definiteness × RC_Break interaction as in the main analysis, whereas some analyses (particularly in the smaller data subsets with potentially reduced power) show only a marginal 3-way interaction or simply a Relatedness × RC_Break interaction or a Relatedness × Spillover interaction. Crucially, we see no emergence of new effects or interactions with Definiteness.

The uniqueness rating study mentioned in relation to Experiment 1 (see Appendix C) included Experiment 2 items and showed that all but one of these items (chair/leg) met our criterion of plausibility of both unique and non-unique contextual instantiation. The chair/leg item, as noted, involves a robustly relational noun and, hence, is expected to be felicitous when used with a definite, despite contextual non-uniqueness.

5.5 Discussion

The strategy in this experiment is to use reading times at the RC_Break region as an indicator of the presence or absence of (early) bridging. Unlike Experiment 1, which provides evidence of an “all things considered” late-stage interpretation, Experiment 2 provides a window into incremental processing. The findings of this experiment reveal a more complex picture than Experiment 1. With respect to the low-related condition, Experiment 2 adds to the evidence of Experiment 1 that in a discourse situation providing no contextual support for bridging, definites are no more likely to bridge than indefinites. That condition provides evidence that this is not merely an “all things considered” effect, but shows that the initial encounter with a low-related definite does not generate an early attempt to bridge. Thus, there is no evidence that a definite drives a bridging attempt, independently of contextual support for such a reading. However, in the high-related case, the situation is more complex. Although Experiment 1 indicates that interpreters are equally likely to give bridged interpretations to high-related definites and indefinites at a late stage in interpretation, things look different at the earlier processing stage targeted in Experiment 2. Only high-related definites show a slowdown in reading times at the RC_Break, indicating an early commitment to a bridged reading.

This interaction is unsurprising on the one-phenomenon view, which takes bridging to be a general interpretative response to multiple cues provided in the input. Arguably, it is also predicted on the two-phenomena view, because, as noted in Section 1, even those who take bridging to be a response to an apparently antecedentless definite also recognize that bridging is constrained by plausibility, so such definites with an “obvious” anchor for bridging will be more likely to bridge than definites with no such anchor. However, the idea of bridging as definiteness-driven suggests that we are more likely to see an early effect of definiteness, regardless of relatedness, with bridging attempts abandoned as broader coherence considerations come into play in later interpretation. Our data show no such effect, and this is consistent with results from more fine-grained experimental paradigms which we discuss in the following section.