4.1 Syntactic structure

We assume that BN objects occur at their base position within the VP whereas rā objects undergo scrambling (cf. Browning & Karimi 1993; Karimi 2003), cf. (15a,b). Subjects are typically moved to the specifier of TP but may stay within vP, cf. (c).

The non-extended vP forms a maximal prosodic domain, predicting accent on kharid in (15)(a), ketāb in (b) and sag in (c) (cf. (4c), (4b) and (6b), Modarresi 2014). We follow Diesing (1992) in assuming that there is an existential closure over the extended verbal predicate, more specifically over the smallest vP, indicated by “∃” in (15).

Dayal (2011) argues against an analysis of PIN objects as non-scrambled, in-situ objects. However, her motivating examples all involve focus movement, which is different from scrambling (cf. Frey 2010): Focus-moved expressions are interpreted in their in-situ position. In Persian, focus movement does not require RA marking, and the PIN properties like number neutrality and narrow scope, are retained.⁴

4.2 Interpretation in DRT

We will assume Discourse Representation theory (DRT) as in Kamp & Reyle (1993). In DRT, sentences and discourses are interpreted as discourse representation structures (DRSs), pairs of a set of DRs and a set of formulas, represented in boxes. For example, the arguments in (17) introduce the DRs x₁, x₂ anchored to Maryam and a book of cardinality 1, and tense introduces the DR e₁ anchored to an event of x₁ picking up x₂, cf. (20)(a). We ignore the contribution of past tense.

Sentence (18) takes (20)(a) as input, generating a new DRS (20)(b) in which the DR x₂ is picked up, and a new DRs x₃ and e₂ are introduced.

Sentence (19) introduces e₃ and picks up x₁ and x₂ via null anaphora for the subject, as expressed by singular agreement, and the object; cf. the DRS (20)(c), where the contribution of foran ‘immediately’ remains unexpressed

DRSs are structural representations of textual information that have truth conditions. Formally, truth conditions are expressed with respect to models of worlds consisting of a domain of entities and events that have certain properties and stand in certain relations to each other. A DRS is true with respect to such a model if there is a way to anchor all the DRs of the DRS to entities in the model such that all the conditions are satisfied in the model. For the DRS (20)(c) this means that there must be an assignment function g that maps the DRs x₁, x₂, e₁, x₃, e₂, e₃ to entities in the model such that the following holds:

If there is such a function g, the DRS after (19) is true in this model, otherwise false. There can be several such functions, as there might be different books or different friends of Maryam that satisfy the conditions, or different events involving the same participants (we make the somewhat artificial assumption that names are unique).

DRT provides a representation format for donkey sentences in which the DR of an indefinite or eventive expression is introduced in the restrictor of a conditional or a quantifier and taken up in its nuclear scope. An understanding of the anaphoric options in donkey sentences is critical for the proposal to be developed here. Consider the following example (22) and its representation (23)(a).

Quantification introduces a complex DRS condition with the connector ⇒. The DR for the book, x₂, is introduced in the antecedent DRS and can be taken up in the consequent DRS, but not in subsequent sentences. The truth conditions are as follows:

As a consequence of this interpretation rule, the DR x₂ is not available outside the complex DRS condition. However, certain types of anaphoric reference are in fact possible; (22)(a) could be continued as follows:

Kamp & Reyle (1993: 4.1.2) propose that such cases involve a process of abstraction over the DRSs of the complex DRS condition followed by a summation over one discourse referent. Applied to our example, this leads to the DRS (23)(b). The DRS construction rule licenses the introduction of a condition x = Σy [DRS ⋃ DRS′] after a condition DRS ⇒ DRS′, where x is a new DR, [DRS ⋃ DRS′] is the union of the two DRSs, and y is a DR in the discourse referents of this union, with the following interpretation:

This ensures that the DR x₄ is anchored to the sum of all entities d such that d is a book, and there is an event in which Maryam picks up d, and there is a friend of Maryam and an event such that Maryam gives d to that friend. Depending on the model, x₄ can be anchored to one or more books; this is responsible for the number-neutral interpretation. This DR x₄ is introduced in a position that it is accessible for subsequent update, as indicated in the last line of (23)(b).

4.3 The interpretation of BN Objects

Krifka & Modarresi (2016) interpret existential closure by a monadic DRT quantifier ∃. It receives an existential interpretation and disables the regular anaphoric uptake of any indefinite or event expression within the scope of the quantifier. For example, (1)(a) has the syntactic form (26) and is interpreted as (30)(a) below.

A condition of the form ∃[DRS], as it occurs in (30)(a), is interpreted as follows:

For (30)(a), the ∃-condition is true for an assignment g if and only if the following holds:

The conditions x₂ = book-of(e₁) and |x₂| = 1 express that x₂ is the unique single book related to the event e₁. The BN ketāb is interpreted as a functional definite (cf. Löbner 1985), and the cardinality of the DR x₂ is 1, due to the singular feature of the count noun ketāb (the plural form is ketāb-hā). However, number neutrality can be derived from the interpretation of ∃ in (27), which allows for more than one extension of the assignment g.

The sentence (1)(b), Maryam yek ketāb kharid ‘Maryam bought one book’, also is compatible with Maryam buying more than one book. But then the use of the number word yek ‘one’ introduces alternatives like Maryam do-ta ketāb kharid ‘Maryam bought two books’ and triggers the scalar implicature that stronger alternatives could not be truthfully uttered. The DRS quantifier ∃ does not come with alternatives, hence (26) does not have this implicature.

This said, cases in which only one book was bought are semantically simpler than others, inviting an interpretation in which x₂ and e₁ can only be anchored to a single entity and event. This tendency will be counteracted by general expectations, as in sentences like Maryam havij kharid, ‘Maryam bought a carrot/carrots’, as it is implausible, given our world knowledge, that Maryam bought only one carrot.

The interpretation of the BN object similar to a functional definite is made plausible by the existence of weak definites in languages like English, cf. (8), which actually mark such nominals as definites. Is also motivated by language-internal evidence, as a BN that is interpreted outside existential closure typically receives a definite-like interpretation, cf. (1c). In this case, the preceding discourse, situation or background knowledge must contain an entity with respect to which there is a unique entity of the required kind. For example, if a book was introduced before, then ketāb, interpreted outside the scope of existential closure, may refer to that book, as illustrated in (29) and the corresponding DRS in (30b).

Notice that the DR for the object in the second clause is interpreted as the unique book in the antecedent DR x₄, which is the sum of the previously introduced book x₂ and the record x₃. Hence, a uniform interpretation of bare nouns is possible.

Being outside of the scope of ∃, ketāb-rā cannot depend on e₁. But it can be interpreted as relative to the sum of all salient DRs x₁+x₂+x₃, introducing the unique book among them, x₂, and identifying it with the new DR, x₅. In general, RA-marked bare nouns are interpreted like functional definites that identify the unique entity in their reference situation or in the world knowledge assumed to be shared by the participants of conversation.⁵ When applied to a specific DR, the possessive form, e.g. jeld-esh/-eshoon ‘its / their cover’, cf. (22b,c) is used.

Consider now the option for anaphoric uptake after (26), as in (31):

The DRs introduced in the scope of existential closure like x₂ in (30a) are not immediately accessible. But the abstraction and summation rule (25) applies, cf. (30c), where x₃ is the sum of all books that were bought by Maryam.⁶

One consequence is that BNs are interpreted as maximal. In (30), the DR x₃ is anchored to the union of all books that Mary bought, within the described discourse universe. This might appear surprising, but maximality effects of pseudo-incorporated nominals and weak definites have been noticed in Dayal (2011) and Schwarz (2014). Maximality can be detected in the contrast between indefinite object and BN in example (32).

The indefinite antecedent ye khaneh (with unstressed, non-numeral reading of yek) is not interpreted as implying that Ali has only one house, whereas the BN object implies reference to the sum of all houses that Ali has.

4.4 The interpretation of YK objects and of complex predicates

YK objects are regular indefinites that are not dependent on another DR. As Fodor & Sag (1982) showed, indefinites may scope outside of island. This means that the DR of an indefinite can be introduced in the highest DRS (cf. Kamp & Reyle 1993: 288ff), leading to the following two interpretations of examples like (33):

The readings are truth-conditionally equivalent but Reading 2 introduces an accessible DR, x₂. Notice that Reading 1 could be expressed by the simpler clause with a BN object, (30a). Hence, we should expect that Reading 1 tends to be blocked by the BN clause, following considerations of economy similar to (Fox 1998): Existential closure ∃ just expresses the existence of an extension of the variable assignment that verifies the embedded DRS, not the existence of precisely one such extension. Hence the number restriction |x₂| = 1 expressed by yek is uninformative. Hence even in the absence of RA marking, indefinite singular objects are rather interpreted following Reading 2, resulting in an accessible DR x₂. This prediction will be put to test in Section 5.

BN objects can also be interpreted as parts of complex predicates, which have an idiomatic interpretation. Regular BN objects and BN objects in complex predicates are syntactically distinct (cf. Modarresi 2014; Megerdoomian 2012). One difference is that BNs in complex predicates do not allow for anaphoric uptake:

(34) Maryam Maryam divar-rā wall-OM rang zad. color hit.3SG ‘Mary painted the wall.   #Gheimat-e-sh   price-of-it geroon expensive bood was.3SG ‘Its price was high.’   Gheimat-e price-of rang color geroon expensive bood was ‘The price of the paint was high.’

The expression rang zad ‘paint hit’ is a complex predicate with the meaning ‘paint’. The BN rang ‘color, paint’ cannot be taken up directly, different from BN objects considered so far. What is possible, however, is to refer to the paint by the referring expression description gheimat-e rang ‘the price of the paint’. This is typical for associative or “bridging” definites (cf. Charolles 1999), which is licensed here as painting events are associated with paint. This is captured in the DRT representation above as follows: First, a DR for Mary, x₁, and for the wall in the given situation s, x₂, is introduced. The BN of a complex predicates refers to the object related to the event but does not introduce a DR. The complex predicate introduces an event variable, subject to existential closure as usual. The subsequent sentence can pick up this event via abstraction and summation, and the associative definite can introduce a DR that is uniquely related to that event.⁷

4.5 Predictions concerning anaphoric accessibility

Our modelling of BN objects and YK objects yields predictions about their anaphoric potential, under the assumption that directly introduced DRs with YK objects are more salient than DRs with BN objects that are introduced by abstraction and summation. Unfortunately, there is little empirical work that deals with summation anaphora (except for so-called complement anaphora, cf. Nouwen 2020 for overview). However, it is plausible that the more complex construction of DRs should reduce the availability of such antecedents. Hence hypotheses BN-0 and BN = YK, cf. (14), should be rejected, supporting BN < YK in (35), our main hypothesis. As BN antecedents do not denote implicit participants that would be only accessible by associative anaphora, hypothesis BN-DD is rejected as well.

The predictions concerning hypotheses BN-NL! and BN-NL, that BN antecedents require or prefer null (NL) anaphora, is a more complex issue. Kamp & Reyle (1993) distinguish between atomic and non-atomic DRs, where atomic DRs can be picked up by SG anaphora, and non-atomic DRS by PL anaphora. Under the assumption that YK antecedents introduce atomic DRs, they should be incompatible with PL anaphora, and should prefer SG over NL anaphora as the presupposition of SG anaphora are satisfied, following the maximize presupposition principle, cf. hypothesis YK-SG.

As for BN antecedents that involve abstraction and summation, their DRs could be either ambiguously singular or plural, or else vague, assuming number-neutral DRs (a notion also proposed in Kamp & Reyle 1993). The type of DR depends on whether the antecedent clause invokes a single object or multiple objects (‘buy boat’ vs. ‘buy button’), or whether it lacks such biases (‘buy book’). We tried to construct examples without bias for the experiments to be reported here. This predicts that BN antecedents should be picked up by NL anaphora, leading to hypothesis BN-NL in (14). However, under the single anchor preference of Section 4.3, we should expect that BN antecedents tend to introduce atomic DRs, leading to a counteracting preference of SG anaphora. As we cannot estimate the strength of the tendencies for NL vs. SG anaphora, BN-NL/SG just states that both are allowed equally; hypothesis BN-noPL states that PL anaphora are avoided.

As suggested by an anonymous referee, hypothesis BN-noPL could be due to the singular number feature of BNs. However, then we should see no difference to the singular YK nouns. Hypothesis BN>YK-PL rejects this explanation.

5.1 Previous studies

There are a number of experimental studies on anaphora in the context of incorporation structures. In an early study, Ward et al. (1991) investigate the anaphoric potential of morphological incorporation (e.g. deer hunting vs. hunt deer) with self-paced reading and find that participants read subsequent sentences with pronouns referring to deer faster in the second condition.

Scholten & Aguilar-Guevara (2010) research the behavior of BNs, weak definites, regular definites and indefinites in Dutch. Participants had to choose between a pronoun or a definite containing the same noun as the antecedent NP. Regular indefinites were taken up far more often by pronouns than by regular definite NPs.

Oggiani (2011) looks at BNs vs. regular indefinites in Spanish, applying a technique in which participants are asked to form small texts containing constituents like tener casa ‘have house’ and tener una casa ‘have a house’. She finds that the object is picked up about only half as often when it is realized as a BN.

Law & Syrett (2017) investigate the anaphoric uptake of BNs vs. overtly marked singular or plural nouns in object position in Mandarin (singular nouns with number word ‘one’ plus classifier, plural nouns with number word ‘three’ plus classifier). Referring to Modarresi (2015), they used stimuli where world knowledge was either not biased or biased towards single or multiple entities. The second sentence contained a singular pronoun or a plural pronoun in subject position, immediately following the object in the preceding clause. An online self-paced reading task showed that anaphoric uptake of (singular) indefinite antecedents is processed more easily than uptake of BN antecedents.

5.2 Experiment 1: Self-Paced-Reading

Inspired by Law & Syrett (2018), we measured the ease of anaphoric uptake for BN vs. YK objects by self-paced reading. This taps directly into the processing of read language and should give us reliable results about the grammatical representation.

We constructed antecedent sentences with the intention to avoid bias towards a singular or a plural interpretation of the BN object and investigated cases in which the subsequent sentence had a null (NL), singular (SG) or plural (PL) anaphor, resulting in a 2 × 3 design (2 antecedents, 3 anaphora). A sample item in six conditions is presented in (36).

As for the hypotheses in (14), BN-0 predicts slowing down with BN objects whereas BN = YK predicts no such slowdown (rather a boost as BN objects are shorter than YK objects). BN-NL! would be supported if BN followed by NL is processed faster than BN followed by SG/PL, in case this difference does not obtain for YK followed by NL vs. YK followed by SG/PL.

64 native Persian speakers participated in an experiment designed with Ibex-Farm, both online and in the lab. There was no significant difference between the two groups; we report the overall results for both groups. The stimuli consisted of 48 sentence items similar to (36). We tried to construct these sentences with frequent words that are encountered in every-day Persian conversation.¹⁰ Comprehension questions after more than half of the trials checked for attention of the participants. One example of a comprehension question following (36) is Ki ketāb kharid? ‘Who bought a book/books?’ followed by a multiple-choice task. In the items, the anaphoric element does not follow the antecedent immediately and is not realized in the beginning of the sentence. Furthermore, antecedent and item were always in object position. There were 48 trials per person. Trial types were presented in a Latin-square design with 6 conditions of similar structure in six different lists. Each participant saw each sentence under just one condition.

Participants read the sentences fragment by fragment in Persian orthography, from right to left, by pressing the space key. Response times between key presses were recorded. Participants answered a comprehension question, with a high accuracy rate (mean = 94.23%), indicating that they paid attention.

We had to delete data from six participants because their mean reading times were suspiciously slow or high. As usual, the measured reading times had a thick right tail, so we used a boxcox transformation to get a normally distributed response variable. A log likelihood profile (R-package) showed the best value for the boxcox parameter to be lambda = 0.4. The results are shown in Figure 1 from the word before the anaphoric expression, at the word containing the anaphoric expression, and the following two words.

Figure 1

Results of self-paced reading test.

We saw an effect of word length of the anaphora (NL < SG < PL) across BN/YK antecedents. The difference between BN followed by SG vs. YK followed by SG was not significant. We concluded that self-paced reading is not a suitable task for investigating the processing of anaphoric reference, at least not in the setup we have chosen.

5.3 Experiment 2: Acceptability Judgements

In a second experiment, we used the same setup and materials as in Experiment 1, except that the reading tasks were followed by an acceptability question on a Likert Scale from 1 (poor) to 7 (good). We wanted that the self-paced readers pay more attention to the content of the items and have additional acceptability data. There were 38 participants, different from the ones in Experiment 1, partly in a lab and partly on-line. We got comparable results as in Experiment 1 for the reading times and do not report them here. As the ratings of participants differed widely (some judging sentences generally better, others generally worse), the judgements of each participant were z-transformed. Figure 2 gives a whisker diagram of the z-transformed data (e.g., +1 indicates one standard deviation difference from the mean); Figure 3 then indicates the means of the indicated continuations. The cases are presented in the same order, where NL/SG combines the NL and SG cases. Cases with BN antecedents are given in white, cases with YK antecedents are given in grey; the order or presentation is the same in the two figures.

Figure 2

z-transformed ratings, with means indicated by ×; y-axis: standard deviations from mean; YK antecedents in grey.

Figure 3

Means of z-transformed ratings, relevant significant differences indicated; y-axis: standard deviations from mean; YK antecedents in grey.

Starting with the non-PL cases, we find that anaphora to BN antecedents were generally rated worse than anaphora to YK antecedents (BN-NL/SG vs. YK-NL/SG, non-paired t-Test assuming equal variances: p < 0.005), contra hypothesis BN = YK. When comparing the non-transformed averages of each participant across items using a paired t-test, the difference is also significant (p < 0.05). The means of BN-NL/SG was positive, contra hypothesis BN-0. Together, this result supports the main hypothesis, BN < YK. BN-NL vs. BN-SG is not significant, consistent with hypothesis BN-NL/SG; YK-NL vs. YK-SG is approaching significance (p = 0.064), supporting hypothesis YK-SG. As for the PL cases, notice that they both are judged negatively (significant differences to all non-PL cases, not indicated in graph), supporting hypothesis YK-SG and BN-noPL. In addition, YK-PL is judged more negatively than BN-PL (p < 0.001), supporting hypothesis BN>YK-PL, even though we found it surprising how negatively BN-PL cases were actually judged.

5.4 Experiment 3: Anaphora Choice

In the third experiment we tested the anaphoric potential of BN and YK objects in a forced-choice selection of anaphoric expressions, a controlled production experiment. This should reflect linguistic competence more directly than a rating experiment. We wanted to test whether there is a requirement or preference of BN antecedents to select for NL anaphora, as claimed in hypotheses BN-NL! and BN-NL, or no such preference, after hypothesis BN-NL/SG.

Participants were presented with a sentence containing a BN or YK object, and a continuation sentence followed by a blank to be filled with multiple choices of NL, SG, and PL anaphora. As a sample item, consider (37).

We constructed 36 test items with two conditions and no obvious bias to singular or plural interpretation, including 8 fillers. There were 153 native Persian speakers that voluntarily participated in this experiment using an online survey platform, answering an average of 16 questions. The stimuli were presented in six different lists (participants also participated in Experiment 4, but it was made sure that no participant saw the same sentence twice). Each list included both conditions, with an average of four fillers; the items were randomized using Latin square design. The results are indicated in Figure 4.

Figure 4

% Anaphora choice (NL/SG/PL) dependent on given antecedent (BK/YK).

BN antecedents are taken up by NL or SG anaphora about equally often, but significantly less often by PL anaphora, supporting hypothesis BN-NL/SG contra hypotheses BN-NL! and BN-NL (Chisquare test, p ≈ 0). The infrequent uptake by PL anaphora supports hypothesis BN-noPL. In comparison, YK nouns are significantly more frequently taken up by SG than by NL anaphora, supporting hypothesis YK-SG. PL anaphora are taken up quite rarely. If they are taken up, then only with BN antecedents, supporting hypothesis BN>YK-PL.

5.5 Experiment 4: Antecedent Choice

Experiment 3 did not show whether the participants favored the use of BNs as antecedents of anaphora because the antecedents were given. We reversed the design and investigated the forced choice of antecedents (BN vs. YK object), when the anaphor in the subsequent sentence is given (as NL, SG or PL). Otherwise, the stimuli were the same as in Experiment 3.

There were 30 items with three conditions, including 6 fillers. The same 153 native Persian-speakers as in Experiment 3 participated in this experiment as the second part of the experiment. The stimuli were presented in six different lists. Each list included all the three conditions in randomized order, including an average of four fillers in each list, using a Latin square design. After reading the whole sentence, the participants had to choose whether the BN or the YK noun was the most appropriate antecedent. Results are presented in Figure 5.

Figure 5

% Antecedent choice (BN/YK) dependent on given anaphora (NL/SG/PL).

This experiment shows that YK objects make better antecedents, except in the case of a PL anaphor, which leads to a semantic clash. We were surprised that YK objects were selected at all. The reason could be that YK objects in general make better antecedents, but we suspect that the task of going back in the text, choosing an antecedent, reading the text, choosing another antecedent, reading the text in this version, and selecting the better option was quite complex, at least for some participants.¹¹ Focusing on the non-PL cases, the experiment clearly rejects BN = YK, that BN and YK nouns are equal in their anaphoric potential (Chisquare, p < 0.001). It definitely rejects hypothesis BN-0, the widespread assumption that BN cannot serve as antecedents.¹² thus supporting the main hypothesis BN < YK.

5.6 Experiment 5: Free Sentence Completion

In a final experiment, we investigated which anaphoric forms are used by participants spontaneously in a context favoring anaphoric uptake, contrasting BN and YK-marked antecedents. This task does not investigate the reflection of participants about language, but asked that the participants perform a natural linguistic activity. In particular, it also leaves open the option of no anaphoric uptake at all. A sample item is (39).

(39) does not have a specific bias towards a singular or plural interpretation, as it is as likely to read one or more than one newspaper in a library. The stimuli consisted of 24 items with two conditions, randomized in a Latin Square design in different lists. There were altogether 252 participants that took part in an online experiment. Participants read an average of three sentences (in the neutral context; there were also other items for singular and plural contexts, as well as fillers) and were asked to type a suitable continuing sentence. We collected 754 test items after exclusion of incomplete answers. Every sentence was analyzed separately to see if and how the participants referred back to the antecedent object noun. Naturally, there was a greater variety in the anaphoric responses. The results in Figure 6 visualize NL anaphora, singular anaphoric reference with pronouns or clitics (Pro-SING), singular anaphoric reference will full DPs (Full DP-SING), plural anaphoric reference with pronouns or clitics (Pro-PLUR) and plural anaphoric reference with full DPs (Full DP-PLUR). Associative plurals and reference to kinds were very rare and are not reported here.¹³

Figure 6

Free completion task; x-axis: anaphoric device.

We see that BN objects are taken up by anaphora more frequently than not (BN objects are not taken up only about 36% of the time), clearly disproving hypothesis BN-0. But YK objects are taken up more frequently than BN objects, arguing against hypothesis BN = YK. However, this difference is not quite significant (Chisquare: p < 0.07). Together, this is strong support of the main hypothesis, BN<YK. Anaphoric uptake of BN objects is rarely by full DPs, against the associative anaphora hypothesis, BN-DD. BN antecedents are about equally often picked up by NL and SG anaphora, disproving hypotheses BN-NL! and BN-NL and supporting hypothesis BN-NL/SG. Interestingly, YK antecedents are about equally taken up by NL and Pro-SING antecedents, not supporting the second part of YK-SG, in contrast to experiment 3. PLUR anaphora occur very rarely, supporting hypothesis BN-noPL; the few cases are anaphoric to BN antecedents, supporting hypothesis BN>YK-PL.