2.1 Theoretical considerations

Jackendoff (1985) observes that there is a substantial class of CEPs in English that alternate between a CC expressing a belief and a CC expressing an intention, like in the examples in (3), which can be paraphrased as in (4):

The same pattern is also observed, for example, with decide (ibid.: 447; Grano 2019: 126):

Since both uses of convince (and decide) are relatively close in meaning and, moreover, this alternation is found systematically with a substantial class of verbs,¹ Jackendoff argues against attributing the alternation to an idiosyncratic polysemy of the verb convince (ibid.: 446–447). The fact that similar alternations occur in different languages all over the world, as we will see in the next section, also corroborates this argumentation. Furthermore, Grano (2019: 128) gives the zeugma test in (6) as evidence against persuade being polysemous: if beliefs and intentions were different senses of a polysemous persuade, (6) would sound funny, similar to John runs marathons as well as a successful company. Since (6) is perfectly fine, persuade seems to be underspecified for the belief/intent distinction rather than polysemous.

In order to obtain one or the other interpretation, the type of CC seems to be crucial: a belief is expressed by a finite CC and an intention by a to-infinitive. Although a close paraphrase of an intention like in (7a) by means of a finite CC as in (7b) is possible, they do not express exactly the same thing, as the non-contradictory nature of both (7c) and (7d) shows.

In fact, the closest paraphrase of a non-finite CC by means of a finite CC in this context still expresses a belief: Mary caused John to form the belief “I should leave”. In praxis, beliefs about what one should do commonly lead to the corresponding intentions, but this is not inevitably the case. (7b) may thus implicate (7a), but it does not entail it (see Grano 2019: 125). Conversely, (7d) shows that it is also possible to form an intention without the corresponding belief, i.e. a person might also act contrary to their beliefs about what they should do.

However, it would be incorrect to think that in English finite clauses always express beliefs, whereas non-finite clauses always express intentions. With other CEPs, both CC types can be synonymous (Jackendoff 1985: 447):

Jackendoff (1985: 457) therefore divides CEPs that take both finite and non-finite CCs into three classes: those that always express beliefs, like claim, those that always express intentions, like promise and those that alternate, like convince. Thus, there is no one-to-one mapping between form and meaning of the CC but the meaning of a certain type of CC depends at least partly on the CEP. There are, however “marked” and “unmarked” or typical and atypical associations: in English, the typical realization of a proposition is a finite clause and the typical realization of an action is a non-finite clause. Consequently, CEPs realizing a proposition complement as a non-finite CC only should be rare or non-existent, just like CEPs realizing an action complement as a finite CC only. If one verb can have both a proposition and an action complement, the former will be realized as a finite CC, the latter as a non-finite CC (ibid.: 457–458).

Grano (2019), focussing on the CEP persuade, aims to develop a formal semantic model that can explain the semantic link between belief and intent readings as well as the crucial contribution of the distinction between finite and non-finite CCs in English. He analyses the verb persuade as meaning ‘cause to have a rational attitude’, with rational attitudes encompassing both beliefs and intentions. As for the CC, he postulates that English infinitival CCs that allow the addition of a subject marked with the preposition for (like John intends for Bill to be happy) express preference modality, whereas finite clauses express doxastic modality (ibid.: 129). Although he does not give any reasoning for this, he takes the infinitival CC of persuade to be of the for-to type, too (ibid.: 132). Thus, when combining the semantic contributions of the CEP and the CC, this yields the observed pattern: a rational doxastic attitude is a belief and a rational preference attitude is an intention (ibid.: 131–132).

On a more general level, beliefs and intentions fall into the two main semantic classes of attitude verbs: representationals (Bolinger 1968; Anand & Hacquard 2013; Hacquard & Lidz 2019) or attitudes of acceptance (Stalnaker 1984), which express a judgment of truth, and preferentials, which express a preference (Hacquard & Lidz 2019: 84, among others). The split between these two classes is reflected in the morphosyntax of the respective CCs in Romance languages, English, German and Mandarin in that representationals take CCs with the morphosyntactic properties of declarative main clauses, while the CCs of preferentials differ from declarative main clauses in mood, word order, finiteness or the availability of (epistemic) modals, aspect markers or so-called “root transformations” (Hooper & Thompson 1973; Anand & Hacquard 2013; Hacquard & Lidz 2019; Huang et al. 2022). Using the terminology of Cristofaro (2003), representationals take balanced CCs and preferentials take deranked CCs. The same pattern can be observed in CEPs from rather diverse languages that alternate between a belief and an intent reading, as we will see in the next section.

Thus, to sum it up, there seems to be a substantial class of attitude verbs in English that are underspecified for the distinction between representational and preferential attitudes and receive this specification from their CC, building on a more general association between form and meaning of CC types.

2.2 Cross-linguistic patterns

CEPs alternating between a belief and an intent reading are found not only in English but in several languages all over the world and it is striking that, although there is a great diversity in the structures employed, in all the examples given in (10)–(15), belief is expressed by means of a balanced or at least tense/aspect-marked CC, whereas a deranked CC that lacks such marking (except for the perfective marking on the Russian infinitive in (13b)) expresses intent.

In Basque in (10), for example, a belief complement of the CEP konbentzitu ‘convince’ is expressed by means of a finite clause inflected for TAM and the two core arguments, just like an independent clause. The intent reading, on the other hand, is expressed by a VN in the purposive, which is not inflected for any verbal categories and could not be used as the predicate of an independent clause.

Rapanui, too, uses a CC in the form of an independent clause (Kieviet 2017: 523) for the belief reading in (11a) and a CC without any aspect marking bearing the subordinator mo, which also forms purpose and conditional clauses (ibid.: 542), for the intent reading in in (11b).

Similarly in Yimas, a belief complement of kacapal- ‘forget’ is expressed by means of a juxtaposed independent clause in (12a), and the intent reading in (12b), on the other hand, by a nominalized verb bearing the noun class marker of the TALK class:

Russian at first sight shows the same pattern, a balanced CC for the belief reading in (13a) and an infinitive for the intent reading in (13b), but it is conspicuous that the intent complement contains perfective marking, which sets the Russian example apart from the aspectless intent CCs in the other examples. Aspect in Russian (or Slavic languages in general) is, however, closer to derivation than to inflection (“grammatical derivation” according to Breu 2000: 23) and has thus to be analysed differently from purely inflectional aspect marking.

In Tommo So in (14), while the intent reading with the CEP ‘be afraid’ is expressed by a verb form bearing only the infinitive ending, the belief reading is expressed by a more complex construction: the CC is inflected for aspect and person and has in fact the same form as an independent polar question (McPherson 2013: 382, 453). It is embedded as a complement of the verb gɛ̀ ‘say’, which in turn functions as a medial verb of a verb chain with the verb níŋ-íyɛ́ ‘be afraid’ as the final verb. McPherson’s consultants analyse this structure as follows: “gɛ̀ lends a sense like ‘believe’, as in, ‘Believing he would hit me, I was afraid’” (ibid.: 454).

In Hinuq in (15), both CCs are deranked, i.e. their predicate is in a form that is different from the predicate of an independent clause, but still they differ in the presence vs. absence of tense and aspect marking: whereas the intent reading in (15b) is expressed by a purposive converb that is only marked for gender, in the belief complement in (15a) the resultative participle expresses relative past time reference, contrasting with the habitual participle used for relative future, present and generic time reference (Forker 2013: 608). The construction consists of an auxiliary and an imperfective participle and thus corresponds to the “compound past” of independent clauses (ibid.: 216).

The verb muʔnima in Yakkha (Sino-Tibetan, Kiranti; India, Nepal) also seems to follow the same pattern: it is listed as meaning ‘forget to do’ with an infinitival CC and ‘forget about something’ with an inflected CC in Schackow’s (2015: 465) grammar.

This differential treatment of belief and intent CCs aligns not only with the observation that representational attitudes are associated with balanced CCs and preferential attitudes with deranked CCs, but also with the observation that the morphosyntactic complexity of CCs in languages roughly corresponds to their semantic complexity (e.g. Givón 1990; Cristofaro 2003; Wurmbrandt & Lohninger 2023). Beliefs are Propositions, the most elaborate complement type in Wurmbrandt & Lohninger’s (2023) classification, which bear a truth value and have independent time reference. Intentions, on the other hand, are Situations: they can be located in a time different from that of the matrix clause, but their time reference is predetermined by the CEP and they lack a truth value (ibid.: 188–189). Since balanced CCs are usually more complex than deranked CCs, the two morphosyntactic classifications of CCs overlap, but they should not be conflated: balancing vs. deranking is a binary distinction, whereas morphosyntactic complexity is a scale. For example in Hinuq in (15), both CCs are deranked, but they still differ in morphosyntactic complexity. The question remains open to what extent the morphosyntactic distinction between beliefs and intentions is based on the structural complexity of the CC type or its (non-)correspondence to independent main clauses. While the Hinuq example, where the CCs differ only in complexity, points to the former, Basque offers evidence for the latter, as we will see in Section ‎5.3.

Hacquard & Lidz (2019) hypothesize that the correlation between form and meaning observed in the CCs of attitude verbs is an important cue for children acquiring the abstract meanings of these verbs. Apparently, it can also be exploited to use the same CEP to express both representational and preferential attitudes.

However, the division is not always as neat as it appears in the data presented so far. In fact, Basque, like English, has two types of CCs which can in principle occur expressing either beliefs or intentions, namely indicative clauses and VNs in the absolutive or instrumental. Although the most common association is for indicative clauses to express beliefs and for VNs to express intentions, as illustrated in (2) above, both are also found with the respective other meaning: (16) is an example of a VN belief, whereas (17) and (18) are indicative intentions. Note that indicative clauses expressing intentions have a fixed TAM marking: with subject control CEPs like in (17), the CC obligatorily contains future marking, whereas with an object control CEP like in (18), it contains the modal behar izan ‘must’.

Unlike what Jackendoff (1985: 457–458) claims for English, in Basque VN beliefs and indicative intents seem to occur also with alternating CEPs: (19a) shows ukatu ‘deny’ with an intent CC and (19b) and (19c) with belief CCs of two kinds, the first one a VN² and the second one an indicative clause. (20a), on the other hand, shows bururatu ‘occur, come to mind’ with a belief CC, (20b) with an intent CC in the form of a VN and (20c) an indicative CC that is presented by Artiagoitia (2003: 664) as a “paraphrase” of (20b), leaving it open to what extent they are in fact synonymous.

The use of the same CC type in different semantic contexts does not contradict theories of a form-meaning correlation in terms of complexity, since a CC type that is suitable to express Propositions is not necessarily banned from expressing semantically less complex complements (Wurmbrandt & Lohninger 2023: 215), although it is certainly an interesting fact that a CC type that can in principle be used to express Propositions might still more typically be used with less complex complements, as is the case with Basque VNs and English infinitival complements. The existence of some atypical uses of CC types is also unproblematic for Hacquard & Lidz’s (2019) hypothesis that children’s acquisition of attitude predicates is guided by morphosyntactic form, as long as the overall tendencies are clear. The pattern does, however, at least at first sight, pose a problem for compositional analyses of alternating CEPs that locate the meaning difference in the type of CC (see Yan 2025a: 20–26 for an overview).

However, a difference in morphosyntactic form can also manifest itself below the level of CC type: in Mandarin, the CCs of representational and preferential attitude verbs differ in the presence or absence of overt subjects, modals and aspect markers (Huang et al. 2022), and Yan (2025a) similarly lists several elements inside the CC like TAM and particles that can affect the reading in what she calls information-preference and information-ability alternations across different languages.

Thus, two questions arise: 1. How acceptable are indicative intents and VN beliefs in general, do they occur freely or are they limited to specific CEPs? 2. If the same CC type can express both beliefs and intentions, which other factors might be relevant for disambiguating between the readings instead of CC type?

In the following section, we will therefore have a look at the CC types used with CEPs expressing beliefs and/or intentions in Basque, semantic factors distinguishing beliefs and intentions, as well as the tendencies in form-meaning mapping that can be observed in corpus data.

3.1 CC type

As we have seen in the section before, the form of the CC seems to be a major factor cross-linguistically in distinguishing between belief and intent readings.

In Basque, the four major types of CCs that are found with belief/intent CEPs are finite indicative CCs, finite subjunctive CCs, VNs inflected in the absolutive or instrumental and VNs inflected in the purposive.

Finite indicative clauses, the only balanced type, consist of a clause in an indicative-like³ mood, inflected for mood, tense and aspect as well as up to three arguments, plus a complementizer. The clause has the same form as an independent clause except for the complementizer that is suffixed to the predicate:

Finite indicative CCs are used mainly with utterance, perception, epistemic and other representational attitude predicates like esan ‘say’, entzun ‘hear’ or uste izan ‘think’.

Finite subjunctive clauses are also inflected for up to three arguments, but not for aspect. Their mood value is fixed and the tense value agrees with the tense of the matrix clause, as shown in (22). Subjunctive CCs are primarily used with desiderative and directive CEPs (Artiagoitia 2003: 640).

VNs are formed in Standard Basque by adding the suffix -t(z)e to the verb stem. They can be inflected in all cases and have slightly different syntactic properties depending on the case value. Absolutive VNs are probably the most versatile CC type in Basque. They are found in a wide variety of contexts ranging from utterance predicates as in (23) to implicative predicates as in (24), but seem to express most typically complements that are semantically less complex than Propositions (see Artiagoitia 2003: 661–665).

Instrumental VNs can be seen as a variant of absolutive VNs, since the instrumental is used for demoted absolutive arguments in “lexically constrained antipassives” (Zúñiga & Fernandez 2021: 629–635). This is demonstrated in (25) with the verb gogoratu ‘remember’, which can be used in a transitive construction with an ergative experiencer and an absolutive stimulus as in (25a) or in an intransitive antipassive-like construction with an absolutive experiencer and an instrumental stimulus as in (25b). Instrumental VNs are, however, by far less common than absolutive VNs and might differ slightly in the kinds of semantic contexts they appear in.

Absolutive and instrumental VNs can distinguish between past/perfective and non-past/imperfective (Artiagoitia 2003: 666). The past/perfective is formed by combining the perfective participle of the main verb with the perfective participle of izan ‘be’, which is then inflected like a noun, as can be seen, for example, in (20b) above. As (23) shows, absolutive and instrumental VNs do not impose any restrictions on the overt expression of their subjects. Purposive VNs, on the other hand, typically do not appear with overt subjects, which is, however, due to semantic reasons since they are primarily used with directive CEPs, which usually exhibit object control (Artiagoitia 2003: 704). They cannot distinguish tense or aspect but can be said to express prospective aspect, as they are used exclusively for events that are to happen after the matrix event (Euskaltzaindia 2021: 1227).

While subjunctive clauses and purposive VNs are restricted to intentions, absolutive and instrumental VNs as well as indicative clauses can in principle appear expressing both beliefs and intentions, as we have seen in Section ‎2. Assigning different modalities to the different CC types, like in Grano’s (2019) analysis of the complementation patterns of English persuade, might thus work well for subjunctive clauses and purposive VNs, whereas with indicative CCs and absolutive and instrumental VNs CC type alone is not decisive.

3.2 Semantic properties of belief and intent clauses

Besides the more language-specific structural factors, there are also language-independent semantic factors disambiguating between beliefs and intentions, namely the more restricted semantic properties of intentions:

1. Intentions are attitudes towards actions (Jackendoff 1985: 451–452).

2. The attitude holder is able to carry out the action volitionally or is at least in some way responsible for its realization (Grano 2017: 595–597).

3. The action is to happen after the moment the attitude holder forms the intention.

This leads to several restrictions on Aktionsart, tense/aspect and agent reference of the embedded clause: the tense/aspect marking has to be compatible with a relative future and/or prospective interpretation and prototypically the predicate is dynamic and agentive with its agent coreferential with the attitude holder. If the latter two criteria are not met, an intent reading may still be possible⁴ as in (26), taken from Jackendoff (1985: 456), which implies that Bill has control over when the addressee receives the letter. In other cases, especially if the CC is marked for another tense/aspect,⁵ an intent interpretation is infelicitous, as in (27)⁶ (ibid.).

There are, on the other hand, no such apparent restrictions on beliefs:

This means that some CCs by virtue of their Aktionsart, tense/aspect value or agent referent are only suitable to express beliefs.⁷

3.3 Patterns in the corpus

In order to get a first impression of the complementation patterns of different belief-intent CEPs in Basque and the factors that might play a role in licensing different constructions, a small corpus study was conducted using data from Egungo Testuen Corpusa (ETC). This corpus contains 355 million tokens from texts published in the 21^st century, mainly newspapers but also novels, non-fictional books and Wikipedia. It has the disadvantage that it does not provide context beyond the sentence, but occurrences from the magazine Argia and sometimes other newspapers can be found in their respective archives and Wikipedia is publicly available.

For the study eight alternating verbs that occur with both indicative clauses and absolutive (or instrumental) verbal nouns were selected: the directive speech act verbs gogorarazi ‘remind’ and konbentzitu ‘convince’, the commissive speech act verb zin egin ‘swear’, the commissive/directive speech act verb proposatu ‘suggest’, three non-communicative verbs referring to the acquisition or existence of beliefs and intentions, namely bururatu ‘occur, come to mind, have the idea⁸’, erabaki ‘decide’, pentsatu ‘think, plan’, as well as the verb ukatu ‘deny’.⁹ The sample was inspired by the verbs listed by Jackendoff (1985), but also includes other CEPs that appear to follow a similar pattern. The details on the sampling and annotation process can be found in Appendix A.

Table 1 and Table 2 present the occurrences of the different CC types with the two readings for each of the eight verbs.¹⁰ Since the focus is here on the two CC types allowing both belief and intent readings, “VN” refers only to absolutive and instrumental VNs¹¹ and all other CC types (mostly purposive VNs, subjunctive clauses and non-finite interrogatives) are classified as “other”. While the directive CEPs gogorarazi and konbentzitu preferably or even exclusively use these unambiguous CC types to express intentions in the sample, other CEPs prefer absolutive VNs or, in the case of zin egin, even indicative clauses for intentions.

Although absolutive and instrumental verbal nouns as well as indicative clauses do occur with both beliefs and intentions, there are very clear preferences. Table 3 combines the totals from Table 1 and Table 2, excluding the “others” column. As Table 3 shows, the overwhelming majority of indicative clauses (namely 85,5%) express beliefs, whereas the overwhelming majority of absolutive/instrumental VNs (93,6%) express intentions and, conversely, the overwhelming majority of beliefs (96,5%) is expressed by indicative CCs and the majority of intentions (76,1%) by VNs. Most cases of VN beliefs and indicative intents are due to one verb each, as can be seen in Table 1 and Table 2: 14 out of 15 VN beliefs are complements of ukatu ‘deny’, while 63 out of 69 indicative intents are complements of zin egin ‘swear’. Ukatu is what Cattell (1978) calls a response-stance verb: it expresses a reaction to an assertion that was already brought forward in the discourse context and the claim reported by its CC is thus given, i.e. present in the common ground (see Krifka 2008: 262).¹² This is illustrated in (29): while a volunteered-stance verb like esan ‘say’ is felicitous reporting a new assertation, ukatu is only felicitous as a reaction to a preceding assertion as in (29a) and infelicitous if no corresponding assertion precedes, as in (29b), which would work only if the claim that the meetings are secret was brought up in some earlier part of the discourse.

Cross-linguistically, such CCs tend to show morphosyntactic similarities with nominals (Bogal-Albritten & Moulton 2018: 215) and it is thus not surprising that the use of a VN is more frequent in this context than with the other CEPs, the content of whose CCs is non-given. As for zin egin, on the other hand, the reasons for its deviating behaviour remain unclear.

Thus, while there is no strict mapping between form and meaning of a CC type, there are nevertheless very strong preferences that motivate classifying indicative intent clauses and VN belief clauses as atypical form-meaning pairs.

Another difference in preference for either a belief or an intent complement can be observed among the CEPs. As Table 4 shows, although all eight verbs can have CCs expressing either beliefs or intentions, the frequency of those meanings is rather different. Some CEPs seem to be more belief-prominent, taking belief CCs more often than intent CCs, whereas others are more intent-prominent.

Since the properties of belief VNs are of major interest in order to see in which regard they might differ from intent VNs, their number was increased by sampling also the remaining six VN belief clauses of konbentzitu (see Appendix A for details on the procedure).

In this extended sample, it is striking that all VN beliefs are either past tense like (30a) or stative like (30b) and (30c). This suggests that atypical belief clauses are found typically in contexts that are difficult or impossible to interpret as intentions, but it might also be an artefact of the annotation process since without much context, past and stative clauses were the ones most easily recognizable as beliefs. Syntactic considerations seem to play a role, too, since 16 out of 21 VN beliefs have a subject that is coreferential with the attitude holder in the matrix clause, as in (30a) and (30b), whereas, for a quick comparison, this is the case in only 8 of the first 21 indicative belief complements of ukatu. Overt non-coreferential subjects do, however, also occur, as (30c) shows, and are in principle grammatical.

Although these observations are preliminary because of missing context in the examples and possible annotation biases, the patterns observed in the corpus suggest the following two generalizations regarding the form-meaning mapping of the complements of belief-intent CEPs:

1. There is a clear link between form and meaning in that indicative clauses generally express beliefs and VNs generally express intentions.

2. Still, VNs may be used to express beliefs if they are stative or past tense.

In order to gather more data on these factors, a questionnaire was designed to explore which CC type can express one or the other meaning under different conditions.

4.1 Factors to be tested

The correlation between morphosyntax and semantics can be studied from two perspectives: with semantics as the independent variable, the question is which structures are appropriate to express a given meaning, and with morphosyntax as the independent variable, the question is which meanings a given structure can express. Here, the first perspective was chosen.

Following the two generalizations proposed at the end of Section ‎3.3, the experiment was designed to explore two hypotheses:

1. VNs are less acceptable in belief contexts and indicative CCs in intent contexts.

2. VNs in belief contexts are more acceptable if they are past tense or have a lower responsibility value, i.e. have properties that make them more difficult or impossible to construe as intentions.

Responsibility refers to properties of the complement on the scale in (31). This scale is based on the ease with which the attitude holder can cause the realization of the embedded state-of-affairs, inspired by Grano’s (2017) concept of a responsibility relation between the attitude holder and the embedded state-of-affairs.

The logic behind the assumption that it is easier for an agent to cause a state than to cause an action carried out by a non-coreferential agent is the fact that a prototypical agent is acting volitionally and can thus resist causation by simply not acting in the desired way. For example, it is easier to cause someone to be bored than to clean the kitchen because the latter requires a conscious decision to act on the part of the embedded subject. States with a coreferential subject, on the other hand, the fourth logical possibility, were excluded from the scale because they do not form a homogeneous group with respect to responsibility but differ a lot in the degree of control the attitude holder has over bringing them about. For example, being awake can be controlled to a much higher extent than being ill, and while the former is close to dynamic predicates with a coreferential agent, with the latter there is no big difference between its use with a coreferential and a non-coreferential subject.

The ten conditions included in the questionnaire are given in Table 5. For a discussion why not all 12 logically possible combinations were included, as well as a motivation of the choice of CEPs, see Appendix A.

4.2 Design of the questionnaire

For each of the combinations in Table 5, one context was created, as well as two additional intent contexts, one with pentsatu and one with bururatu, in order to yield more data on the acceptability of indicative clauses in intent contexts than only two contexts would provide. Although creating several contexts for each feature combination would have reduced the risk of idiosyncratic properties of the context or the stimuli influencing the results, including only a small number of items enabled a more qualitative, explorative approach, focusing as much as possible on the comments of the participants in order to get a better understanding of the relevance of the tested and possibly further factors. This seemed more appropriate for a first exploration of the patterns.

The contexts and the targeted meanings of the stimuli are given in Appendix A.

Since the constructions to be tested are the dispreferred ones, a production task would probably more often than not simply yield the preferred constructions. The factors were thus tested with a judgement task. Participants were presented a context description and as stimuli one sentence containing an indicative CC and the same sentence containing a VN CC, always listed with the dispreferred construction given first, i.e. VN–indicative in belief contexts and an indicative–VN in intent contexts, and they were asked which of the two options described the situation correctly. They could choose one, both or none of the options and had the opportunity to write a comment in a comment field. For the option(s) they did not choose, they were asked why they did not think it describes the situation correctly. The possible answers were “it sounds strange”, “it means something else” and again a comment field. The procedure is exemplified with one context in Appendix A.

Offering also the preferred construction as an option allowed to control for other factors that might lead to unacceptability: if participants do not accept the preferred construction either, this indicates that their rejection is due to some other factors that are unrelated to the CC type.

The experiment was conducted entirely in Basque. Giving the context description in the target language without using the target structure (i.e. the CEP to be tested) already in the description sometimes lead to a slight semantic discrepancy between the description and the stimuli,¹³ which was noted by some attentive participants. This problem could have been avoided by giving the context in Spanish and French, but on the other hand, this might have primed the structures most similar to the ones in the contact languages.

The twelve tasks were divided into two questionnaires, ensuring an equal distribution of each value of the tested parameters across the two questionnaires, and each participant was assigned one of them randomly. The reason for dividing the tasks into two questionnaires and thus not presenting all contexts to all participants was the hope that a shorter questionnaire would encourage the participants to take more time to explain their answers and possible further thoughts in the comments.

The questionnaire was created using SoSci Survey (Leiner 2024) and made available on www.soscisurvey.de. It was distributed predominantly via social media.

The questionnaire was filled in by 1584 participants from all seven provinces of the Basque Country. For details on their demographics, see Appendix B.

5.1 Acceptability of atypical CCs

Table 6 shows the percentages of participants that accepted both options, only indicative clauses, only VN clauses or neither in each of the 12 contexts. Table 7, on the other hand, shows the overall acceptance of indicative and VN clauses in each context, calculated on the basis of those cases where at least one of the two options was chosen as describing the context correctly. This is in order to control for factors other than the form of the CC that might have influenced the judgement. The following discussion of the data is based on these adjusted numbers.

Table 8 shows what reasons participants gave for rejecting the atypical construction, i.e. the VN in contexts a)-h) and the indicative clause in contexts i)-j). “Sounds strange” and “means something else” where offered as choices, “other” refers to cases where participants chose neither but gave their reasons in the comment field. In many cases this was simply “it’s wrong”, in others it was something along the lines of “it’s ok, but I prefer the other option” or “it’s understandable but I wouldn’t say it like that”. “Means something else” was chosen especially frequently in context b) and f), the two belief contexts that had the tense and responsibility values also found with prototypical intentions, namely future reference and a coreferential agent. Why context j) is the only intent context where the indicative clause is considered “strange” instead of “meaning something else” by more than half of the participants, on the other hand, did not become clear from the comments.

What is most conspicuous in Table 6 is that participants were in general more inclined to accept indicative clauses in intent contexts than VNs in belief contexts. While the former were, with one exception to be discussed below, accepted by slightly more than half of the participants, the latter generally had a very low score (with two exceptions). The reason for the high acceptance rate of indicative clauses in intent contexts is, however, probably primarily pragmatic, since intentions and beliefs about what one will do largely coincide in practice. In (32), for example, if Harkaitz and Ander plan to name the cat Pirritx, they probably also think that they will do so.

Since “an intention report comes along with the expectation that the attitude holder believes the named outcome to be within the realm of possibility” (Grano 2017: 590), not believing in the realization of one’s intentions is in many contexts pragmatically odd, but of course one can never be certain that nothing impedes the realization of one’s intentions. Participants who did not accept the indicative clause in intent contexts in fact often argued that it expresses a higher degree of certainty than the VN complement and that, for example, Harkaitz and Ander only have a plan but cannot know for sure that this will indeed happen. One participant commented on indicative clauses in intent contexts: “orokorrean, zehatzagoak dira es[a]ldi hauek. bai ala bai betetzekoak.” (“In general, these sentences are more precise, they are to happen inevitably.”)

For others, the control of the attitude holder over the action in the complement got lost in one way or the other when an indicative clause was used. Quite a few participants stated that it sounded like the CC had an agent different from the attitude holder: “Ez dirudi katuak Harkaitz edo Anderrengandik izena jasoko duenik, beste perstona batzuek izena jarriko diotela dirudi, eta Harkaitz eta Anderrek hori pentsatu dutela dirudi.” (“It does not sound as if the cat is going to receive its name from Harkaitz or Ander, it sounds as if others will name it and Harkaitz and Ander have thought that.”)

The same thought was also on the minds of the participants who commented on (33) that Ibai was lacking the ergative marker, interpreting the sentence as ‘It came to Joana’s mind that Ibai will invite her tomorrow for another coffee’: apparently, for them the construction made sense only if the attitude holder Joana was not the agent of the embedded clause.

Another possibility going in the same direction is construing the attitude holder as not having control over what will happen to them. One participant commented on (34): “Ematen du zoriaren esku utziko dutela geldituko diren edo ez.” (“It sounds like they will leave it to chance whether they will meet or not.”)

These intuitions show two things: first, for the speakers in question, indicative clauses can only express beliefs, and second, for many of them a belief reading works only if the attitude holder is not the agent of the embedded action, because otherwise they probably have difficulties to see why the complement would be construed as a belief and not as an intention.

Participants who did accept the indicative clauses, on the other hand, probably interpreted them as beliefs, too, but did not see a big difference between the belief and the intent reading. In order to tease them apart, one would have to come up with contexts where an intention is more clearly different from the corresponding belief.

One context where the difference is clearer is (35), where the coreference between the attitude holder and the embedded agent is only partial, so that the attitude holder can be less certain about the outcome than in cases where the intended action is carried out by themselves alone: intending for all to do something and believing that all will do it are clearly two different things.

(35) has by far the lowest acceptance rate of all indicative clauses in intent contexts, namely 14%.¹⁵ Like with the other indicative clauses in belief contexts, here, too, at least one participant got a reading with disjoint reference: “Bainatuko direnak beste batzuk dira, baina ez ideia hori izan duena” (“The ones who are going to swim are other people, not the one who had this idea.”) For some others it sounded like an action in the remote future: “Etorkizun urruna dirudi, momentuko erabakiaren ordez” (“It sounds like a distant future instead of a decision in that moment.”) Most participants, however, argued that if one person has the idea, this does not mean that all of them will in fact do it, thus again arguing with the higher certainty attributed to belief clauses: “Denak ba[i]natzeko gogoa izango dutela ziurtzat jotzen du” (“It presents it as certain that all of them will be up for swimming.”), “Proposamen bat egin du, ez predikzio edo agindu bat.” (“S/he has made a suggestion, not a prediction or a command.”)

Thus if the belief reading is more clearly different from the intent reading, this is immediately reflected in a much lower acceptance of indicative clauses. The fact that the attitude holder does not have full control over the fulfilment of the intention in context k) does, on the other hand, not seem to be a problem when a VN is used: this construction was accepted by 90,4% of all participants (see Table 6).

It seems thus that for most speakers indicative clauses invariably express beliefs not intentions and their acceptability (and usage) in contexts where an intention can be construed as well is due to the large overlap between these situations.

Such an overlap between a belief and an intent reading was less available in the belief contexts. There was, however, one case where apparently many participants slipped to the intent reading, namely (36), where the use of bururatu was probably a bit too forced.

Many participants remarked that they would not use the verb in this context and some said that the event of coming to Amaiur’s mind had already happened earlier, referring to the moment when Amaiur had the idea to travel to Egypt, which would be an intent reading. This is further supported by the fact that almost 75% of those who chose the VN did not choose the indicative clause. The very high acceptance rate of the construction with a VN, namely 48,1%, thus probably stems mostly from intent readings.

In the other belief contexts, however, participants often justified their rejection of the construction with a VN arguing that this would sound like an intention. For example, one participant described (37) as follows: “Inork ezer esan barik zerbait egitea erabaki duela esan nahi du.” (“It means that she has decided to do something without anyone saying anything.”)

The same is true to some extent even when the embedded agent or subject is not coreferential with the attitude holder (i.e. contexts c), d), g) and h)), although in these contexts the percentage of participants stating that the sentence “means something else” drops from 75,7–82,6% to 33,2–49,4% (see Table 8). On (38), participants commented that it sounds as if Alaitz wants to deliberately make the concert boring for the children. As one participant puts it: “Aditzera ematen du Alaitzek umeak asperrarazi nahi dituela gehiagotan ez etortzeko.” (“It sounds like Alaitz wants to bore the children so that they don’t come more often.”)

In a similar vein, on (39) a participant commented: “Txoriek zer jango duten ezin du Xubanek erabaki. Gainera, ez da hori berak nahi zuena.“ (“It’s not Xuban who can decide what the birds will eat. Besides, that’s not what he wanted.”)

Apparently, the use of the VN is for many speakers so strongly associated with the intent reading that this overrides the semantic restrictions on intentions. The association persists even if speakers cannot make any sense of this interpretation. On (40), which has also a non-coreferential subject, a participant commented: “Garazik ezin du anaiaren poztasuna kontrolatu, semantikoki ez du zentzurik.” (“Garazi can’t control her brother’s happiness, it doesn’t make any sense semantically.”) Another one could only imagine the construction expressing something else than the targeted belief reading even though they could not see what: “Arraroa da eta, aldi berean, ondo egotekotan (arrarotasun horretan “ondo”), beste zerbait esan nahiko luke” (“It is strange and at the same time, if it was correct (“correct” in its strangeness), it would mean something else”)

Thus, both the very low acceptance of VN clauses in belief contexts and the reasoning given by participants for their judgements show that VNs are, at least as complements of the two CEPs used in the questionnaire, so closely linked to intentions that speakers are reluctant to interpret them as beliefs even when an intent reading does not make any sense to them. Together with the low acceptance of indicative clauses in an intent context that differentiates more clearly between an intention and a corresponding belief, these results indicate a strong link between CC type and interpretation.

5.2 Properties influencing the acceptability of VN complements in belief contexts

Contrary to the second hypothesis formulated in Section ‎4.1, VNs in belief contexts yielded higher acceptance rates in contexts with a coreferential agent than with the other two responsibility values, as Table 9 shows, thus exactly in those cases where the CC could easily be interpreted as an intention as well. However, at least in context f) the reason for this acceptance rate seems to be precisely the fact that the clause was interpreted as an intention, as discussed above. It is quite possible that a large part of the acceptance rate of context b) also stems from participants who did not read the context carefully but thought that the sentence itself sounded natural. This thus needs further investigation. Stative CCs and CCs with a non-coreferential agent, on the other hand, seem to have an equally low acceptance rate.

The results are much clearer regarding tense: except for context f), where most participants choosing the VN probably did so with an intent reading in mind, the acceptance of VNs in belief contexts is higher in all past tense contexts than in all non-past contexts, as summarized in Table 10, although the gap between the past context with the lowest acceptance and the non-past context with the highest acceptance is probably negligible. The median acceptance of VNs in past tense contexts is six times as high as in non-past belief contexts.

Nevertheless, it is apparent from Table 10 that there are notable differences inside the classes of past and non-past contexts. These seem to be due not to a semantic but a syntactic factor: the expression of the embedded subject. As already observed in the corpus study, overt subjects seem to be clearly dispreferred with VNs, since almost all clauses with a (coreferential) null subject got a better score than almost all clauses with an overt non-coreferential subject. Controlling for the factor of tense, this is true of all clauses, as can be seen in Table 11.

The VN clause with the highest acceptance of all belief contexts, amounting to almost half of the participants, is thus the one in (41), which is both past tense and has a (partially) coreferential non-overt subject.

In the contexts with a non-coreferential embedded subject, many participants did in fact state that the VN would only work if the subject of both clauses was the same. For (39), a participant commented: “Gaizki dago edozein gauza adierazteko. Ergatiboa (txoriek) traba egiten du beste esanahia izateko.” (“It is wrong for expressing anything. The ergative (the birds) is an obstacle for it to have another meaning.”) With non-past VNs the reasons might be semantic as well as syntactic: for most speakers the VN can only express an intention, but it can be rather difficult to construe an intent reading when the embedded agent is different from the attitude holder. The fact that the same effect can also be observed with past tense VNs, which cannot express intentions anyway, suggests, however, that there is in fact a dispreference of overt subjects with VNs for syntactic reasons.

Contrary to this observation, however, most comments on (42) suggest that the problem is not so much the overt subject but rather the CEP.

(42) has the second highest acceptance rate of VN CCs in belief contexts, but nevertheless 87,2% of participants rejected it (not counting those that rejected both options). Some commented that the structure cannot be used when the embedded subject is different from the matrix subject, but more often participants saw the problem with the CEP: “Galdu izana aditzarekin[ ]ez dator bat” (“Galdu izana does not fit with the verb”), “Izana + pentsatu arraroa” (“Izana + pentsatu is strange”), and they gave examples of other matrix predicates with which the CC would work, some of them given in (43). Note that the CC occupies all kinds of different syntactic contexts in these examples and in (43c) it is a P argument, just as with pentsatu.

Semantically, however, the suggested contexts have in common, and in this differ from pentsatu, that their CC is not assertive, but desiderative in the ‘hope’ reading of (43c) and factive in the other three cases. The latter fits well with the observation that ukatu ‘deny’, as discussed briefly in Section ‎3.3, occurs with VN beliefs much more frequently than the other CEPs in the sample: although ukatu is not factive, it has in common with factive verbs that the content of the CC is discourse-given.

The only alternative matrix clause offered for context e) (the past tense context of bururatu), given in (44), is factive as well:

A difference between belief-prominent pentsatu and intent-prominent bururatu with respect to the availability of atypical CCs, on the other hand, could not be observed. Table 12 compares the acceptance rates of atypical CCs with the two CEPs. There are notable differences in some lines, but they can generally be explained by other factors: in the past tense, the difference is due to the presence or absence of an overt subject, in the contexts with a coreferential agent it is due to the more widespread construal of the CC as an intention in context f) and in the intent contexts it is due to the clearer difference between the belief and the intent reading in context k). Of course, since the study included only one context for most of the combinations of factors, these results are rather preliminary and need to be confirmed with more data.

5.3 Further discussion

The results show that the association between beliefs and indicative clauses on the one hand and intentions and VNs on the other hand is quite strong. What looks like indicative intent clauses at first glance, at closer examination turns out to rather express beliefs that overlap pragmatically with intentions, although a more detailed examination would be necessary in order to see whether this is true in all cases. VNs, on the other hand, are only very marginally acceptable as belief complements of belief/intent CEPs, at least if the content of the CC is not discourse-given. Thus, an analysis of the alternation based on a difference in CC type can capture the largest part of the pattern in Basque, too. What it does not capture, however, are interactions between CC type and specific properties of the given CC, namely the fact that VNs in belief contexts improve notably when the CC is past tense and when they appear without an overt subject.

The hypothesis that this effect of specific properties of the CC on the acceptability of VN beliefs is motivated by a need for sufficient disambiguation between the two readings is, however, confirmed partly at most. Although past tense VN beliefs being more acceptable would be in line with the fact that an intent reading is ruled out in these cases, mere semantic implausibility of the intent reading turned out not to be sufficient to yield a belief reading. In addition, belief VNs are in fact not only avoided in contexts where they could be mistaken for intents: other volunteered-stance (Cattell 1978) predicates like esan ‘say’ or uste izan ‘think, be of the opinion’, which only have one reading, do not combine with VN CCs either. Thus, a mere need for disambiguation does not seem to be the primary force behind this pattern. Instead, absolutive VNs are apparently strongly dispreferred for expressing discourse-new assertions. They are in principle structurally adequate to express Propositions, as evidenced by their use with discourse-given complements of verbs like ukatu ‘deny’ or in the contexts listed in (43a, b, d), but what they lack is the structural similarity to declarative main clauses, which seems to make them less suitable to express representational attitudes. This observation supports not only Hacquard & Lidz’s (2019) claim that representational attitudes are cross-linguistically linked to the properties of declarative main clauses, but is also evidence that this link cannot simply be reduced to a difference in structural size between Propositions and Situations. It strengthens thus Hacquard & Lidz’s (2019: 77) explanation that the association between declarative main clause syntax and assertive complements is due to the fact that declarative main clauses are in fact the syntactic structure used for non-embedded assertions.

The use of VNs for discourse-given complements, on the other hand, is in line with the cross-linguistic tendencies of factive or given complements to be more similar to nominals morphosyntactically (see Bogal-Albritten & Moulton 2018: 215–216 for an overview). While factive CEPs do not express a judgement of truth but presuppose the truth of their complement and thus do not belong to the class of representational attitude verbs (Kiparsky & Kiparsky 1970: 147; Bolinger 1968: 17–18), response-stance verbs do in fact express representational attitudes while at the same time having a discourse-given complement. Interestingly, this is reflected in the complementation patterns of ukatu, which allows both indicative CCs, in line with expressing a representational attitude, and absolutive VNs, in line with having a discourse-given complement.

The fact that past tense VNs are nevertheless accepted in non-given belief contexts by up to half of the speakers consulted indicates that they should not simply be analysed as the past counterpart of VNs but rather as a CC type of their own – an analysis that is justified morphologically, since unlike non-past VNs, they do not involve the nominalizing suffix -t(z)e but are based on the perfective participle (see Section ‎3.1).¹⁶ The two nominalizations are commonly assumed to belong to the same category because of their shared morphosyntactic properties and distribution (Zabala & Odriozola 1996: 240), but their different behaviour in belief contexts indicates that nominalized perfective participles differ, at least for some speakers, from VNs in their semantic properties. If they are indeed a different CC type, their use in belief contexts is unproblematic for analyses locating the semantic difference in the CC, since unlike VNs, nominalized participles only express beliefs and never intentions.

The effect of givenness and subject coreference on the acceptability of both types of nominalizations, on the other hand, cannot be explained by postulating different CC types. An adequate analysis of these patterns must be able to take into account the interaction of different factors that seem to be able to override each other. This concerns not only the effect of both factors on the acceptability of VNs in belief contexts but also the interaction between the two:

the comments on context a), mentioned in Section ‎5.2, suggest that once the preference for VN Propositions to be discourse-given is satisfied, the preference for avoiding an overt subject becomes irrelevant. These patterns suggest a hierarchical ranking of constraints, as in the Optimality Theoretic analysis of the Mandarin verb xiang, which alternates between the readings ‘think’ and ‘want’ depending on several different properties of the embedded and the matrix clause, in Yan (2025b). As for the association of a certain type of attitude with a certain type of CC, this means that although there are canonical form-meaning mappings, they may be neutralized in certain contexts – comparable to Jackendoffs (1985: 447) observation that infinitival CCs, although being canonically linked to intentions, can express beliefs with certain verbs, like claim.

This study has, of course, several limitations: first of all, as the corpus study showed, different CEPs might differ significantly in their complementation patterns and the conclusions drawn here thus apply primarily to the two CEPs studied. Furthermore, since it included only one item for each combination of factors (except for the intent contexts i) and j)), more data is needed in order to fully tease apart the contributions of the studied as well as possible other factors. Since the contexts did not explicitly target intentions without beliefs, the finding that indicative clauses in intent contexts actually express the corresponding beliefs instead of intentions remains somewhat speculative. Another open question is to what extent the acceptability of the atypical construction was affected negatively by the fact that participants were presented with both the typical and the atypical construction on the same page.

Corpus study

For each of the studied verbs, the corpus was queried for the respective lemma and the order of the occurrences was randomized, then the first 100 occurrences with a CC¹⁷ were annotated for form and meaning of the CC. The semantic annotation of CCs that have the necessary semantic properties to be interpreted as intentions (as discussed in Section ‎3.2) has to be taken with a grain of salt, however. The semantic context was often not clear enough to completely rule out one or the other reading, so that the decision for one or the other reading might in some cases reflect interpretation during annotation rather than actual usage.

In order to obtain a slightly larger sample of belief VNs, all remaining 27 occurrences of instrumental VN + konbentzitu in the corpus were considered as well, which yielded 6 additional VN belief clauses. This verb was chosen for the following reason: Since past and non-past VNs cannot be queried by means of the same regular expression because they are so different in form (see Section ‎3.1), a sample that is small enough to allow to go through all occurrences of both past and non-past VNs is necessary in order to be able to compare their proportions in the set of belief VNs. Konbentzitu occurred with a preceding instrumental VN only 29 times in the whole corpus and constituted thus such a small sample.

Questionnaire

In principle, combining the three factors context, tense and responsibility yields the 12 logically possible combinations shown in Table 13. However, not all of the values are equally compatible: as discussed in Section ‎3.2, intentions are always in the future relative to the matrix state-of-affairs and are more difficult to construe the less obvious the responsibility relation is.

Furthermore, not all factors are logically independent regarding the question whether or not the embedded state-of-affairs has properties compatible with an intent reading: since non-future CCs cannot express intentions anyway, responsibility can only influence the ease with which a complement is construed as an intention if the CC has or is compatible with future reference. For the purpose of this study, responsibility was thus disregarded in non-future contexts. Intent contexts, on the other hand, were limited to actions to be carried out by the attitude holder in order to not introduce further confounds.

Belief/intent prominence of the CEP did not show any effect in the corpus data, but in order to control for this factor, each of the combinations of tense and responsibility was paired once with a more belief-prominent and once with a more intent-prominent CEP. The aim was to choose CEPs that occur with both complementation patterns relatively frequently in order to ensure that a construction does not sound odd to speakers simply because the CC type is not commonly used with the CEP in question. This might be the case, for example, with instrumental VNs of konbentzitu or indicative CCs of proposatu, which are rather infrequent as shown in Table 1 and Table 2. Among the more intent-prominent CEPs, the one with the most equal distribution of indicative clauses and VNs is bururatu ‘occur, come to mind’. Among the more belief-prominent CEPs, it would be ukatu ‘deny’, which, however, as mentioned in Section ‎3.3, deviates from other CEPs in this class in argument structure and givenness of the content of its CC. So instead the CEP with the second most VN occurrences among the belief-prominent CEPs was chosen, namely pentsatu ‘think, plan’.

The contexts and the targeted meanings of the stimuli presented to the participants are given in (45):

All CEPs in the stimuli were in the perfective present, which expresses perfect, narrative present and hodiernal past and is equally compatible with belief and intent readings, unlike the imperfective that might favour a stative belief reading with pentsatu. The attitude holder was always third person. Word order was always CC-CEP.

The procedure in the questionnaire is exemplified in (46) and (47) for context a). The participants were given a context description as in (46), followed by the question ‘Which sentence describes the situation correctly? (If you think both are correct, choose both.)’ and the two stimuli, as in (47), as well as a comment field.

When they clicked “forward” and had not chosen both of the sentences, the (first) sentence they had not chosen appeared on the next page together with the question “Why does this sentence not describe the situation correctly?”. If they had also not chosen the second option, the same page appeared afterwards for the second option.