1. Setting the scene

Recursion in language is considered a procedure that can be repeated ad infinitum, creating structures of unbounded extent and complexity (Hauser, Chomsky & Fitch Reference Hauser, Chomsky and Tecumseh Fitch2002, Pinker & Jackendoff Reference Pinker and Jackendoff2005, van der Hulst Reference van der Hulst and van der Hulst2010, Karlsson Reference Karlsson and van der Hulst2010b, Hollebrandse Reference Hollebrandse, Amaral, Maia, Nevins and Roeper2020). A quintessential example of recursion are sentences with multiple embedded clauses, such as John likes Mary, who likes Tom, who likes Michael, who likes Sarah, and so on. While multiple embeddings of the same type (like the relative clauses in this example) are rare in actual usage, written language often features sentences with successively embedded clauses of different types; see (1) with a complement clause, an adverbial clause of time, and a relative clause.

In this study, we investigate long-distance temporal reference, a phenomenon that occurs exclusively in complex sentences containing deeply embedded clauses (DECs).Footnote ¹ We present the phenomenon in Section 2 and then try to identify the grammatical contexts in which it arises.

DEC is a clause which is embedded in a clause that itself is embedded. The clause in the innermost brackets of (1) occurs at a depth of three, i.e. it is a third-order embedded clause. Sentences containing such clauses can be formalized as [C [C1 [C2 [C3 […]]]]], where C = main clause, C1 = first-order embedded clause, C2 = second-order embedded clause, C3 = third-order embedded clause, etc. A DEC is a clause below C1, i.e. a clause occurring at an embedding depth of two or more.

An attentive reader may notice that we equate ‘embedded’ and ‘subordinate’ clauses. Matthiessen and Thompson (Reference Matthiessen, Thompson, Haiman and Thompson1988) argue that adverbial clauses, such as in I am not going out [because it’s raining], are not embedded within the main clause, although they are dependent on it. They reserve the term ‘subordinate’ for such clauses and distinguish them from both embedded clauses (e.g. complement or relative), and coordinate clauses, which involve neither embedding, nor dependency (see also Givón Reference Givón1990:Chapter 19).Footnote ² Here, we adopt a more traditional view of embedding (reflected in Haspelmath Reference Haspelmath, Haspelmath and König1995), which regards adverbial clauses as embedded too. We acknowledge that the issue boils down to the degree of structural interlacing between the subordinate and the main clause. Given that the distinction between ‘subordinate’ and ‘embedded’ is mainly useful for analyzing different types of adverbial clauses, which are not our focus here, we refrain from drawing a boundary and use these terms co-extensively.

We focus on the temporal reference of clausal contents in sentences with DECs. Temporal reference is traditionally categorized as either absolute (or ‘deictic’) or relative (or ‘anaphoric’) (see e.g. Demirdache & Uribe-Etxebarria Reference Demirdache and Uribe-Etxebarria2007:20). These notions have primarily been used to account for the temporality of the main clause and the C1 clause (e.g. Xrakovskij Reference Xrakovskij2016), but they can also be applied to describe the temporal relations of DECs.

Absolute temporal reference places the state of affairs (SoAs) described in the clause in the past, present, or future relative to the speech time, i.e. to the moment in which the speaker produces her utterance. In (2), the temporal reference of C1 is absolute; the most natural interpretation of its contents is that the SoAs find the friend is located in the future from the speech situation, not from the moment when the SoAs not believe occurs. The temporal reference of C2 is also absolute: whom we saw last time in Tübingen is estimated as past from the moment of speech, not from the future moment indicated in C1. Otherwise, we might think that the speaker has not yet seen their friend in Tübingen but will before Ulli meets him. This is not a possible interpretation of the sentence in (2).

Relative temporal reference locates the SoAs described in the clause in the past, present, or future relative to some other point in time, i.e. to a moment which does not coincide with the moment of speech. This other point in time is specified in a superordinate clause. In (3), the SoAs expressed by C1 is located in the past from the moment of speaking, that is, Shakespeare believed is past from the speaker’s ‘now’, and thus, this clause has absolute temporal reference. C2, with its pluperfect form, on the other hand, expresses past from the past moment specified in C1. Therefore, C2 has relative temporal reference, specifically, relative past.

Recapitulating: absolute temporal reference locates a SoAs in the past, present, or future in relation to the time of the utterance; relative temporal reference locates a SoAs in the past, present, or future in relation to another other point in time (i.e. the SoAs is anchored to a different temporal reference point, not the moment of speech). Both types of temporal reference influence the choice of grammatical tenses in the language.

2. Object of study

We focus on a third type of temporal reference, which can only be examined in conjunction with the main types mentioned above. This type occurs only in DECs, i.e. in second-order or deeper embeddings, and is rarely discussed in the literature. It is instantiated when the temporal reference of a clause is estimated relative to a SoAs in another clause that is not its immediate superordinate clause but is higher in the embedding hierarchy. For instance, if the target clause is C2, its temporal reference would be past, present, or future relative to the time of the SoAs expressed in the ultimate main clause C. Our study is inspired by the work of Alexander Letuchiy, who briefly discussed constraints on tense marking and temporal reference in tri-clausal constructions in Russian and shared some ideas about the contexts favoring this type of reference (Letuchiy Reference Letuchiy2018, Reference Letuchiy2021:543–546). Admittedly, Letuchiy was not the first to notice the phenomenon; it has been mentioned by Sgall (Reference Sgall, Wagner and Widgen1990), who presented the following example.

In this sentence, C1 has relative temporal reference. This clause expresses future relative to a moment in the past specified in C. C2 also has relative temporal reference: the expressions of future is hearing and would hear, both possible in this clause, do not convey future relative to the moment of speech, that is, we do not have absolute future here. But the future tense in C2 does not convey posteriority in relation to the SoAs of the immediately higher clause C1 either; in fact, the SoAs of C2 precedes the SoAs of C1. The only adequate reading of the temporal relations in this sentence is that C2 expresses future relative to the moment of time specified in C, the ultimate main clause. Letuchiy (Reference Letuchiy2021:543) calls this type of temporal reference ‘absolute-relative’. We regard it as relative, as it anchors the clausal contents to a moment in time that does not coincide with the speech act moment and is specified in another clause. Therefore, to avoid confusion with Comrie’s (Reference Comrie1985:64) notion of ‘absolute-relative tense’ we use the term long-distance (relative) temporal reference (LDTR). Since LDTR is always relative, we place ‘relative’ in parentheses. Absolute temporal reference, on the other hand, is always ‘short-distance’ as the temporal characteristics of the clause (regardless of how deeply embedded) are directly accessed from the speech moment.

Letuchiy studied LDTR using mainly constructed examples from Russian. We will investigate it in sentences from three Finno-Ugric languages, collected from corpora and organized into samples of sentences with DECs. Although there are few occurrences of the phenomenon in the data, these allow us to make some observations about its distribution, for example, across types of clauses and temporal orientations (to the past, present, or future). This is an exploratory observational study on a rare yet distinctive phenomenon, aiming to elucidate its characteristics within a precisely defined population (described in Section 3). We leverage all available insights (irrespective of theoretical commitment) to enhance our understanding of LDTR. Research on the nuances of temporal reference in complex sentences, conducted within the frameworks of formal semantics and generative linguistics, proves particularly valuable in this context.

It is easy to study the phenomenon in sentences like (4), where all clauses contain indexical expressions (a month ago, in two weeks, in ten days) that help to disambiguate their temporal reference. We often lack such clues, and the grammatical structure of the sentence alone does not suffice to determine the temporal reference of a clause. A major problem is that ‘temporal reference’ is not related to ‘tense’ in a trivial manner; the type of reference cannot always be singled out based on grammatical tense. In Section 3, we present some ambiguous contexts that were excluded from the study because they would have skewed the analysis. Suffice it to say that assigning temporal reference to a clause involves choosing between ‘good’ (realistic) and ‘bad’ (counter-intuitive) matching of the order of events described in the sentence with the combinations of forms expressing tense and temporality.

Another caveat is that we apply the notions of ‘absolute’ and ‘relative’, typically used to account for functional differences between grammatical tenses, to the description of temporal relations between clauses (see Section 4.1 for details). Therefore we decided to treat the temporal reference of the ultimate main clause C as absolute, even though the tense in this clause can be relative or absolute-relative (in terms of Comrie Reference Comrie1985:64; see also Metslang Reference Metslang1991 on Estonian). For example, when studying LDTR in C2s, we are primarily concerned with C as an orientation point for the temporal reference of C2, rather than with the internal temporal structure of C.

The aim of this study is to make observations about LDTR by monitoring the semantics of the sentences in which it occurs. We use attested examples, and even though our sample is small, the patterns we found would not have been noticed through introspection alone. The main research question is:

The answer to this question is the first step towards answering the more difficult question: What properties of the clause (the DEC) and the higher clauses in the sentence evoke LDTR?

3. Data

In this section, we introduce the languages, the samples we extracted from the corpora, the variables for which we annotated the sentences, and the criteria used in selecting examples of LDTR.

We study long-distance temporal reference in Estonian, Moksha Mordvin (henceforth Moksha), and Komi Zyrian (henceforth Komi). These languages belong to the Finnic, Mordvinic, and Permic branches of Uralic, respectively, and are genealogically quite distant from each other. All three are considered SVO languages, but while in Estonian SVO is dominant, Moksha and Komi have more flexible word order, with SVO only slightly prevailing over the older Uralic SOV order (see Hamari & Ajanki Reference Hamari and Ajanki2022, Kuznetsov Reference Kuznetsov2022, Metslang Reference Metslang2022, Vilkuna Reference Vilkuna2022). Moksha has the least differentiated tense system, with two (synthetic) past tenses, in addition to the unmarked non-past, whereas Estonian has three (one synthetic, two analytic), and Komi up to six past tenses (two synthetic, four analytic). Komi is the only language with a future tense, which, however, is overtly marked only in the third person. Estonian most frequently uses European-type bifinite clausal subordination, whereas Moksha most frequently uses non-finite subordinate clauses (i.e. monofinite subordination constructions); Komi falls in between.Footnote ³ Estonian has been in contact mainly with German, Russian, and Latvian; Moksha with Russian, Tatar, and Chuvash; and Komi mainly with Russian.

As previously noted, LDTR occurs only in sentences with deeply embedded clauses. Such sentences are characteristic for written, rather than oral, varieties, and therefore a few words are in place here about the literary varieties of Estonian, Moksha, and Komi. Estonian represents a typical East-Central European national language, with a vibrant literary tradition and texts in all genres. In contrast, Moksha and Komi literary languages never became widely used registers of communication among their speakers. Their language planning began in the early Soviet period, and the body of printed texts in these languages remains very small compared to that of Estonian.

Some of these circumstances can be assumed to impact the temporal reference of clauses. In some Uralic languages spoken in Russia, complex sentences with finite subordinate clauses have been claimed to be a product of standardization and language planning (see Buzakov Reference Buzakov1973:8–9 for Mordvin languages and Edygarova Reference Edygarova2022:122–124 for Udmurt). Uralic languages exhibit a vast diversity in their reliance on finite versus non-finite subordination, and while finite clauses can have either absolute or relative temporal reference, non-finite clauses tend to have relative temporal reference cross-linguistically (Klein Reference Klein1994:131).Footnote ⁴ Similarly, the temporal reference of a clause is more easily disambiguated in a more differentiated tense system (as in Komi). Conversely, a less differentiated tense system (as in Moksha) leaves more ambiguity in interpreting temporal reference. Languages with a long literary history and presence in all genres are likely to feature complex sentences with greater average embedding depth than languages primarily used for oral communication.

In this study, we are interested in observations on long-distance temporal reference with general significance. Given that LDTR is scarcely studied, we appreciate the opportunity to present evidence from outside the Indo-European family. The differences outlined above make Estonian, Moksha, and Komi a concise variety sample for such a study. At the same time, these languages are related (albeit remotely), and share structural features potentially relevant to the phenomenon studied, such as right-branching embedded clauses and being non-SOT languages.Footnote ⁵

The data comes from samples assembled for a broader investigation into the properties of DECs. Sentences with DECs in Estonian, Moksha, and Komi were collected from printed text corpora: ENC 2019 for Estonian, CCLM for Moksha, and CZC and Korp for Komi, all from fiction and press subcorpora. The Estonian sample includes 337 sentences with 444 DECs, the Moksha sample 210 sentences with 241 DECs, and the Komi sample 219 sentences with 273 DECs (the samples are available in the data source Kehayov & Todesk Reference Kehayov and Todesk2025). Thus, we have 958 DECs, occurring at an embedding depth between two and five, in which we could study the phenomenon of LDTR.

Each sentence with DECs was entered on a separate line in an Excel table, and its embedded clauses were coded for the following variables (see Kehayov & Todesk Reference Kehayov and Todesk2025):

1. (a) embedding depth (C1, C2, C3, etc.),

2. (b) position of the clause relative to the superordinate clause (right-embedded, left-embedded, center-embedded),

3. (c) occurrence of a subordinating conjunction in the clause,

4. (d) morphological form of the verb,

5. (e) type of the clause (complement, relative, adverbial),

6. (f) temporal reference (absolute, relative, [potentially] LDTR).

For this study, we reviewed all candidates for LDTR and retained only unequivocal examples of the phenomenon. We excluded the instances where LDTR was suspected to occur in C2, but the ultimate main clause C was in the present tense, marking a moment or time interval overlapping with the moment of speech. This was necessary because, in C2s, LDTR can only be determined in contexts where the SoAs of C takes place at a time different from speaker’s ‘now’. When the tense in C is present, temporal reference relative to the time of C cannot be distinguished from absolute temporal reference (Demirdache & Uribe-Etxebarria Reference Demirdache and Uribe-Etxebarria2007:24). However, we included sentences ending at an embedding depth of two, where C is in the historical present; in such cases, the present tense form refers to a moment (or time interval) in the past, specified by a time adverbial in the sentence (e.g. at that moment).Footnote ⁶

In this way, we pinpointed 27 examples of LDTR in the material: six from Estonian, four from Moksha, and seventeen from Komi. The scarcity of data precludes a quantitative study of correlations between LDTR and the other variables. Consequently, the observed associations between the phenomenon and other features of the sentence presented below are only suggestive. In other words, while the ideal purpose would be to determine what is possible, we can, at best, say what seems typical for LDTR.

However, the number of occurrences of the phenomenon within a representative sample, which was assembled according to clear definitional criteria, provides insight into the frequency of the phenomenon in written registers. The number of occurrences of LDTR in each language within the respective populations of DECs (Estonian n = 444, Moksha n = 241, Komi n = 273) suggests that the phenomenon is very rare.

4. Occurrence of long-distance temporal reference in the data

Among the variables listed in Section 3, LDTR appears to correlate only with the temporal reference of the immediately higher clause (f), and the type of clause (e). Therefore, we start with some general remarks about the relationship between embedding depth, temporal reference, and type of clause.

Jurkiewicz-Rohrbacher & Kehayov (Reference Jurkiewicz-Rohrbacher and Kehayov2024) found a correlation between the embedding depth of the clause and its type and temporal reference in a different, more restricted population than the one used in this study. A subordinate clause, whether finite or non-finite, had to contain a subordinator to be considered a clause, i.e. a separate level of embedding. In this population, they observed that DECs are more likely to have relative temporal reference than C1s, which more often have absolute temporal reference than DECs. They also observed that the proportion of relative clauses increases with the increase in embedding depth: DECs are more likely to be relative clauses than C1s. Conversely, the proportion of complement clauses decreases: DECs are less likely to be complement clauses than C1s.

In (post-)generative research, there has been much discussion about the limitations on the availability of absolute and relative temporal reference in different subordinate clauses. For example, Demirdache & Uribe-Etxebarria (Reference Demirdache and Uribe-Etxebarria2007) argue that while complement clauses can have either absolute (‘independent temporal construal’) or relative (‘dependent temporal construal’) reference, relative clauses usually have absolute temporal reference. Among relative clauses, non-restrictive ones are often associated with absolute temporal reference, while restrictive ones are linked to relative temporal reference (see Hatav Reference Hatav and Binnick2012, Rathert Reference Rathert and Binnick2012). Regarding adverbial clauses, Demirdache & Uribe-Etxebarria (Reference Demirdache and Uribe-Etxebarria2007) claim that when-clauses always have absolute temporal reference.Footnote ⁷

Less attention has been given to embedding depth as a factor influencing temporal reference. However, Grønn (Reference Grønn2020) observed that Russian before-clauses have absolute temporal reference when they are C1s (and their C has absolute reference), but when they are C2s and subordinate to an infinitival C1, they tend to have relative temporal reference. Letuchiy (Reference Letuchiy2021:545) makes a similar observation for while-clauses.

Regarding contexts that attract long-distance temporal reference, Letuchiy (Reference Letuchiy2021:548) notes that a relative temporal interpretation is not possible in a when-clause occurring as C2 if its SoAs precedes the SoAs of C1. In such cases, LDTR occurs. See (5) from Russian in which the future umrët conveys posteriority with respect to the SoAs in C. The form umer, which would express anteriority in relation to C1, is not possible in this context.

Letuchiy (Reference Letuchiy2021:548) also notes that LDTR is preferred in C2 over relative temporal reference if C1 is a relative clause. Additionally, he observes that LDTR is more acceptable in non-referential adverbial and relative C2s than in referential C2s (Letuchiy Reference Letuchiy2021:555).

These observations are based on sentences constructed for the researcher’s purposes, which makes us somewhat skeptical about their validity, given the number of relevant features (and unknowns) in all clauses. Aware of the limitations of this introspective approach and to avoid tunnel vision, we approached the topic from a different angle. We searched for LDTR in a corpus of sentences containing DECs, the only environment in which LDTR occurs. Although the number of examples we found is small, they are natural and likely represent ‘real-life variety’. There are 26 sentences in the data, containing a total of 27 clauses with LDTR; in one sentence (from Komi), there are two clauses with LDTR. These sentences form the population in which we study the distribution of LDTR.

4.1 Embedding depth and temporality

The observed combinations of embedding depths, temporal reference (absolute, relative, LDTR), and temporal orientation (to the past, present, or future) are shown in Table 1.Footnote ⁸ Each type of combination is assigned a roman numeral.Footnote ⁹

Table 1.

Embedding depth and temporal reference in sentences with LDTR

We will exemplify some of the types in the table, but first, a few words about the language we use when discussing temporal relations across clauses. Since we are interested in temporal relations between clauses rather than within clauses, we treat the temporality of a clause as a plain, elementary parameter. We recognize that the temporality of clausal contents cannot be accounted for without resorting to additional parameters, such as Reichenbach’s (Reference Reichenbach1947) notions of E (event time), R (reference point), and S (point of speech), or Klein’s (Reference Klein1994:7) distinction between ‘time of situation’, ‘utterance time’, and ‘topic time’, to mention a few influential frameworks for analyzing tense and temporality. But for the sake of simplicity, we will disregard the fact that the ‘tense’ of a clause can be broken down into (an interplay of) further parameters. Accordingly, the types of temporal relations in sentences with LDTR will be discussed as if clausal contents did not have complex internal temporal structure but are elementary units in the analysis of complex sentences. This approach can be anticipated in the notions in Table 1, where, for example, the temporal reference of C1 is past or future relative to C.

The types in Table 1 that will be exemplified will also be supplied with schematic representations of the temporal relations between their clauses. A key to Figures 2–9 below is provided in Figure 1.

Figure 1.

Key to the schematic representations of temporal relations between clauses in sentences with LDTR. It introduces all the concepts necessary for understanding Figures 2–9.

Figure 2.

Schematic representation of Type I sentences with LDTR.

Figure 3.

Schematic representation of Type III sentences with LDTR.

Figure 4.

Schematic representation of Type IX sentences with LDTR.

Figure 5.

Schematic representation of Type VI sentences with LDTR.

Figure 6.

Schematic representation of Type VIII sentences with LDTR.

Figure 7.

Schematic representation of Type X sentences with LDTR.

Figure 8.

Schematic representation of Type XIV sentences with LDTR.

Figure 9.

Schematic representation of Type XII sentences with LDTR.

In Figures 2–9, { C1 }, for example, represents the point or interval of time during which the SoAs of C1 occurs; { C1 } has a temporal reference and can serve as an anchor for the temporal reference of another clause. As shown in Figure 1, the temporal references of different clauses, arranged on the horizontal line depicting time axis, will be indicated by different types of curved lines. The roman numerals in Figures 2–9 below correspond to the types in Table 1.

The most common combination in our data (Type I in Table 1) is where C expresses absolute past, C1 relative future, and C2 long-distance relative past; see example (6) from Komi, whose temporal structure is shown in Figure 2. In this sentence, C expresses that someone hoped for something before the moment of utterance. Hopes are oriented towards the future, so the complement C1 is located in a time span after the time of hoping. C2, containing the adverb already, places the SoAs described in it prior to the time of hoping. It is clear that the temporal orientation of C2 is past relative to C, not past relative to C1, and that we have LDTR here. A reading of this sentence where the brothers and sisters already have families before Boris can meet them, but not before the third-person protagonist expresses her hope, is not possible in this context.

Table 1 above shows that despite the small sample, most possible combinations of embedding depths and temporal values are attested (at least for sentences ending with a C2). LDTR can be past, present, or future relative to a clause, whose content is situated in the past, present, or future. The clause immediately above the clause with LDTR can also be in the past, present, or future.

Observe (7) from Moksha, where C refers to the past from the moment of narrating, C1 is situated in the past relative to C, and the center-embedded C2 is in the present relative to C; this is Type III in Table 1, illustrated in Figure 3. The main clue for interpreting the present tense of C2 is provided by the time of C; the mother works at the kiosk at the time the man is talking about the event of her finding a cash shortage. She also works there when she finds the shortage, but ‘finding’, with its punctual meaning, is not well-suited to providing a reference point for estimating the time of the long-term activity of being employed somewhere. C, with its durative ‘explain; talk’, is a better anchor for the temporal reference of C2. Furthermore, the predicate of C1 follows C2 in the flow of the utterance. Thus, at the moment the interlocutor hears C2, the only available cue for interpreting the time reference provided by the present tense in this clause is the time of C. This example also demonstrates that LDTR occurs in nested (center-embedded) clauses.

Similarly, in (8) from Komi, illustrating Type IX (shown in Figure 4), C is situated in the future from the moment of speaking, C1 is in the future from C, and C2 in long-distance past relative to C. The SoAs of C is the best reference point for the temporal interpretation of C2. Stating that the SoAs of C2 is in the past from C1 is not an exhaustive clue for interpreting the temporal structure of the sentence, because this SoAs cannot be prior to C1 but subsequent to C.

In (9), from Estonian, C is in the past from the speech situation, C1 in the future relative to C, and the interval covered by C2 overlaps with the SoAs in C. Thus, C2 has long-distance present temporal reference. This is Type VI in Table 1; it is represented in Figure 5.

Example (10) from Komi includes several clauses with the same tense form. C is set in the past from narrator’s present, C1 is in the past from C, and C2 is also in the past from C, but not from C1. C2 is a when-clause, expressing simultaneity with the SoAs of C1. If the reference point for the interpretation of C2 was the time of C1, C2 would not be in the past tense. Therefore, the past tense in C1 must be motivated by its location in time relative to C. This is Type VIII in Table 1, illustrated in Figure 6.

In Table 1, we have six sentences where LDTR occurs at an embedding depth of three, and one where it appears at an embedding depth of five. Consider (11) from Estonian, where C is in the past from narrator’s now, C1 is present relative to C, C2 is future relative to C1, and C3 is in the long-distance relative past with respect to C1. This is Type X, illustrated in Figure 7. It is clear that were able to pay school fees in C3 expresses an event that happened before the SoAs in C1. C2 does not provide sufficient clues for the temporal interpretation of C3. The SoAs ‘being able to pay school fees’ is prior to the decision that hits middle-income parents, but it must also precede the statement of the Parents’ Association.

Finally, note that a clause can be temporally oriented toward a clause which is not one but two clauses away; see Type XIV in Table 1, illustrated in Figure 8, where the reference point for the temporal interpretation of C3 is the SoAs of C. Example (12) comes from Komi. Here C is in the past from the speech time, C1 is in the future relative to C, C2 is in the future relative to C1, and C3 in the present relative to C (not C1!). The temporal indexical adverbial talun ‘today, now’ in C3 conveys simultaneity with the urge expressed by Viktor Basagin in C.

The variety of possibilities in Table 1 leaves the impression that embedding depth (C, C1, C2, C3, etc.) and temporal orientation (to the past, present, or future) are independent variables. This does not seem to apply to temporal reference. In all our examples, the temporal reference of the immediate superordinate clause of the clause with LDTR is relative (see Table 1). This is noteworthy given that, in our samples of sentences with DECs, the proportion of C1s exhibiting relative temporal reference is 43% in Estonian, 63% in Moksha, and 66% in Komi.Footnote ¹⁰ At the same time, Table 1 lists 20 sentences in which LDTR occurs in C2, and in all of them, the corresponding C1 displays relative temporal reference. This suggests that the observed association between LDTR and relative reference in the superordinate clause is unlikely to be a mere reflection of the overall frequency of relative temporal reference in C1s. Another observation from Table 1 is that in all seven instances where LDTR occurs at an embedding depth of three or more, the immediately higher clause expresses relative future.

4.2 Type of clause

In this section, we monitor the relationship between clause type and embedding depth. Table 2 shows the type of the clause featuring LDTR and its matrix clause (i.e. the immediately higher clause in the hierarchy). We have three clause types (complement, relative, adverbial) and two levels of embedding depth (‘clause with LDTR’ and ‘matrix clause of the clause with LDTR’), resulting in nine possible combinations.

Table 2.

Clauses with LDTR and their matrix clauses: clause type combinations

Jurkiewicz-Rohrbacher and Kehayov (Reference Jurkiewicz-Rohrbacher and Kehayov2024) found that in Estonian and Moksha, complement clauses tend to be C1s, whereas relative clauses tend to be deeply embedded. Given that most of the clauses immediately preceding the clauses with LDTR are C1s (see Table 1), it is not surprising that Table 2 shows a high frequency of combinations where the first clause is a complement clause. Likewise, the low frequency of combinations with ‘relative’ in the first column of Table 2 can be attributed to the preference of relative clauses for deeper positions in the embedding structure.

However, the distribution in Table 2 seems too robust to be solely due to the correlation between embedding depth and clause type. Out of 27 clauses with LDTR in the sample 20 are C2s (Estonian n = 5, Moksha n = 4, Komi n = 11), and from these 16 (Estonian n = 5, Moksha n = 3, Komi n = 8) are subordinated to a complement C1. Meanwhile, the share of complement C1s from all C1s in our corpus of sentences with DECs is 49% in Estonian, 32% in Moksha, and 51% in Komi.Footnote ¹¹

The opposite is true for relative clauses. As shown in Table 2, a clause with LDTR was subordinated to a relative clause only once. In this case (from Komi), the clause with LDTR was C3, and its matrix clause C2. This means that none of the 20 C2s with LDTR were subordinated to a relative C1, although in our corpus of sentences with DECs relative clauses account for 20% of the C1s in Estonian, 19% of the C1s in Moksha, and 14% of the C1s in Komi.Footnote ¹² Future research should investigate whether and why clauses with LDTR seem to prefer complement and avoid relative matrix clauses.

Conversely, Table 2 indicates that clauses with LDTR are more often relative (n = 14) than complement (n = 3), which also is a hint for prospective research.

Note that the temporal reference does not vary; it is constant in all combinations of clause types. As specified in Section 4.1, all clauses in the first column of Table 2 have relative temporal reference and all clauses in the second column have LDTR. Three of the combinations in this table have already been exemplified. Examples (6), (7), and (11) illustrate the combination ‘complement–relative’. In (6) we see a combination of a that-clause and a non-restrictive relative clause, in (7) and (11) of a that-clause and a restrictive relative clause. Examples (9) and (10) show the combination of a complement and an adverbial clause. In (9), we have a that-clause and a concessive clause, in (10) a wh-complement clause and an adverbial clause of time. The combination ‘adverbial–relative’ is instantiated in examples (8) and (12). In both, the matrix clause is an adverbial clause of purpose, but in (8) the clause with LDTR is a restrictive and in (12) a non-restrictive relative clause.

The other attested combinations are shown in (13)–(15) from Komi. Example (13) demonstrates the ‘complement–complement’ combination: C1 is an infinitival complement clause and C2 a wh-complement clause. The temporal relations here correspond to Type I in Table 1 (shown in Figure 2 above). The event of calling happened in the past, the action called for (C1) is posterior to this event, and the gerund in C2 locates the SoAs ‘breaking the laws’ in the past relative to C.

Example (14) presents the only sentence where the immediate matrix clause of a clause with LDTR is a relative clause. The clause with LDTR is an adverbial clause of time. The temporal relations in this sentence match Type XII in Table 1; see Figure 9 below. LDTR appears at an embedding depth of three. C refers to the future from the moment of speaking (see the indexical expression in a week or two), but also C1 has absolute temporal reference. The events expressed in C and C1 overlap (even coincide), and if C1’s time were relative to the time of C, we would expect the present tense in C1. The future in C1 locates the contents of this clause after the utterance time. C2, on the other hand, has a relative temporal reference, conveying posteriority to the SoAs of C1. The light fur of grass stretches upwards after the land is covered with it. C3 is an until-clause with LDTR. The moment in which the SoAs of this clause occurs is within the time span of C2; it marks the endpoint of the duration of the SoAs of C2. Because of this temporal inclusion, if C2 were the reference point for the estimation of C3’s time, we would, again, expect the present tense in C3. Therefore, the future tense in C3 must be motivated by reference to the time of C1.

Finally, in (15) we see the combination ‘adverbial–complement’. The temporal relations here correspond to Type I (see Figure 2). C1 expresses a purpose aimed at the future from the past time in C. The complement clause C2 brings up a fact from the past that happened before the situation in C. It is not possible to interpret this sentence as having the SoAs of C2 ‘not leaving one in trouble’ occur before the situation in C1 but after the situation in C. Therefore, the SoAs of C1 cannot serve as an orientation point for the temporal reference of C2; instead, the orientation point is the SoAs of C.

The combinations not found in the data are ‘relative–complement’, ‘relative–relative’, and ‘adverbial–adverbial’. Their absence is likely due to the limited size of the sample.

5. Summary and implications of LDTR

We explored the temporal semantics of complex sentences and examined the long-distance temporal reference, a feature unique to deeply embedded clauses. We looked for factors in the embedding structure that evoke LDTR.

The frequency of LDTR in the samples we studied endorse the assumption that it is easier to spot the phenomenon in languages with more elaborate tense systems, like Komi. At any rate, the phenomenon is rare, occurring only in a few deeply embedded clauses, and given how uncommon sentences with DECs are, we assume that it appears in less than 1% of complex sentences in written language. Although Moksha and Estonian yielded only a few instances of LDTR in our data, they nonetheless contributed to our understanding of the phenomenon. We needed languages with different typological profiles to demonstrate that LDTR is not an idiosyncratic feature tied to specific structure. While Komi possesses a grammatical future tense, which – as shown in examples (6), (8), and (14) – offers a clear signal of temporal orientation to the future, Estonian and Moksha, which lack such a signal, still provide sufficient cues for identifying LDTR (see examples (9) and (11), both involving reference to the future). Furthermore, certain combinations of embedding depth, temporal reference, and temporal orientation presented in Table 1 were not attested in Komi at all; Types V, VI, and X appeared only in Estonian. Similarly, example (7) from Moksha was the sole sentence in our data demonstrating that LDTR is compatible with center-embedding.

Overall, we have too few examples of long-distance temporal reference to make definitive claims regarding its licensing conditions. Moreover, we observed a variety of combinations of morphological tenses and temporal orientations (to the past, present, future) in the data and could not identify any restrictions on LDTR in this respect.

Remarkably, though, in all examples the temporal reference of the immediate superordinate clause was relative. At the same time, in most examples LDTR occurred in C2, which was past, present, or future relative to C, whose temporal reference is absolute. One is tempted to assume that absolute temporal reference is a stronger anchor for the temporal reference of an embedded SoAs than relative temporal reference: an anchor, which overrides the adjacency of the immediately superordinate clause. But this conclusion is premature. We observed in Table 1 that in most cases where LDTR occurs at an embedding depth of three or more, the reference-point clause (usually C1) has a relative temporal reference. Therefore, we cannot conclude that LDTR is bound to the first clause upwards in the embedding hierarchy with absolute temporal reference.

We noticed that when LDTR occurs at an embedding depth of three or deeper in our data, the immediately higher clause encodes relative future. We also found that temporal reference can operate over a very long distance; in one sentence, the anchor for the temporal reference of a third-order embedded clause was the ultimate main clause. It is unlikely that this insight could have emerged from introspection or constructed examples alone.

As for clause type, we observed that when a DEC has a long-distance relative temporal reference, the immediately higher clause is likely to be a complement and unlikely to be a relative clause. Conversely, LDTR is more often hosted by relative than complement clauses. This can be partly explained by the fact that relative clauses become more frequent as embedding depth increases, while complement clauses become less frequent. However, the correlation between the type of the clause with LDTR and the type of its matrix clause is so strong in our data that it cannot be explained by this tendency alone.

We refrain from speculating about the functional (or other) constraints behind the observed tendencies, as these need to be tested and validated within a larger sample. Instead, we come back to the mere fact of the existence of LDTR. Deeply embedded clauses manifest a feature (LDTR) that never occurs with C1s. While other features of DECs are governed from their matrix clauses, their temporal interpretation might depend on an earlier clause in the embedding cycle. Does this observation challenge the notion of recursion mentioned at the outset of this research report?

The core tenet of recursion holds that a finite set of rules is sufficient to account for infinite embedding (see e.g. Chomsky Reference Chomsky1965:15–16). However, if the rules required to derive first-order embedded clauses are insufficient for generating deeper embeddings – necessitating an additional rule to account for LDTR – does this not imply that clauses at different embedding depths are derived from distinct rule sets?

More specifically, recursion is understood to involve the repeated application of a rule to its own output, where ‘[m]aterial introduced by any lower recursive cycle is always contained in the material introduced by the immediately higher cycle’ (Karlsson Reference Karlsson and van der Hulst2010b:51). In syntactic theory, recursion refers to ‘embedding a constituent in a constituent of the same type’ (Pinker & Jackendoff Reference Pinker and Jackendoff2005:10), or, in a slightly different formulation, embedding a constituent in a constituent ‘of the same complexity’ (van der Hulst Reference van der Hulst and van der Hulst2010:xxi).

Recursion thus implies that the choice of structural features available to each next clause in the embedding cycle remains constant, that is, the same (distributions of) features should be available to clauses at any embedding depth. This is why even comprehensive grammars of major European languages typically address clausal subordination by discussing only first-order embedded clauses: applying the same procedure results in the same structure, so one should not expect to find anything new in second-order (or deeper) clauses compared to first-order embedded clauses.

However, linguistic evidence indicates that this expectation does not fully hold: the depth of embedding does, to some extent, affect the grammar of subordinate clauses (Jurkiewicz-Rohrbacher & Kehayov Reference Jurkiewicz-Rohrbacher and Kehayov2024). In the Chomskyan tradition of thought, structural tendencies associated with the increase in embedding depth have been explained in terms of processing limitations external to the syntax. In other words, differences between clauses that correlate with their depth of embedding have been solely attributed to the greater processing demands imposed by deeper structures (see Karlsson Reference Karlsson2010a). Yet we observe LDTR – a feature available to deeply embedded clauses but unavailable to C1s. This is a qualitative shift rather than a mere quantitative variation in structural preferences. Moreover, it would be counterintuitive to claim that LDTR arises as a response to processing limitations, given that long-distance dependencies tend to increase, rather than alleviate, cognitive load.

Recursion also implies that the properties of sentences with multiple embeddings can be reduced to bi-clausal relations between an embedder and embedded clause, that is, the properties of three-, four- etc. clausal constructions are predictable from the properties of bi-clausal constructions (Letuchiy Reference Letuchiy2021:525–529). However, this is not what we observe in our data. A second-, third- or higher-order finite embedded clause has properties depending on its immediate superordinate clause but may also have properties (like LDTR) governed by a clause higher up in the hierarchy.

One might argue that the foregoing objections overlook a crucial point: in the case of LDTR, we have to do with semantics rather than syntax. In Generative Grammar and its derivatives, semantics and syntax are distinct components of language. The recursive property of language is primarily stated for syntactic structure, so temporal semantics cannot provide hard evidence against recursion.Footnote ¹³ Still, whatever mechanism is posited for LDTR within this tradition of thought, it would have to be able to bypass the hierarchically intervening clause and make reference to a higher clause.

When the aforementioned complexity condition is taken into account, the problem becomes more evident. Recursion presupposes that clauses cannot be more complex than the clauses they are embedded in. But DECs exhibit two types of relative temporal reference: relative to their immediate matrix clause and relative to a higher clause. Aren’t they more complex than first-order embedded clauses, for which only the first type of relative temporal reference is available? Isn’t the embedded more complex than the embedder, if a structural choice is available to the former but not to the latter? Furthermore, isn’t the long-distance temporal reference cognitively and structurally more complex than the short-distance relative reference?

In this study, we examined how distant the anchor (or antecedent) of temporal reference can be, and what factors constrain or facilitate such distance in discourse. Future research should situate LDTR within the broader context of other indexical categories, such as pronouns and evidentials. The analogy between temporal and pronominal reference, originally proposed by Barbara Partee (Reference Partee1973), has developed into the binding approach to tense, which remains highly influential in formal linguistic theories. More recently, the view that evidentials are inherently indexical has gained traction in semantic research (e.g. de Haan Reference de Haan, Cienki, Luka and Smith2001, Hanks Reference Hanks2012, Bergqvist Reference Bergqvist2023). Like pronouns and tense, evidentials exhibit deictic properties (they can refer to aspects of the utterance situation), are context-dependent (e.g. with respect to epistemic authority – who knows what, and how), can convey reference shift (e.g. in reported speech), and have anaphoric potential (referring to previously established sources or modes of knowing). These similarities warrant the investigation of LDTR across a broader range of deictic categories.

With these insights and questions, we pass the baton to future research. We hope to have contributed to a more comprehensive understanding of long-distance relations in both semantics and syntax.