2.1 The perfect as a universal category

General studies in comparative linguistics or language typology have revealed that the perfect is a universal grammatical category, found in a large number of natural languages. While the present paper does not provide a formal semantic derivation of perfect interpretations, a substantial body of work has developed explicit compositional analyses of perfect constructions (see e.g. Verkuyl 2022). The discussion in this section is intended to remain at a descriptive and theory-neutral level, as the syntactic analysis proposed below does not hinge on a particular semantic implementation. The realization of the perfect differs cross-linguistically both formally and in range of meanings. Several studies (Comrie 1976; Anderson 1982; Bybee 1985; Dahl 1985; Bybee et al. 1994; Lindstedt 2000) advanced the understanding of the perfect and its universal properties.

A defining feature of the perfect is the current relevance of a past situation (cf. Comrie 1976: 52; Bybee et al. 1994: 54; Lindstedt 2000: 366). The perfect is used when speaking about something that has occurred or existed previously but is significant in the present; “present" here refers both to the actual moment of speech and to a generalized present. Often, the reason for this relevance is that a new situation or state has arisen from the past event, which is now being discussed.

By using a form corresponding to the perfect, the speaker links the past event to the present, but does not link it to other past events (it is detached from other past facts, Lindstedt 2000: 366, citing Maslov 1990). From this follows a second, negative criterion for identifying the perfect: it is not used to narrate the past, i.e., to tell a story composed of interconnected past events (Lindstedt 2000: 366; Bybee et al. 1994: 54).

Many Indo-European languages exhibit compound present perfect forms with more restricted uses, with English present perfect being the best studied one (Alexiadou et al. 2003; Mittwoch 2008; Dahl 2021; among others). Recently, a variety of studies have addressed variation in the uses of the perfect in Germanic and Romance languages, mainly from the perspective of formal semantics (Lindstedt 2000; Portner 2003; De Swart 2007; Bertrand et al. 2022; Van der Klis et al. 2022). This article differs from previous studies in two respects. From a theoretical perspective, it does not aim at defining temporal and aspectual uses of the perfect from a semantic perspective, but it represents an application of the syntactic operation Agree (Chomsky 2001), used to capture the syntactic computation of the elements that make up the compound perfect form (the present auxiliary and the past participle form), as also proposed, in different fashions, by Arsenijević (2006) and Biskup (2024); the analysis is embedded in the Minimalist generative framework (Chomsky 1995). From an empirical perspective, the article focuses on the Latvian present perfect (henceforth “perfect”; Nau 2005; Arkadiev & Wiemer 2020; Arkadiev & Daugavet 2021) in the indicative mood and active voice. The Latvian perfect has been widely overlooked in formal linguistic approaches to tense and aspect; however, it will be shown that it exhibits the properties of a prototypical perfect, in its core use, allowing for formal variations in the spell-out of the auxiliary. We will propose that such variation depends on an underlying difference in the syntactic computation, analyzing it in terms of feature valuation and interpretability, adopting an exoskeletal view of syntax.

2.2 The meaning of the Latvian perfect

The Latvian perfect exhibits distinct surface realizations, which differ in both interpretation and morphosyntactic structure. At a purely formal level, the Latvian perfect can appear in three forms: the full perfect (1a) (present auxiliary + active past participle); the oblique perfect (1b) (invariable evidential auxiliary + active past participle); the short perfect (1c) (“bare" active past participle).

Only the full and oblique perfects are unambiguous in their meaning. The full perfect is traditionally analyzed as a regular perfect, while Latvian grammars generally analyze the oblique perfect as an instance of “reported" mood. Here we will propose that the two forms are built from the same syntactic blocks and the distinction is captured at the featural level. Short perfects raise a well-known classificatory issue in the literature. One possibility is to treat them as reduced variants of the full perfect with a dropped present auxiliary, or of the oblique perfect with a dropped evidential auxiliary. Another possibility is to analyze them as evidential counterparts of the simple past (Holvoet 2001). In this paper, we adopt neither auxiliary-drop analysis. Short perfects will be treated as a distinct syntactic configuration lacking tense features and auxiliary altogether. At the same time, we acknowledge that short perfects can function as evidential equivalents of past tense forms in discourse. This evidential use, however, does not entail formal identity with the simple past, as short perfects remain participial constructions. Throughout the paper, the term “perfect” is used as a descriptive cover term for predicative uses of active past participles in Latvian, following a long tradition in descriptive grammars. We do not assume that these constructions form a semantically homogeneous category. In particular, they may encode perfect, resultative or evidential meanings, or combinations thereof. Crucially, existing work on the Baltic perfect (Arkadiev & Daugavet 2021; Daugavet & Arkadiev 2022; Kapkan 2024) emphasizes a distinction between resultative and current relevance uses. We do not deny the importance of this distinction. However, for the purposes of the present paper, both are subsumed under a broader category of ‘perfect-related’ interpretations, as the syntactic analysis proposed here does not hinge on this finer-grained semantic split.

The core meaning of the Latvian full perfect involves reference to something that has happened previously, but the focus is on the current relevance of the event itself, which is therefore assumed to be somehow significant at speech time; in this context, the perfect generally appears in its full compound form, with a present tense auxiliary and a past participle, both agreeing with the subject of the sentence, as shown in (2):

The situation described with the perfect is linked in the text to present events, as the word tagad (‘now’) highlights: (2) clearly expresses a contrast between the past (described using the simple past) and the present. These linguistic facts support the view that the perfect in Latvian should be classified as a present tense, even though it includes the meaning “something that happened earlier”.

The actual time when the event occurred is not specified: sentences using the current relevance perfect never contain past temporal adverbials that refer to a specific time (such as dates, years, or clock times) and at least with respect to the core meaning of the perfect, it can be confidently stated that such temporal expressions are not compatible (3). Conversely, temporal expressions are fine with a past tense, as shown in (4).

A closely-related meaning of the Latvian full perfect is the experiential one (5):

Like in the case of current relevance, (5) refers to something that has occurred or existed at least once in the past, until the present moment, as highlighted by the use of dzīvē (‘in life’), which links the past with the present, as it implies “both in the past and now”. The same is true for the adverb nekad (‘never), implying “not in the past and not now” (6).

Also in the experiential perfect meaning, the present perfect appears in its full form.

In addition, Latvian allows universal perfect constructions to refer to extended temporal intervals under cumulative or lifetime readings.³ This reading is typically licensed by adverbials such as since. As in previous contexts, the universal perfect requires overt temporal anchoring and in Latvian it is expressed with the full perfect form (7).

As it was presented at the beginning of this section, Latvian exhibits also oblique and short perfect forms. Such perfect forms are generally used when talking about past events the speaker has learned about indirectly, either inferred from visible circumstances or heard/read about, as discussed for perfect forms in some other world languages (Bybee et al. 1994). Importantly, neither the oblique nor the short perfect can normally be used in the core perfect readings, which generally emerge in presence of the full perfect with an agreeing auxiliary.

In Latvian, the oblique perfect (Kalnača & Lokmane 2021) construction is used to indicate that information about an event was obtained from other people or written sources, that is, the reported reading (see Izvorski 1997). In this context, the auxiliary verb form is replaced by the invariable form esot (8).

In the literature, such constructions are sometimes analyzed as a separate oblique mood (Kalnača & Lokmane 2021); however, in line with Holvoet (2007), we will propose a more fine-grained analysis of this form, which qualifies as a perfect, with the modal meaning specifically signaled by the special invariable form of the auxiliary esot. Importantly, the -ot morpheme that marks this evidential value is not confined to the present stem es-, but can also attach to the future auxiliary stem būs-, as in (9). This suggests that the oblique auxiliary is not tenseless but rather carries an abstract tense feature. The evidential meaning arises from the -ot suffix, while the auxiliary stem itself continues to encode temporal reference (9).

As noted in Kalnača & Lokmane (2021), the invariable form esot is often dropped, giving rise to a “short” form of the perfect, made up of a bare past active participle, without any auxiliary verb (10).⁵

As shown in (10), such forms freely allow for definite past expressions to co-occur with the short perfect, unlike the case of the full perfect. Besides, the perfect event is not related to the present via current relevance, while it is connected to other events in the past, forming a narration. Nau (2005) treats these forms as instances of a reported reading and does not classify them as genuine perfects, placing them on a par with the esot forms. However, unlike the oblique esot perfect, the auxiliary is absent altogether in the short perfect; as we will discuss in the analysis, the lack of an overt auxiliary signals the absence of any tense feature in the structure. Nau (2005) acknowledges that the boundary between perfect and reported uses is not sharply defined: it is not always clear whether the speaker’s intention is to report someone else’s words and she notes the possibility that reported contexts may instead reflect a relative tense reading. In other words, the existing descriptions leave the status of these forms somewhat indeterminate, which reflects the empirical difficulty of pinning down the precise contribution of perfect morphology in these contexts. It is precisely here that corpus evidence becomes essential: by providing a systematic account of distribution and interpretation, we are able to sharpen the definition of the Latvian perfect and draw a clearer boundary between core perfect and reported readings.

In what follows, we examine evidence from the BASTA corpus to uncover the frequency and functional range of each perfect type.

3 Corpus data

This section presents an analysis of the Latvian perfect tense based on the BASTA (Baltic and Slavic Tense and Aspect) corpus (Klimek-Jankowska et al. 2026). The corpus uses a Translation Mining methodology (Van der Klis et al. 2017) to analyze cross-linguistic variation in the use of the perfect by comparing parallel translations of three chapters (1, 16, 17) of J.K. Rowling’s Harry Potter and the Philosopher’s Stone in Slavic and Baltic languages. Here we focus on the Latvian data in the corpus (from the Latvian version of the book, Harijs Poters un Filozofu akmens, Rowling 2021). In total, we extracted 54 examples annotated as perfect, which we additionally labeled according to perfect types (full, oblique and short), whether they are used to express a reported reading and whether they convey a core perfect reading (current relevance, experiential, universal, or none).⁷ The data reveal strong correlations between these features that help clarify the distribution and semantic contribution of different perfect forms.⁸

The distribution of perfect forms is summarized in Tables 1, 2 and 3:

The tables provide a general overview of the distribution of counts across perfect types (Table 1), reported reading (Table 2) and core perfect readings (Table 3), offering a broad summary of the data.⁹ These cases account for the apparent mismatch between the totals in Tables 2 and 3. However, examining the correlations between these variables reveals a more nuanced and detailed picture. By analyzing how different perfect types relate to the reported and to the core perfect readings, how the reported use intersects with such readings and how perfect types interacts with the reported use, we uncover specific patterns and distributions that are not apparent from the summary alone.

Table 4 shows that the full perfect type is always associated with core perfect readings (current relevance, experiential and universal). The oblique and short perfects are never associated with such readings.

Table 5 shows that the full perfect type is not used in the reported reading, while the short perfect mostly occurs with the reported reading, with only two occurrences being used in non-reported contexts. The oblique esot perfect type appears only with the reported reading (4 counts).

The results reveal a striking pattern in how perfect forms function in Latvian. Full perfects appear exclusively in non-reported contexts and strongly correlate with core perfect readings: current relevance (11), experiential (12) and universal (13).

In contrast, short perfects are excluded from core perfect readings and are almost entirely restricted to reported readings¹⁰ (14). Only in two cases, the short perfect was used in non-reported reading. In these contexts, the short perfect is found in an embedded clause and its tense reading allows for a past perfect interpretation (15).

Examples (14) and (15) show that short perfect forms may receive experiential or resultative interpretations (as also shown in Daugavet 2023) in embedded or reportative environments. We take these cases to involve pragmatic enrichment, facilitated by embedding under verbs of saying or thinking, rather than grammatical encoding of present perfect meaning by the short perfect itself.

A similar trend holds for oblique perfects, which also occur only in reported contexts (16).

Notably, the presence of a core perfect meaning is heavily tied to non-reported information. Reported contexts show a clear dissociation from current relevance, experiential and universal readings.¹¹

Overall, the corpus data point to a systematic division of labor between perfect types. The full perfect exclusively supports canonical perfect interpretations, confirming its role as the default marker of current relevance, experiential and universal readings. By contrast, the short and oblique perfects almost categorically pattern with reported readings and they never appear in contexts requiring current relevance. This asymmetry strongly indicates that auxiliary choice and auxiliary omission in Latvian perfects are not stylistic variants but encode different featural configurations. In particular, the distribution motivates an analysis in which the full perfect involves genuine tense anchoring, whereas the short and oblique perfects lack such anchoring and instead introduce evidential features.

4 Theoretical background

The empirical generalizations established in Sections 2 and 3 raise a theoretical question: how can the variation among Latvian perfect forms be captured in terms of syntactic feature configurations? To address this, it is necessary to establish the theoretical assumptions that will guide our account, before turning to the analysis of the Latvian perfect. The analysis is couched in a Minimalist framework with an exoskeletal bias, in the sense that syntactic structure is built independently of lexical content. This does not entail the adoption of a DM framework. Rather, it reflects the assumption that functional structure and feature composition play a primary role in determining interpretation, while lexical items are inserted into pre-established syntactic positions. Therefore, the goal of this section is not to present new proposals, but to gather in one place the key ideas from the literature that we will rely on. The subsections that follow review exoskeletal approaches to grammar, the distribution of tense and aspect features on verbal heads and recent work on feature valuation, interpretability and agreement.

The discussion proceeds in three steps. First, we outline the assumptions of an exoskeletal, syntax-first model of grammar, which provides the general architecture for our approach. Second, we consider proposals about the distribution of tense and aspect features on verbal heads, which will be central to explaining the contrasts among the full, oblique and short perfect. Third, we review the mechanisms of feature valuation, interpretability and agreement that allow us to connect morphosyntactic variation with semantic interpretation. Together, these assumptions establish the framework within which the Latvian data will be analyzed in Section 5.

4.3 Summary

To summarize, Section 4 has reviewed a set of theoretical proposals that together provide the toolbox for our analysis of the Latvian perfect. We briefly restate the points we drew from the literature and explain why each is needed for the analysis that follows.

From the exoskeletal tradition (Borer 2005; 2014) we take the architecture-level claim that syntactic structure is projected by functional heads while lexical roots are conceptually rich but grammatically unmarked until merged. This syntax-first perspective motivates our focus on features (rather than lexical idiosyncrasies of participles or auxiliaries): the empirical alternations in Latvian should be captured by different feature-configurations in the clausal spine, not by distinct lexical entries. From Minimalist work on labeling and economy (Chomsky 2013) we also retain the idea that Merge and feature-driven operations are the central engines of syntactic building, with labeling and interface pressures determining which feature relations (e.g. between Asp, v and T) are syntactically privileged.

Within this architecture, Wurmbrand’s (2012) proposal that verbal heads (V, v) may bear uninterpretable tense or aspect features provides the first step for our analysis. It gives us a way of treating participial morphology as hosting tense/aspect features that interact with higher heads. We do not adopt Reverse-Agree as Wurmbrand formulates it, but we retain the insight that V/v can carry temporal properties, which is crucial for analyzing the interaction between participle and auxiliary in the Latvian perfect.

Next, the notion of parasitic agreement (e.g. De Vincenzi & Di Domenico 1999; Van Urk 2015) helps explain why some features appear only when carried along by other relations. This perspective is particularly useful for modeling the Latvian participial tense feature: it is not independently licensed but can be transmitted parasitically together with aspectual agreement to higher heads, where it is ultimately resolved.

Zeijstra’s (2014) reinterpretation of interpretability as a formal syntactic property provides the framework for keeping uninterpretable features present and available for valuation, without causing the syntactic structure to crash. This flexibility allows us to treat participial tense features as uninterpretable but still operative, a key step in defining the core perfect reading syntactically.

Pesetsky & Torrego’s (2007) model then supplies the mechanics for multi-step agreement chains. By treating agreement as sharing rather than simple value transfer, it explains how multiple unvalued features - for example, on the participle, Asp and v - can all remain active until resolved by a valued feature on T. This provides exactly the structure we need to derive the contrasts between the full, oblique and short perfect.

Finally, Rezac’s (2010) last-resort feature insertion is used here as a constraint on feature introduction: otherwise inert features may be activated only when required to complete a syntactic dependency, not to prevent semantic failure at the interfaces. We will use this mechanism to capture why the participle can supply a tense feature when the auxiliary is absent: its insertion is not an idiosyncratic lexical fact, but a formally constrained repair strategy licensed by the grammar. Crucially, in our approach, last-resort insertion of features applies only when an otherwise well-formed syntactic structure would crash due to an unvalued but interpretable feature at the phase edge.

Taken together, these strands supply a coherent, feature-centered architecture: (i) the exoskeletal syntax locates features on functional heads rather than lexical roots; (ii) verbal heads may enter with T/Asp-type features; (iii) uninterpretable features can remain active until valued (Zeijlstra 2014); (iv) agreement proceeds via feature valuation (Pesetsky & Torrego 2007); and (v) last-resort insertion is available only in tightly constrained contexts (Rezac 2010). The approach adopted here is also closely related to the Morphology-as-Syntax framework (Collins & Kayne 2023; Zyman 2026), which similarly rejects a rich pre-syntactic lexicon and derives morphological structure through syntactic operations. Our proposal is compatible with this perspective in treating morphological exponents as the spell-out of syntactically derived feature configurations, without committing to a specific implementation of roots or vocabulary insertion. With this apparatus in hand, we can now derive the three Latvian perfect forms as distinct configurations of presence/absence and valuation/non-valuation of tense, aspect and evidential features.

5.1 The current relevance, experiential and universal (full) perfect

We propose that in the core perfect readings, the underlying featural representation of an auxiliary corresponds to an interpretable tense feature, valued as present; we represent it as T:pres. Conversely, the participle is featurally represented as an interpretable aspect feature Asp, which can bear different values: in this case, we assume that the participle carries a perfective aspect feature Asp:pfv (see Cinque 1999); the participle also parasitically carries an uninterpretable tense feature, valued as past; we represent it as uT:past.¹⁷

The Latvian current relevance perfect is syntactically computed as follows. In line with what discussed in Section 4.2.1, we propose that V and v are merged with some verb-related features. In this case, we assume that V is merged as a feature bundle, minimally specified for an interpretable aspect feature Asp:pfv and an uninterpretable tense feature uT:past. Conversely, an interpretable present tense feature T:pres is merged in v. AspP, encoding an interpretable unvalued aspect feature Asp:_, is merged over vP (18). No claim is made that individual features are merged separately; rather, V is fully specified for these features, in line with exoskeletal assumptions.

The Asp head probes down looking for a goal carrying an interpretable aspect feature, finding Asp:pfv in VP: Asp:pfv values Asp:_ on the Asp head. The uninterpretable past feature on the participle parasitically raises to Asp together with the aspect feature (19).

TP, encoding an interpretable unvalued tense feature T:_, is finally merged (20). Although multiple interpretable tense features are present in the structure, we assume that semantic temporal interpretation is determined only at the highest syntactic position where a valued interpretable T-feature is present, namely T.¹⁸ Lower interpretable tense features (e.g. on v) are formally interpretable but do not contribute independently to temporal interpretation, in line with the view that interpretability is a syntactic property rather than a guarantee of semantic interpretation (Zeijlstra 2014).

The T head probes down looking for a goal carrying an interpretable tense feature. The first such feature it finds is T:pres on the v:¹⁹ T:pres values T:_ on T as present, as shown in (21).

At PF, the T:pres feature on T is spelled out as the present auxiliary (aux in (21)), while the Asp:pfv, uT:past bundle in AspP is spelled out with perfective participial morphology (ptcp in (21)). Since both T and AspP valued φ-features, full agreement is morphologically realized.

The structure in (21) is faced with a problem at this point. The uninterpretable past feature carried by the participle parasitically raised to Asp but does not participate in any agreement relationship.

In Chomsky’s (2001) model, this is a problem, as uT:past would interfere with the interpretation of the sentence. We argue that it is precisely the presence of a past feature on the participle that makes it possible to derive the current relevance reading. Following Zeijlstra (2014), the difference between interpretable and uninterpretable features is only formal in nature and therefore only visible to syntax. Feature checking does not lead to deletion of uninterpretable features and their presence at LF does not make the structure crash. Similarly, interpretable features are formal syntactic properties that fulfill a checking requirement imposed by the presence of uninterpretable features. In this model, the uninterpretability of features is not defined as a semantic property, but just as a formal requirement that the uninterpretable feature needs to stand in a c-command relation with an element carrying a semantically interpretable counterpart of the same feature. The uninterpretable feature on Asp is therefore not problematic for the semantic component. We propose that an additional mechanism is established between T:pres on the auxiliary and the uT:past on the participle. Following Partee (1973), the participle functions as a bound variable. In binding terms, uT:past turns the participle into an anaphoric tense, which, as per Condition A, is bound in its local domain and interpreted as past relative to speech time expressed by T:pres on the present auxiliary.

5.2 The oblique perfect

For the second perfect type, the oblique perfect with the invariable auxiliary esot, we propose that the syntactic computation follows parallel steps, but the type of features involved in the agreement operation exhibit fine-grained differences, resulting in the reported reading associated with the construction.

In particular, our analysis needs to exclude the current relevance, experiential and universal interpretations and situate the event described by the perfect form in the past.

We argue that the difference is not to be searched in the way the structures are computed, but rather in the featural composition of auxiliaries. Recall that parasitic features are uninterpretable features activated as a last resort and cannot independently probe or value interpretable features. Syntax distinguishes parasitic from non-parasitic features via their formal specification, not via contextual inference.

Crucially, we propose that the lack of current relevance, universal or experiential readings for the oblique perfect is captured syntactically as the lack of a present value on the invariable auxiliary esot; the auxiliary carries an interpretable feature T:_, which is however not able to establish a relationship with the past tense expressed by the participle because it lacks a tense value itself.

The participle is first merged in V, carrying an interpretable aspectual feature valued as perfective, Asp:pfv and an uninterpretable but valued tense feature uT:past. Above this, v is merged, bearing an interpretable but unvalued tense feature T:_ and a valued uninterpretable modal feature uMod:evid. The valued past tense feature on V is parasitic and uninterpretable. As such, it cannot serve as a valuer for the unvalued T-feature on v, despite structural proximity. This follows from the general restriction that uninterpretable features cannot value interpretable ones. Over vP, an aspectual projection AspP is merged, where Asp carries an unvalued aspect feature Asp:_, as shown in (22).

The aspectual head in AspP probes its c-command domain for a matching interpretable aspect feature, locating Asp:pfv on the participle. This feature values the aspect probe in Asp as perfective. In addition, the uninterpretable tense feature uT:past parasitically raises to Asp along with the valued aspect feature. This Agree operation results in the configuration in (23).

Subsequently, TP is merged above AspP. The T head introduces an unvalued tense feature T:_, as shown in (24).

T then probes its c-command domain for a matching goal. The first matching feature it finds is T:_ on v. In Pesetsky & Torrego’s (2007) terms, feature-sharing applies; however, this interpretable feature is also unvalued, so agreement does not result in value assignment and the probe in T remains active. Nonetheless, the complex feature bundle [T:_, uMod:evid] on v parasitically raises to T as a result of feature sharing, resulting in the configuration in (25).²⁰

At this point, the structure would not converge, since T requires a valued tense feature but only encounters an uninterpretable one on Asp. Following Rezac (2010), however, uninterpretable features can exceptionally be activated as a last resort when needed to rescue the structure.²¹ We analyze the participial uT:past feature in precisely this way: although normally inert, it can be pressed into service under constrained conditions to provide the value that T requires. Therefore, T probes again and encounters the exceptionally activated uninterpretable tense feature uT:past on Asp. This probe-goal relation allows T to be valued as past, as shown in (26).

Following Cinque (1999), a ModP layer is merged above TP. The Mod head carries an interpretable unvalued modal feature Mod:_, as shown in (27). Mod:_ on Mod is interpretable, as it contributes evidential meaning.

The modal head probes its c-command domain and, being unable to find an interpretable goal, it identifies the uninterpretable modal feature uMod:evid on T for Last Resort Agree. Importantly, T probes separately for tense and modal features, as a feature-specific cyclic Agree: valuation of the tense feature activates T, enabling a subsequent probe for modal features. This interaction allows Mod to be valued as evidential, as illustrated in (28).²²

This final representation captures the steps involved in licensing the evidential perfect. The reported interpretation arises through the valuation of Mod by an evidential feature originally introduced on v, while the temporal interpretation as past is mediated via the feature-sharing dependency between Asp and T. At PF, the tense head valued as T:past but bearing an evidential feature is realized as the invariable auxiliary esot. The lack of agreement morphology on the auxiliary follows from the absence of valued φ-features on this head, while agreement on the participle remains unaffected.

5.3 The short perfect