Constructional sources of durational shortening in discourse markers

1 Introduction

Interest in the study of discourse markers has rapidly grown over the last decades since Schiffrin’s (1987) seminal paper. The bulk of the research has identified recurrent crosslinguistic semantic ∼ pragmatic changes that take place as the lexical source for the discourse marker grammaticalizes (consider, e.g., Traugott’s theory of [inter]subjectification [Traugott 1982, 2003, 2010]). The grammatical and semantic changes involved in the creation of discourse markers are widely agreed upon. In contrast, the associated phonetic and phonological changes of the grammaticalizing forms have not received much attention in previous literature (Schubotz et al. 2015: 376). Studies of discourse markers that do include analyses of pronunciation variation (e.g., Bybee and Scheibman 1999; Bybee et al. 2016; Ernestus and Smith 2018; Gonen et al. 2015; Plug 2005) report differences in realizations between the discourse marker and its lexical source.

Previous studies have established a link between discourse markers and articulatory reduction. Compared to lexical items with propositional meaning, units that guide communicative inferences and frame discourse exhibit greater phonetic reduction. Segment lenition is a feature often cited for discourse markers in a variety of languages (Drager 2016; Local 2003; Rivas and Brown 2010), as well as segment or syllable deletion (Bybee and Scheibman 1999; Ernestus and Smith 2018; Plug 2005). Discourse markers not only evidence greater lenition and deletion of segments but compared to non-discourse marker counterparts are also typically durationally shorter (Gonen et al. 2015; Local 2003; Martinuzzi and Schertz 2022). For example, Bybee et al. (2016) note that in addition to greater reduction and deletion of segments, the mean duration of the onset consonant /s/ in the Spanish discourse marker o sea ‘I mean’ is shorter than in the lexical source (third-person singular subjunctive sea). This durational shortening can apply to individual segments and to overall duration of the lexical item (Gonen et al. 2015).

These differences in the degree of reduction and lenition between discourse markers and lexical sources, in fact, hold even when comparing ‘homophonous’ or etymologically linked lexical items. For example, in a study of English, Schubotz et al. (2015) examine the word pair you know used both with a propositional meaning (‘to have knowledge of’) as well as with a discourse function. The discourse marker is realized with a reduced vowel as opposed to a full vowel to a greater degree than when the word dyad is used with a literal sense or meaning.

This widely reported finding of greater reduction in discourse markers over lexical counterparts has been given different explanations. One linguistic factor commonly considered as a source for the variants of the word forms produced by speakers is lexical frequency. The connection between reduction and high frequency is well established, whereby high-frequency words reduce to a greater degree and before words of lower lexical frequency (Bybee 1999; Jurafsky et al. 2001; Phillips 2006). In cases of discourse markers specifically, the greater degree of lenition and shortening is often attributed to a higher frequency of the discourse marker compared to its lexical source (Bybee et al. 2016; Schubotz et al. 2015). Frequency, therefore, has been identified as a source for the reduction, be it through articulatory routinization (Bybee 2001), greater use in casual contexts (D’Introno and Sosa 1986; Gonen et al. 2015) or by facilitating the accumulation of production biases (Bybee et al. 2016). These results can be accounted for within an exemplar model whereby experiences with words directly shape the lexical representation in memory (Bybee 2001; Pierrehumbert 2001).

Relatedly, increased frequency accompanies the grammaticalizing forms (Bybee 2003) that give rise to discourse markers. In very specific contexts of use, words and word combinations that serve as sources of discourse markers undergo decategorialization, loss of propositional meaning, and phonological reduction (Hopper and Traugott 2003). With the loss of propositional meaning of the word or word combinations, these items acquire textual and (inter-) subjective meanings (Traugott 2010). Unlike textbook cases of grammaticalization, discourse markers broaden their syntactic scope (Lehmann 2002). Nevertheless, the development of discourse markers entails processes closely associated with grammaticalization, such as decategorialization, semantic bleaching, subjectification, divergence, persistence, and layering (Brinton 2007). Discourse markers, thus, may be argued to exhibit greater phonetic and phonological reduction as part of the grammaticalization process.

In this study, we provide evidence of articulatory reduction of a discourse marker in Spanish (es decir ‘that is to say’) that cannot be straightforwardly accounted for by appealing to these explanations. We examine the Spanish word decir ‘to say, to tell’, which is used conversationally both as a lexical verb with propositional meaning and as a discourse marker with procedural meaning within the reformulating construction [es decir] ‘that is to say, I mean’. We will show that, although it is neither highly grammaticalized nor frequent, reformulator decir undergoes phonological reduction to a greater degree than the more frequent lexical source decir ‘to say’.

Through careful consideration of conditioning factors known to constrain durations of words in target contexts (i.e., speech rate of the target-word context, predictability of following words, number of phones, distance from pause), we demonstrate that durational differences between tokens of decir cannot be explained solely by conditioning factors of the target context. Even considering the possibility of the accumulation of these factors in memory, the identity of the target type, whether it be decir with propositional meaning or decir within the construction [es decir], significantly predicts duration.

This result suggests separate lexical representations for the types of decir, each containing fine-grained phonetic detail reflective of use. Finding that specific lexical items accumulate effects of usage patterns that appear in subsequent productions is in line with recent work (e.g., Tang and Shaw 2021). What’s more, however, our findings allow us to propose an alternative explanation for the reduction of es decir and discourse markers generally. We argue that the reformulator construction itself contains lexical specifications derived through interactional uses (Bybee and Thompson 2022), which in this case promote articulatory reduction including durational shortening.

2 Background

3 Data

To determine whether significant acoustic differences exist within the non-finite verb decir used both inside the reformulator construction [es decir] as well as outside of it, it is essential that we analyze naturalistic, spoken data of Spanish. The limited dialectal information we have (Castillo Fadić and Sologuren Ínsua 2017: 90) suggests that the use of [es decir] as a discourse marker in spoken Spanish may be more frequent in Spain than in other varieties. As such, we employ in the analyses data extracted from two corpora of Peninsular Spanish.

The first corpus is the publicly available Project for the Sociolinguistic Study of Spanish from Spain and America (PRESEEA). This corpus is a collection of spoken Spanish representative of different dialects and social groupings of Spanish speakers. Within this project, nevertheless, we limit our data searches to two geographically contiguous Spanish varieties: Madrid and Alcalá de Henares. The former data set is comprised of approximately 185,000 words across 37 different speakers. We extract data from approximately 15 h of recordings. The latter corpus is made up of approximately 150,000 words spoken by 41 different speakers. The data represent approximately 14 h of conversations. From these two sources of Peninsular spoken Spanish, we extract all examples of decir for which there is audio (N = 52).

The second data source is the Nijmegen Corpus of Casual Spanish (Torreira and Ernestus 2012). This corpus is made up of approximately 393,000 words spoken by 52 different men and women from Madrid. The Nijmegen corpus provided approximately 30 h of speech for analysis (N = 336). A summary of the data and speakers is provided in Table 1.

Overall, thus, the tokens for this study come from approximately 60 h of conversational data from Peninsular Spanish.

4 Methods

From these corpora we extracted all fluent, audible tokens of decir without speaker overlap. In order to consider potential effects of target context rate on target productions, we also extracted audio from the pause delimited contexts. Pauses were defined as the cessation of speech >50 ms. The dependent variable for this work was the target duration of all tokens of decir spoken in the corpora. The acoustic inspection of the data was carried out in the phonetic analysis software Praat (Boersma and Weenink 2019). Using synchronized waveforms and spectrographic displays, each token of decir was manually delimited to calculate duration in seconds. In continuous speech, the phonetic contexts preceding and following the target tokens are variable, as are the realizations of the onset consonant /d/ (Amastae 1989; Barrutia and Schwegler 1994; Brown 2018; Cole et al. 1999; D’Introno and Sosa 1986; Eddington 2011; Waltermire 2010) and terminal phone /ɾ/ (Bradley and Willis 2012; Henriksen 2014, 2015; Henriksen and Willis 2010; Hualde 2005). For tokens of decir with an initial consonant realized as a voiced, dental stop (N = 9), the word duration was measured starting at the closure. When the word-initial voiced, dental consonant was realized as an approximant [ð] (N = 302), the word duration calculation began at onset of evidence of constriction as seen through intensity and formant structure. Cases lacking a discernable closure (N = 77) were coded as deleted /d/ targets. The onset of the target was marked at the first regularly repeating cycle of the vowel /e/ in decir. If the preceding lexical context was /e/ (N = 64), the start of the decir target word was calculated from the midpoint of the contiguous vocalic segments. The end of the word was marked at the end of the terminal phone. Tap realizations occur word- and syllable-finally and word-finally when resyllabified into the following vowel-initial word. Target tokens of decir articulated with a tap (N = 214) were marked at the end of the break in the spectrogram (with accompanying reduction of amplitude in the waveform). Target tokens realized without a clear apical occlusion of /ɾ/ (N = 174) were marked at the offset of the short constriction phase prior to greater opening of the following vowel (as determined by definition of the formants and relative intensity) or at the onset of a following consonant-initial word.

In addition to the dependent variable (duration of decir), we code each token for factors demonstrated in previous works to constrain durations of words. Each of these factors is described in the following paragraphs:

Target context speaking rate: In this work, we seek to understand whether any durational differences across separate realizations of decir are accounted for through control of discourse contextual factors. With regard to durational differences, it is well established that local speaking rates strongly predict and correlate with target durations. For instance, increased contextual speaking rates contribute to durational shortening of target segments and words (Arnon and Cohen Priva 2013; Brown et al. 2021; File-Muriel and Brown 2011; Fosler-Lussier and Morgan 1999; Gahl 2008; Hualde and Prieto 2014; Nadeu 2014). Contextual speech rate is nearly universally considered in studies of phonological variation due in large part to the strength of its conditioning effect, yet it has not been widely controlled for in studies of discourse makers. For all tokens of decir, we measure the duration of the context preceding the token and that following the token (bounded by pauses or cessation of speech). We then calculate a measure of the speaking rate in lexical syllables per second. The average pre-context speaking rate in syllables per second for [es decir] is 4.93 (SD = 4.30), and for decir tokens it is 8.58 (SD = 3.77). Considering only the fluent tokens (no targets bounded by pauses), the average pre-context rate in syllables per second for [es decir] is 6.97 (SD = 3.31), and for decir it is 8.52 (SD = 3.73). The average post-context rate in syllables per second for [es decir] is 5.67 (SD = 3.64) syllables per second, and for decir it is 5.85 (SD = 5.29). The non-pause adjacent targets have post-context rates in syllables per second of 6.90 (SD = 5.09) for decir and 6.87 (SD = 2.33) for [es decir].

Predictability: Studies of word variation demonstrate that word durations correlate with word predictabilities (Bell et al. 2009; Jurafsky et al. 2001), with increased predictabilities of words and segments correlating with increased phonetic reduction. These authors argue that predictable words can be retrieved quickly and faster lexical access yields speeded productions. Although alternative hypotheses for the predictability effects have been argued (e.g., Smooth Signal Redundancy [Aylett and Turk 2004, 2006], Uniform Information Density [Jaeger 2010]), in multiple frameworks, increased predictability correlates with increased reduction. Thus, for each token of decir, we code for the word that was spoken before the target. We obtain corpus counts from the oral section of Corpus del español (Davies 2002) for our targets and preceding word forms. Then, for each target token, a predictability measure was calculated based upon the previous word (w|w−1).

Onset presence or absence: The target tokens of decir are spoken in a variety of phonological contexts. The onset consonant (the voiced dental stop /d/) prescriptively has allophonic variation ([d], [ð]) contingent upon the preceding context. Word durations correlate with number of phones, with words containing a greater number of segments having greater duration. Each token of decir is coded for whether the initial /d/ was deleted or not. The average rate for /d/ deletion is 32 % for both [es decir] (N = 9) and decir tokens (N = 68).

Distance from phrase boundary: A characteristic traditionally associated with discourse markers is that they constitute a separate prosodic unit. Based upon the behavior of discourse markers in written texts, it is noted that discourse markers are often pause-adjacent, with clear boundaries. Nevertheless, empirical studies on the prosody of discourse markers do not consistently corroborate this observation. As is noted by Wichmann et al. (2010: 47), discourse markers cannot be distinctively associated with a specific prosodic pattern. Studies based on naturally occurring data (e.g., Cabedo Nebot [2013] for Spanish, and Shan [2021] for Chinese) also show that discourse markers are not systematically delimited by pauses. Nevertheless, in order to control for any potential effects stemming from systematic differences in prosodic phrasing, such as phrase-final (Gahl 2008; Lohmann 2018) and pre-boundary (Lohmann and Conwell 2020) lengthening or domain-initial strengthening (Wagner and Watson 2010), we code each target token for its distance from a preceding pause and a following pause in number of lexical syllables. Any potential effect of pause-boundary on target tokens will be diminished with increasing distance in number of syllables. The average number of syllables preceding [es decir] targets is 9.6 (SD = 11.58) and for decir tokens it is 12.4 (SD = 11.0). Only non-pause adjacent tokens are preceded on average by 11.3 (SD = 8.45) syllables for [es decir] and 12.4 for decir (SD = 10.8). The average number of syllables following the target [es decir] is 13.2 (SD = 12.6) and 10.9 (SD = 11.7) for decir targets. The average number of syllables following non-pause adjacent targets is 17 (SD = 11.8) for [es decir] and 12.8 for decir (SD = 11.7).

Target type: Lastly, based upon the meaning of the token (whether used as a reformulator or whether expressing propositional content), each target token was labeled as either belonging to the discourse marker construction [es decir] or not. This factor is coded to test for the possible independent effect of lexical type on the duration of the decir targets.

5 Results

6 Discussion

This work explored the understudied discourse maker in Spanish [es decir] to test whether there was evidence of reduction when used as a reformulator compared to decir as a lexical item with propositional meaning, as has been demonstrated for other discourse markers in previous literature (e.g., Bybee et al. 2016; Schubotz et al. 2015). We controlled for multiple factors of the target tokens’ context, such as speech rate of the preceding context, the following context, and word predictability, in addition to phonetic features of the target itself (onset /d/ realization). We considered whether significant differences in reduction were due to online factors and appreciably different patterns of use in naturally occurring discourse. While controlling these factors that condition durations of words, a significant effect of the word type is evident. Tokens of decir are shorter when used within the reformulator construction [es decir] compared to uses outside of the dyad.

Knowing that the reformulator decir is shorter, can an explanation be found in usage patterns? That is, conditioning factors present in target contexts may accumulate in memory in lexically specific ways. Previous research (e.g., Brown and Raymond 2012; Bybee 2002; Forrest 2017; Raymond and Brown 2012) suggests that a word’s history of use (forms’ ratio of occurrence) in contexts conditioning phonetic reduction predicts variant forms, independently of the factors operating in the production context. As variant forms of words and constructions are experienced in use, episodic traces of these experiences are stored in memory, and may in turn be selected for future use via a Feedback Loop (Kemmer and Barlow 2000: ix).

For instance, increased contextual speaking rates contribute to durational shortening of target segments and words. As context speaking rates increase, target durations decrease. This is an online (contextual) effect that significantly constrains target durations (Arnon and Cohen Priva 2013; Fosler-Lussier and Morgan 1999; Hualde and Prieto 2014; Nadeu 2014), and indeed is significant in these data (Table 3). However, words differ in their likelihood of use in speaker-relative fast (vs. slow) speech rate contexts (Brown and Raymond 2014). Words used more often in fast speech have more opportunity to reduce, and this reduction accumulates in memory (Brown et al. 2021). The durational shortening of words then can reflect both the online contextual speaking rate in addition to the independent effect of the accumulation in memory of words long-term histories of use in these contexts. As such, words’ cumulative exposure to specific discourse contexts is crucial to understanding the variation (Brown and Raymond 2012; Bybee 2002).

We ask whether an explanation for the durational differences between the target types (decir and [es decir]) can be found in a lexicalized accumulation in memory of significantly different patterns of use in contexts conditioning reduction. That is, we consider whether [es decir] is used proportionally more often in phonetic contexts promoting reduction compared to [decir], the effects of which would accumulate in memory giving rise to the shorter duration of decir in [es decir]. The results of this data exploration are summarized in Table 5.

As Table 5 makes evident, the usage patterns for decir and for [es decir] are distinct concerning the average conditioning contexts we consider in this study. We examine the average rate of contexts preceding both target types [decir, es decir], as well as the mean speech rate of contexts following the targets. The average of previous and following speech rates are higher for the propositional decir. Higher speaking rates promote durational shortening of targets (see results Table 5). Thus, if there is an accumulation in memory of words’ histories of use in relatively fast speech (Brown et al. 2021), such an effect is not driving the durational shortening of [es decir] in our data.

Research has shown that adjacency to pauses can affect durations (e.g., Gahl 2008). Pauses both preceding and following target words and phones can have a lengthening effect. Thus, we examine in our data the distribution of the tokens with regard to previous and following pauses and calculate their ratio of occurrence in each context. The proportion of tokens adjacent to both a previous and a following pause (which has the effect of lengthening or slowing articulation rate) is higher for [es decir] than decir. Again, here, the likelihood of pause adjacency does not predict shorter decir realizations in the discourse marker.

Jurafsky et al. (2001) note that words are durationally shorter when used in lexically predictable contexts. Aylett and Turk (2004, 2006 find a direct correlation between predictability and durational shortening, and Seyfarth (2014) demonstrates that words that are commonly predictable (low informativity words) are more prone to durational reduction, even in unpredictable contexts. An average predictability measure of the decir tokens based upon the previous word is summarized in Table 5. The predictability of decir from the previous word (bigram/previous word frequency) is higher (on average) for decir outside of the construction. Higher average predictability would seem to favor durational shortening of the target when not used as the discourse marker, which of course, is not in the result reported for these data.

In each case, importantly, the averages of the conditioning factors summarized in Table 5 favor faster articulations of decir over [es decir]. Thus, shorter durations of decir in the [es decir] construction do not appear to be due to the accumulation of effects stemming from the phonetic conditioning environments examined in this work. Results in Tables 3 and 4 likewise suggest that factors of the online production context known to constrain variation do not account for the different durations of the two types of decir. The shorter durations of [es decir] do not seem to derive, then, from online conditioning or the accumulation in memory of the conditioning factors.

Likewise, lexical frequency effects and reduction due to grammaticalization fail to provide a straightforward explanation for the findings. Tokens of decir are more frequent outside of the reformulator construction. In the corpus-based analysis we conducted, we found 360 tokens of decir, compared to 28 tokens of [es decir]. In the larger, perhaps more representative Corpus de referencia del español actual (Real Academia Española), just 29 % of the tokens of decir (N = 88,438) are found within the [es decir]^[2] construction. Given the reductive effect of lexical frequency, propositional decir would be predicted to be shorter. The difference between decir types also does not appear to have an explanation in lexical frequency.

Similarly, as noted in Section 2.2, the target is mostly used as a reformulator with a textual meaning and presents scarce evidence of having acquired subjective or expressive meanings, which would be associated with highly grammaticalized discourse markers (Traugott and Dasher 2005). The duration of [es decir] does not seem to be due to a high degree of grammaticalization. Thus, a unit need not be especially frequent or grammaticalized in order to be significantly reduced. In other words, frequency and grammaticalization do not always account for durational shortening in naturally occurring discourse.

In an exemplar model of lexical representation (Bybee 2001; Pierrehumbert 2001) fine-grained phonetic detail is stored in memory. The nature of the lexical representations, then, reflects usage patterns and experience. Studies that identify lexically specific patterns of reduction and variation provide evidence in support of a view in which words are stored in memory with vast amounts of detail. The contents stored in memory reflect experience with words in production and perception. These details include memory of phonetic realizations, contexts of use, contextual and phonetic probabilities, as well as detail regarding relevant environmental and extralinguistic factors. The phonetic shapes of words registered in the minds of speakers, which reflect the accumulation of contextual pressures, may then serve as targets for subsequent articulations. Although words are stored in memory, evidence suggests a network of constructions with different degrees of abstraction (Pardal Padín 2020: 502).

From this perspective, then, we would expect that speakers create associations between constructions that present similarities in form and/or function. In this way, [es decir] is expected to be related to other lexical items that fulfill a similar discourse function as a reformulator including o sea ‘I mean’, esto es ‘that is’, and en/con otras palabras ‘in other words’. As is mentioned in previous studies (Martín Zorraquino and Portolés Lázaro 1999: 4,123), o sea ‘I mean’ is the most frequent reformulator in Spanish. As a consequence of high frequency of use, the word pair is stored as a chunk in memory (Bybee 2001). Through frequent co-occurrence in use o sea may be combined in memory into one unit of storage with a single syllable carrying the prominent stress. The concomitant lack of prominence across the combination facilitates phonetic reductive processes occurring naturally in the language. As such, discourse markers may record through this chunking process ‘special reduction’ (Bybee et al. 2016).

Lexical items are closely associated with the constructions they frequently occur in (Bybee and Thompson 2022: 12). This entails that the representations of constructions in memory contain information associated with experiences of use including the lexical items, phonetic detail, interactional functions, and context of use. In the case of [es decir], then, the reduction seems to stem from the reformulation use. That is, speakers are aware of a [reformulator] construction at a higher level of abstraction. Reduction is part of the articulatory plan and/or the lexical representation of possibly both the lexical combination (that is, the specific words themselves), as well as the construction. This predicts that novel reformulators or other less frequent ones such as digo ‘I say’ would exhibit greater reduction than their lexical counterparts. Therefore, lexically related pairs of words (sea compared to [o sea], digo ‘I’m saying’ compared to digo ‘I mean’, decir compared to [es decir]) will differ in degree of reduction not necessarily because of the online conditioning, degree of grammaticalization, or high lexical frequency, but because of the reformulative construction having shortening as part of the lexical representation.

7 Conclusions

This study is the first to report a detailed acoustic analysis of two lexically related forms in Spanish (decir and [es decir]) in naturally occurring discourse to determine whether phonetic differences exist across lexical types. We find significant differences in the duration of decir across word types (Table 2). Factors of the target context independently predict target duration along with word type (Table 3), and in fact, duration is a clear predictor of target type (Table 4). These durational differences do not appear to stem from an accumulation in memory of patterns of the likelihood of use in those conditioning environments (Table 5), high token frequency, or grammaticalization, which have been previously used to explain durational differences across homophonous word pairs. We suggest that [es decir] is stored as a lexical unit that contains as part of the lexical representation shorter word durations relative to the lexical form decir and that this durational shortening is part of the [reformulator] construction.