2.1 Deaccentuation as a marker of givenness

In intonation languages such as English, German and Icelandic, stressed syllables of content words are often produced with a pitch event on or near the stressed syllable. This is called pitch accent, which leads to the perception of intonational prominence of that stressed syllable (Pierrehumbert 1980; ’t Hart, Collier & Cohen 1990). Pitch accent can be acoustically measured, exhibiting cues such as a specific tonal event (accent type), longer syllable or segment duration of the accented syllable, larger F0 difference between the maximum F0 and the minimum F0 (i.e., widened pitch register) in the area of the pitch accent, and also louder speech (Lehiste 1970, also see Grice et al. 2020 for English and its varieties). In broad-focus contexts (in which all the items in an utterance are new and not contrasted), each pitch accent in an intonational phrase (i.e., a clause-size prosodic unit) exhibits a downstep relation in English, German, and Icelandic, such that each pitch accent is realized with a pitch height that is relatively lower than the preceding one. This is schematically represented in the hypothetical example in (7), in which each stressed syllable (marked with x) is accompanied with a pitch peak, which is progressively lower towards the end of the utterance:

The lack of accentuation on a stressed syllable in a prosodic position that is accented as a default is considered as deaccentuation. Deaccentuation leads to the weakening of suprasegmental cues on an item in such a way that there is another prosodic unit that is perceived as more prominent and hence accented. Prosodic prominence is a relational notion, in which a portion of speech is produced/perceived as prosodically more prominent than the other parts (Ladd 1978; 1980; 2008; Gussenhoven 1983; Selkirk 1995; and see Ladd & Arvaniti 2023 for an overview). In this sense, givenness related deaccentuation is often considered as part of the strategy for marking focus related prominence.

In languages such as English and German, deaccentuation-related low level flat intonation (often acoustically manifesting itself as lower mean F0 and/or less difference between maximal and minimal F0 values) is often observed in the post-nuclear/post-focal area (which is also referred to as post-focal pitch compression) (Ladd 1978; 1980; 1996; Xu 2011; Xu and Xu 2005). Linguistically given items that linearly precede the semantic focus in an utterance do not systematically exhibit deaccentuation. In the pre-focally given area accent realization and prosodic phrasing is usually retained as if the sentence is uttered in broad focus context (Baumann et al. 2021; Baumann & Grice 2006; Féry & Kügler; 2008; Féry & Ishihara 2009). Although not as common as post-focal deaccentuation, pre-focal intonational reduction is possible (also known as pretonic accent Ladd 1980). In pre-focal deaccentuation, the pitch range difference is as small as post-focal deaccentuation, but unlike post-focal deaccentuation, the pitch level (i.e. the overall mean F0 (fundamental frequency) of the area under discussion) is much higher (leading to a high plateau in the intonational contour). The hypothetical stylised intonational contours in (8) illustrate the visualisation of the intonational contours in the pre-focal area (8b) and post-focal area (8a), in which both areas exhibit deaccentuation, hence downstep across the intonational phrase is not retained.

For intonation languages, the consensus is that focused items often exhibit a nuclear pitch accent, and givenness is often cued with loss of prosodic prominence and deaccentuation (Chafe 1976; Ladd 1978; 1980; 2008; Gussenhoven 1983; Terken & Hirschberg 1994; Selkirk 1995; Schwarzschild 1999; Wagner 2005; Baumann 2006; Cole et al. 2010; Baumann & Kügler 2015). This idea has been recently challenged by acoustic research. Recent findings based on acoustic analysis of data gathered through experiments on intonation languages such as English and German shows that deaccentuation in the post-focal area is not systematically present (Kügler & Féry 2017; Wagner & McAuliffe 2019; Kügler & Calhoun 2020 and references therein). This is surprising especially for the post-focal area, which has been thought to be deaccented as a grammatical means to mark focus (such that prosodic dephrasing/deaccenting in this area lends more prominence to the focused item that immediately precedes it). A degree of accentuation has been experimentally shown to be still present post-focally, even though it is much more diminished than its non-given (broad focus) counterpart. This has been illustrated by Kügler & Féry (2017) for German, based on experimental evidence, as shown in Figure 1.

Figure 1: Downstep of the compressed pitch accents in the post-focal/post-nuclear region, small dots represent post-focal accents on the given area. (Kügler & Féry 2017: 283, Fig.10).

The finding that post-focal givenness does not always lead to a complete flattening of the intonational contour in this area,⁶ coupled with the fact that given items in the pre-focal area do not show any systematic deaccentuation, pose a challenge to any account that assumes a procedural and causal link between ellipsis and deaccentuation. One may therefore conclude that deaccentuation and ellipsis are not inter-linked and the SRD accounts should be abandoned all together. However, such an approach would then miss the long-observed discourse relational dependencies that seem to converge in marking both ellipsis and deaccentuation in a non-trivially similar manner. In fact, even in cases such as German, in which accentuation is not completely lost, there are some cues of givenness in the form of compressed pitch range. Therefore, the observed post-focal prominence is at best “weak prominence” (Kügler & Féry 2017: 283). This can be inspected in the visualisation provided in Figure 1, in which the area after the nuclear pitch accent (all the steps of the downstep that come after the accent marked with a large point) exhibits pitch peaks with a much narrower register, and the overall accentuation is markedly compressed. In other words, although one may not speak of deaccentuation as an intonational cue for givenness in intonation languages, prosodic attenuation (pitch compression and other diminished cues such as shorter duration, smaller F0 register etc.) may still be considered as a correlate of information structural givenness. Therefore, one may still maintain a weakened version of the RD approach, (WRD as schematised in (6b)), in which the subject matter prosodic correlate is not complete loss of accentuation, but the presence of diminished prosodic cues, and lack of nuclear prominence in the post-focal area (see Ott & Struckmeier 2016 for a similar discussion).

WRD would allow us to make a more crosslinguistic generalisation, in which languages that lack pitch accent placement in their grammars, and non-intonation languages that do not mark information structural notions based on accent-related events (such as phrase languages, or most tone languages) may as well be included, as long as they systematically exhibit a form of prosodic attenuation in especially post-focally given area.⁷

What would still be problematic even for WRD is a language that does not exhibit any givenness related prosodic attenuation at all. If such a systematically non-deaccenting language, in which no instance of deaccentuation of information structurally given items is observed, productively exhibits ellipsis, this would then mean that ellipsis is possible in a language without prosodic marking of givenness. Such a language would require us to completely de-link ellipsis and prosodic attenuation in marking givenness.

Without empirical support from experimental data, the validity of any RD-based hypothesis, which is initially based on impressionistic observations or native speaker intuition, remains questionable. To fill this gap in the literature, we carried out a production experiment to investigate the prosodic profiles of pre-elliptic clauses in Icelandic.

We specifically investigated Icelandic, because although it is an intonation language, Icelandic has been suggested to be one such language, in which given items are observed not to be systematically deaccented and are produced in the same way as their non-given counterparts (Nolan & Jónsdóttir 2001; Dehé 2009 and see Section 3 for a more detailed description of the Icelandic intonation). We carried out a production experiment on Icelandic to test whether there is any acoustic or intonational cue that systematically marks givenness. If we find that speakers of Icelandic do not prosodically distinguish utterances in broad focus and narrow focus contexts (of pre-elliptic clauses), then our results will provide strong evidence against WRD, simply because Icelandic productively allows clausal ellipsis (Wood et al. 2016; 2020). If we find prosodic attenuation in the post-focally given parts in pre-elliptic clauses, then a WRD account can still be maintained.

The following section provides an overview of the previous literature’s findings about the lack of deaccentuation in Icelandic and motivates the need for a systematic investigation of pre-elliptic clauses, which leads to the currently reported case study. In sections 3.1 and 3.2, we present the details of the study and methodology respectively. Section 3.3 reports the results of our case study.

3.1 The study

We conducted a prosodic production experiment to determine whether Icelandic is indeed a ‘non-deaccenting’ language. We asked native speakers of Icelandic to articulate a variety of pre-elliptic utterances and we then measured the prosodic properties of three lexical items (i.e., words) in the post-focal domain of each test item (see (11) for an abstract representation of a test condition, using English). To determine whether clauses that are eligible for ellipsis in Icelandic systematically show signs of givenness-related prosodic attenuation, we compared the prosodic properties of these target words when they linearly follow an information-structurally new item (who in (11)) and are therefore given (the post-new condition) to when they comprise part of the broad focus domain and are therefore not given (the all-new condition). Thus, post-new and all-new comprise the levels of the predictor variable info-structure. The prosodic properties we measured (i.e., the outcome variables) were duration, mean fundamental frequency (F0), and the difference between the maximum and minimum F0 points on each target word (F0 register).

Adopting the default assumption that, like other intonation languages, Icelandic displays givenness-related prosodic attenuation, we expected to obtain the following results:

In addition to info-structure, our experiment contained the factor ellipsis-type. This factor has two levels, pre-fragment and pre-sluice.¹⁰ Test items in the pre-fragment condition would produce a fragment response if clausal ellipsis is semantically licensed and applies (see (12)), whereas test items in the pre-sluice condition would produce a sluice if clausal ellipsis is semantically licensed and applies (see (11)).¹¹

We included ellipsis-type as a factor because Güneş & Lipták (2021) suggest that, in English, Hungarian and Dutch, the wh-remnant in sluicing (i.e., who in …but I don’t know who) exhibits a default nuclear accent, as opposed to a focus-related pitch accent. If this is also true for Icelandic, then the post-wh-area in our pre-sluice test items in the post-new condition might not exhibit complete deaccentuation, as the syntactically-defined nuclear accent position may shift when other items follow the wh-word. In contrast to sluicing remnants, the remnants of fragmentary responses are known to be focused, and therefore bear a focus-related pitch accent (see Weir 2014 and the references therein). Thus, we expect to observe:

The experiment had a 2×2 within-item, within-subject design. Sample items for each condition are presented in (13):

The experiment employed 4 lexical sets. Almost all target words (i.e., the boldfaced words in (13)) were minimally trisyllabic and all exhibited initial stress with no lexically-specified long vowels. Lexical words, as opposed to function words were chosen to ensure the prosodic freedom and lack of potential prosodic effects of subcategorization frames that function words may possibly have. Function words are also often prosodically weak (unless information structurally contrasted) and mostly do not inherently exhibit lexical stress which is essentially required in prosodic word formation (Selkirk 1996).

We restricted the number of target words to three in each token. This is because in the post-new case if there were to exist accentuation, and resulting downstep, three steps is ideal to observe the transitions in the downstep. In two-word condition or one-word condition, one cannot reliably attribute the observed prosodic properties to downstep, hence to accentuation. For instance, in the single post-new word scenario (which is similar to Dehé’s (2009) data) the prosodic properties (e.g., longer duration on the stressed vowel) on the single post-focal item may be attributed to the utterance final intonational phrase boundary which is known to have a lengthening effect on the final word. All in all, a longer linguistic string after the narrowly focused item is ideal to assess the presence of any prosodic event, or the lack thereof in the post-focal area – also see Kügler & Féry (2017) for similar motivations for including multiple post-focal phrases when investigating deaccentuation in German. We avoided including more than three lexical items in the post-focal area simply because longer sentences would run the risk of leading to production errors or exhibiting uncontrolled prosodic behaviour due to production planning of elongated utterances (see e.g., Krivokapić (2012) for the role of utterance length in prosodic production planning).

The syntax remained constant across lexical sets. The items in (13) present the first lexical set—the remaining three sets are exemplified below:

The experiment included 3 repetitions of each test item, meaning 48 test items in total. The test items were accompanied by 150 fillers (50 unique filler items, 3 repetitions of each), yielding a 1:3 ratio of test to filler items.

The speech production experiment included the recordings of 10 native Icelandic speakers (6 females), ranging in age from 20 to 56 at the time of the experiment (October 2018). Participation was voluntary and unpaid. Recording occurred in a quiet studio at the Reykjavik University in Iceland, using a professional table-top microphone and Audacity 2.2.2. The average total duration of each recording session was 25 minutes. All participants were uninformed about the study’s aim. The first author was present at each recording session. Stimuli were presented to participants as slides in a presentation software (one item per slide) on a laptop screen. Participants were permitted to repeat items and controlled the transition between slides. The participants were instructed to read the content of each slide content silently and to familiarise themselves with the utterance before producing that slide’s target item. Furthermore, they were instructed to speak as naturally as possible, with a normal speech rate. Before the experiment round, the participants completed a practice round (2 practice tokens). Stimuli were presented to each participant in a different pseudo-randomized order, with no consecutive test items.

Audacity 2.2.2 was used to extract each target utterance as a .wav file. Each sound file was parsed into syllables, words, and sentences using PRAAT (Boersma 2001) by a trained native speaker of Icelandic. 30% of all parsed tokens were second checked by the first author for the accuracy of the interval boundary placement for words and syllables, with no inaccuracies found during this check. Data extraction was performed using ProsodyPro (Xu 2013). 480 observations (each of three T-Words) entered the statistical analysis. For impressionistic analysis, time-normalized mean F0 line plots of test utterances were generated for each lexicalization. Hypothesis testing statistical analysis was carried out by fitting linear mixed-effect regression models (LMEMs) to the data, using the lmertest package in R (Kuznetsova et al. 2017). All LMEMs display a maximal fixed effect structure and the best fitting random effect structure, as determined manually through maximum likelihood testing (lmerTest’s anova function). All LMEMs have anova-coded predictor variables. The model estimates, t values (generated by Satterthwaite’s method), and p values reported below were generated using lmerTest’s summary function.

3.2 Results

3.2.1 Duration

Figure 3 presents the mean durations of each T-word in all conditions and lexicalizations.

A summary of the best fitting LMEM for the duration data is presented in Table 1. In this model, the mean durations of each individual T-Word are averaged over, as we are not interested here in variation between the T-Words themselves. The model summary shows that there was a main effect of info-structure, but no main effect of ellipsis-type and no interaction effect. This confirms our prediction that our target words in pre-elliptic utterances exhibit a shorter duration when given as opposed to broad focused (recall [P1]) and strongly suggests that, more generally, Icelandic employs shortened duration to prosodically mark givenness.¹²

Figure 3: Mean duration (in milliseconds) for each target word in each lexicalization.

Table 1: Summary of interaction model (duration).

Formula = duration ~ ellipsis-type * info-structure + (1 | speaker)
Contrasts	Estimate (β)	SE	t	p
all-new – post-new	25.01	5.48	4.59	<0.01
prefrag – presluice	–6.70	5.48	–1.28	0.22
(all-new_prefrag – post-new_prefrag) –(all-new_presluice – post-new_presluice)	5.80	10.91	0.53	0.60

The absence of a main effect for ellipsis-type or an interaction effect shows that the duration of our T-Words is the same regardless of whether they follow a prominent wh-word or a prominent non-wh DP. Therefore, [P4] is not borne out for duration.

3.2.2 Mean F0

Figure 4 presents the mean F0 for each target word in all conditions across four lexicalizations.

Figure 4: F0 mean (in semitones) for each target word in each lexicalization.

A summary of the best fitting LMEM for the mean F0 data is presented in Table 2. In this model, the mean F0 of each individual T-Word is again averaged over. The model summary shows that there was a main effect of info-structure, but no main effect of ellipsis-type and no interaction effect. This confirms our prediction that our target words in pre-elliptic utterances exhibit a lower overall F0 when given as opposed to broad focused (recall [P2]) and strongly suggests that, more generally, Icelandic employs lower overall F0 level to mark givenness.

The absence of a main effect for ellipsis-type or an interaction effect shows that the mean F0 of our T-Words is the same regardless of whether they follow a prominent wh-word or a prominent non-wh DP. Therefore, [P4] is not borne out for mean F0.

Table 2: Summary of interaction model (meanF0).¹³

Formula = duration ~ ellipsis-type * info-structure + (1 | speaker)
Contrasts	Estimate (β)	SE	t	p
all-new – post-new	–2.24	0.21	–10.49	<0.01
prefrag – presluice	–0.37	0.21	–1.72	0.09
(all-new_prefrag – post-new_prefrag) – (all-new_presluice – post-new_presluice)	0.17	0.30	0.39	0.70

3.2.3 F0 register

Figure 5 presents the mean F0 register for each target word in all conditions and lexicalizations.

Figure 5: mean F0 register (in semitones) for each target word in each lexicalization.

A summary of the best fitting LMEM for the F0 register data is presented in Table 3. In this model, the F0 register of each individual T-Word is again averaged over. The model summary shows that there was no main effect of info-structure and of ellipsis-type and no interaction effect. This does not confirm our prediction that our target words in pre-elliptic utterances exhibit a smaller F0 register when given as opposed to broad focused (recall [P3]) and suggests that, more generally, Icelandic does not employ smaller register as means to prosodically mark givenness. The absence of an effect between all-new and post-new condition indicates that the post-new condition exhibited an accent on each target word.

The absence of a main effect for ellipsis-type or an interaction effect shows that the register of our T-Words is the same regardless of whether they follow a prominent wh-word or a prominent non-wh DP. Therefore, [P4] is not borne out for register, we observed no pitch compression.

Table 3: Summary of interaction model (F0register).

Formula = F0register ~ ellipsis-type * info-structure + (1 | speaker)
Contrasts	Estimate (β)	SE	t	p
all-new – post-new	0.41	0.30	1.35	0.18
pre-sluice – pre-frag	0.51	0.30	1.72	0.09
(all-new_pre-sluice – all-new_pre-frag) – (post-new_pre-sluice – post-new_pre-frag)	0.44	0.60	0.74	0.45

One gets the impression from Figure 5 is that the difference in F0 register between the all-new and post-new conditions is greater for T-Word1 than for T-Word2 and 3. Consequently, it might be the case that a significant difference obtains between the levels of infostr for F0 register in T-Word1. To determine this, we conducted post-hoc pairwise comparisons of LMEM’s estimated marginal means (grouped by T-Word). This indeed revealed a significant effect of modulating infostr on F0 register for T-Word1 but not T-Word2 and 3; see Table 4.

Table 4: Pairwise comparisons of the estimated marginal means for F0 register.

Word position	contrast	t	p
T-Word1	all-new – post-new	1.35	0.04
T-Word2	all-new – post-new	0.79	0.10
T-Word3	all-new – post-new	–0.53	0.28

We offer an explanation for this disparity in mean F0 register differences between T-Word1 and T-Word2 and 3 in the discussion section below.

3.2.4 Time-normalized average F0-line charts

The Figures 6 and 7 below present time-normalized mean F0 line-plots for each lexicalization. Each plot averages over speakers and test item repetitions. From these plots one obtains an impression of the contour properties—e.g. possible downstep, flattening, prosodic chunking, or upstep behaviour—for our test items in the all-new and post-new conditions.

Figure 6: Time-normalized F0-line charts for the entire tokens [Focused Subject (the left-most word)+pre-elliptic domain] in pre-fragment condition, in neutral/all new (black line) and post-focal/post-new (red line) context.

The impressionistic observations that we gather from the inspections of the F0-line charts are in line with the hypothesis testing results we obtained for F0. In particular, when we compare the contour trends in all-new condition (in black) to post-focal condition (in red), we see that both conditions exhibit a downstep of pitch accents, and that the high points of each step of downstep is a bit lower in the post-focal condition. Similarly, the lowest point of each step in the downstep is also lower. This explains why we found no significant difference in the F0 register. As such, in the post-focal condition, pitch accents are fully realized but on a lower level overall, when compared to the all-new condition. We observe no pitch compression, but an overall low-level pitch in post-focal area.

To summarize the results, pitch accents on the lexical item in the given area are realized, yet they are also diminished, exhibiting lower mean overall F0, and shorter duration, when compared to their non-given counterparts.

Figure 7: Time-normalized F0-line charts for the entire tokens [Wh-word hver + pre-elliptic domain] in pre-sluice condition, in neutral/all new (black line) and post-focal/post-new (red line) context.

3.2.5 Focused item: mean F0 and Max F0 comparisons

Inspection of time-normalized F0 line charts show us that focused item in pre-ellipsis clauses, and their all-new context counterpart also show a visually inspectable difference, such that the potential remnants (i.e., narrowly focused items) in the pre-ellipsis condition seem to be produced with an overall higher F0. Although our hypothesis does not make any predictions about the focus-prosodic properties in pre-elliptic clauses, to report an additional observation about the acoustic properties of narrow focus in Icelandic, we ran post-hoc statistical analyses, in which we aimed to inspect the prosodic properties of only the focused word in pre-elliptic clauses and its counterparts in all-new context.

In the pre-fragment condition the potentially focused items are the sentence initial subjects (i.e., Þorvaldur, Hrafnhildur, Bergmundur, and Hansína), which are all tri-syllabic (in which the initially accented syllables are closed syllables). In the pre-sluice condition, the potential target of narrow focus is always the mono-syllabic wh-word, hver. Since the pitch accent is realized initially on the leftmost stressed syllable in our data, then in the pre-sluice condition the only existing syllable (i.e., that of the wh-word hver) carries the pitch accent on this word. However, in the pre-fragment case after the initially realized pitch accent, there are two more syllables within the word that leaves room for a post-accentual fall, which we expect to be steeper in post-new condition. Independent of the ellipsis properties, the two ellipsis-related conditions then are also distinct in terms of the number of syllables in their remnants. We envisage that the syllable number has an effect on mean F0 calculations, such that the post-accent area in the tri-syllabic words will involve a stretch of unaccented syllables exhibiting post-accent fall. We do not expect to see this fall in mono-syllabic words. Therefore, in our analysis we kept the ellipsis condition as a variable, however keeping in mind that it is not the ellipsis but the syllable number that is the relevant type of property that distinguishes pre-fragment and pre-sluice conditions for our post-hoc analysis. In the following we report our findings.

Figure 8 presents the mean F0 and Maximum F0 for only the focused word in all conditions averaged over all lexicalizations.

Figure 8: mean F0 (left) and maximum F0 (right) for only the (potentially) focused word in each lexicalization.

A summary of the best fitting LMEM for the mean F0 data is presented in Table 5 below. The model summary shows that there was a main effect of info-structure and ellipsis-type, but no interaction effect between the two factors. This shows that the focused item in pre-elliptic utterances exhibits a significantly different mean F0 than their counterparts in all-new condition. The plot on the left panel in Figure 8 shows that narrowly focused words are produced with higher F0 than their counterparts in broad focus condition. The plot on the left panel in Figure 8 also shows that the monosyllabic wh-item is produced with a higher overall F0 (both in broad focus and in narrow focus condition) than the multi-syllabic words. Absence of the interaction effect between information structure and ellipsis type shows that these are two completely independent variables.

Table 5: Summary of interaction model for only the focused word (F0mean).

Formula = F0mean ~ ellipsis-type * info-structure + (1 | lex)
Contrasts	Estimate (β)	SE	t	p
all-new – post-new	0.55	0.17	3.17	<0.01
pre-sluice – pre-frag	0.68	0.17	3.91	<0.01
(broad focus_pre-sluice – broad focus_pre-frag) – (narrow foc_pre-sluice – narrow foc_pre-frag)	0.26	0.34	0.75	0.45

A summary of the best fitting LMEM for the maximum F0 data is presented in Table 6 below. The model summary (similar to the mean F0 results for the focused item) shows that there was a main effect of info-structure and ellipsis-type, but no interaction effect between the two factors. This shows similar results to mean F0 findings.

Table 6: Summary of interaction model for only the focused word (F0max).

Formula = F0max ~ ellipsis-type * info-structure + (1 | lex)
Contrasts	Estimate (β)	SE	t	p
all-new – post-new	1.02	0.20	4.88	<0.01
pre-sluice – pre-frag	0.81	0.20	3.90	<0.01
(broad focus_pre-sluice – broad focus_pre-frag) – (narrow foc_pre-sluice – narrow foc_pre-frag)	0.57	0.41	1.36	0.17

The plot on the right panel in Figure 8 shows that narrowly focused words are produced with higher F0 than their counterparts in broad focus condition. The maximum F0 plot Figure 8 also shows that the monosyllabic wh-item is produced with a lower maximum F0 (both in broad focus and in narrow focus condition) than the multi-syllabic words in pre-fragment condition. We link this to the syllable number of the items in pre-sluice vs. pre-fragment condition. That is, the overall high levelled mean F0 in pre-sluice condition is linked to the fact the inspected wh-item (hver) is monosyllabic carrying the high accent, H*, (focal or not) on this word therefore exhibits no stretches F0 fall that one usually observed in post-H* area. On the other hand, since in pre-fragment condition the inspected words all involved three syllables the first one of which is accented, the fall that is observed after the initial H* caused the observed difference in the means in both elliptic conditions, such that pre-sluice data ends up having a higher F0 mean. This contrast clearly changes its direction when we compare maximum F0 values. Here we see that the maximum F0 of the H* in pre-fragment condition is much higher than the maximum F0 on hver of the pre-sluice condition. However, in both ellipsis conditions, there is a significant F0 boosting effect (both in terms of maximum F0 and in terms of mean F0) of narrow focus.

The significantly higher mean and maximum F0 values in narrow focus condition (when compared to broad focus condition) may be an additional prosodic cue that the speakers employ to distinguish non-all-new context from all-new context.

4.1 Prosodic properties of post-focal givenness in Icelandic

Results show that, although Icelandic does not exhibit canonical deaccentuation (partially confirming previous literature), it exhibits givenness-related prosodic attenuation (with shorter duration and lower overall F0 level) on post-focally given items both in post-sluice condition and in post-fragment condition, when they are compared to their counterparts in broad focus contexts. In the post-new condition (in the area that we predict properties related to deaccentuation), we found no significant difference between pre-sluice condition and pre-fragment condition. Therefore, our prediction 4 is not confirmed, such that the prosodic profile of the areas that follow the wh-item in the pre-sluice condition and the focused word in the pre-fragment condition did not differ. The F0 register measurements showed that pitch accents are realized. Therefore, our Prediction 3 is not confirmed, supporting the previous observations about the lack of deaccentuation of linguistically given items in Icelandic. Although, givenness does not seem to be linked to the lack of accents in Icelandic, we found that the target words exhibited diminished acoustic values when they are linguistically given. As such, when compared to their broad-focus counterparts, linguistically given target words in the post-focal position exhibited shorter duration and lower overall F0. Therefore, our Prediction 1 and Prediction 2 are confirmed. The inspection of the time-normalized F0 line charts clearly shows the presence of post-focal pitch-accents, which are produced with a lower level F0 than the pitch accents in the all-new condition.

In most cases in Figures 6 and 7, for the all-new context, we observe that the downstep seems to start with T-Word1 (i.e. the post-verbal Object in the SVO order). Compared to the Subject’s F0 properties for this context, we also observe that there is a slight upstep after the Subject’s peak, and on the Object’s peak (in which either the Object’s peak is as high as the Subject’s peak, or they have equal height) (see Figure 9 below for the schematized representation of the all-new contour). In the pre-elliptic context, intonational peak on the Subject is followed by a downstepped peak on the Object, which then progresses into the following two lexical items that follow the Object creating a continual downstep pattern throughout the entire intonational phrase. Given that an upstep on the Object is only observed in all-new context (on the black lined contours in Figures 6 and 7) we relate this to the default prosodic phrasing in which the Subject is prosodically parsed as an independent prosodic domain and the Verb and the items that follow it are parsed as another prosodic domain as in: (S)(VO X X) (see Figure 9 below). Given that verbs do not receive a pitch accent in all-new contexts (c.f., Gussenhoven 1992 among many others) in Germanic, the higher pitch peak on the Object can be explained in terms of the start of a new prosodic domain. This bi-phrased prosodic structure, and the intonational contour of the all-new condition is schematized in Figure 9 below, in which the two large dots represent the first pitch accents of each phonological phrase. Small dots represent the downstepped pitch accents and valleys in between H tones.

Figure 9: Observed intonational contour and prosodic phrasing of all-new condition. High pitch accent on the Subject and Object.

When we compare our findings to Kügler & Féry’s (2017) observations on post-focal accents in German, we see a difference in intonation between German and Icelandic. While the German post-focal accents are produced with a compressed pitch register, we found no such compression in Icelandic, but an overall lowering in the F0 level. Compare the schematic representations of German post-focal accents (in Figure 1) and Icelandic (based on the findings of our current study) in Figure 10 below:

Figure 10: Low-level pitch accents in the post-focal/post-nuclear region in Icelandic and single phonological phrase formation. Small dots represent post-focal accents on the given area, the large dot represents the nuclear/focal pitch accent.

Due to the presence of pitch accents on the linguistically given post-focal items, our results confirm the findings of Nolan & Jónsdóttir (2001) and Dehé (2009). Our findings differ from Nolan & Jónsdóttir’s (2001) observation, in which post-focal accent was found to be as high as the focal accent that precedes the post-focal item. Our findings, confirming Dehé’s (2009) observation regarding downstep, show that accents on the words that are post-focal and linguistically given exhibit a downstep relation with the nuclear accent and other accents in the post-focal area.

Another prosodic difference between the all-new and post-new condition is prosodic constituency. We observe that in all-new condition the sentence initial Subject is separated from the rest of the utterance (see the representation in Figure 9), whereas, in the post-new case we observe that the linguistically given area is at some level parsed together with the sentence initially focused Subject (see Figure 10). In the all-new case in Figure 8, the Subject and Object are parsed into separate phonological phrases, and the phonological phrase that contains the Object includes recursive phonological phrases, such that each embedded phonological phrase represents a step of downstep within the second maximal phonological phrase. In the post-new case in Figure 9, we observe that Subject and the Object (together with all the other lexical items in the utterance) are parsed into a single maximal phonological phrase, and each embedded phonological phrase flanks a lexical item (this is the string of the verb and the Object in the case of the second embedded phonological phrase). The observed downstep observed across the entire utterance can be explained via a downstep across the embedded phonological phrases, with the Intonational Phrase as the domain of downstep, confirming earlier observations since Árnason (1998).

The prosodic grouping of the given items together with the focused item may be interpreted as a symptom of prosodically weaker and dependent nature of information structurally given constituents, which is also in line with the weakened durational and mean-F0 cues we observed.

4.2 Motivating post-focal accents in Icelandic: syntax, and phonology

It is clear in the F0 line charts that the post-focal accents in the post-new condition on T-W1, T-W2 and T-W3 (represented with red lines) are not “phrase accents” (i.e., the tonal event on the transitional area from final, nuclear pitch accent to the right edge intonational phrase boundary tone) (see a discussion on the post-focal pitch accents vs. phrase accents in Kügler & Féry 2017: 283). Given that focus related pitch accent is often identified as nuclear accent (which is often defined as the right-most pitch accent within an intonational phrase in languages with right-headed intonational phrases), the post-focal accents that we observe in Icelandic (similar German (e.g., Kügler & Féry’s 2017), or Swedish (e.g., Myrberg & Riad 2015)), challenge our current understanding of “nuclear accents/intonational phrase headedness”. How can one motivate the presence of post-nuclear accents in intonationally right-headed languages? In this section we will discuss two possible avenues in motivating the presence of post-nuclear/post-focal accents in Icelandic. These are (i) reminiscent accents due to prosodic phrasing linked to syntactic phrasing, and (ii) rhythmic beat-based accent insertion as a phonological requirement. We leave the discussion open however noting both hypotheses are plausible sources of keeping post-focal accents in Icelandic (and any language that has similar behaviour for that matter).

1. Post-focal accents due to syntax-based prosodic constituency: In line with the discussions in Wagner & McAuliffee (2019), and Kügler & Féry (2017), the post-focal accentuation may be linked to the conclusion that “prosodic phrasing created by syntax is not affected by information structure” (Kügler & Féry 2017: 283). Such that, at the level of syntax-based prosodic constituency, all target words (and other weaker words e.g., the function words that surround lexical items) may be parsed into individual prosodic constituents (even in the post-focal area). It seems that information structural import of an item does not fully override its prosodic properties retained form syntax-based chunking. We may interpret the phonetic reduction of the cues (i.e., the overall F0 level and duration) on the post-focal area as the result of the non-final focus and hence nuclear accent. Therefore, a prosodic weakening in the area that follows focal/nuclear accent is still present.

2. Post-focal accents due to a rhythm-based phonological requirement: The possible import of eurythmic organisational factors in phonology may also be considered as an alternative, if not an additional factor to post-focal constituency idea. In particular, post-focal accentuation may be present solely as a result of a strong phonological requirement (in this case, a rhythm based accent boosting requirement) to maintain rhythmic grouping, even when the items are given, in such a way that the word-level prominence of post-focally given items is boosted to be present at the phrase level. Since the post-focal troughs are low due to the post-focal attenuation, then the pitch peaks are also relatively low. Similar rhythm-based accentuation of linguistically given material in the pre-nuclear/focal area has been previously documented (e.g., Horne 1990). Specifically for Icelandic, previous research highlights the role of rhythmic grouping in various prosodic levels. Árnason (2011: 309) maintains that the fact that Icelandic major word classes are inherently left-strong along with other “more or less regular principles combine to define what we may call the normal scansion of a sentence”.

   Perhaps, in Icelandic, then, strong beats at foot structure are not deleted at a higher level in the prosodic hierarchy if associated with a non-focus item, but they are just weakened, preserving the lexically based “normal scansion of a sentence” (i.e. lexically determined positions of strong beats), while at the same time reflecting focus or information structure (similar to tone languages in such a way that those may be faithful to the shape of lexical tones at phrasal/sentence level, but that information status of items may still be globally marked by f0 such that the contour is flattened or realized at a lower f0 level; or expanded pitch range for focus).

4.3 RD and its place in the ellipsis literature in the light of the findings

In terms of the question whether Icelandic challenges WRD, our results indirectly favour this hypothesis. This is because, although the domain of potential clausal ellipsis in Icelandic does not exhibit “low-flat intonation”, it exhibits prosodic attenuation (manifested with shorter duration and overall lowering of the F0 level).

The fact that deaccentuation (or attenuation) is particularly visible in the post-focal area, also known as post-focal dephrasing or deaccentuation, is naturally linked with a prominent account of the syntactic derivation of ellipsis, i.e. the move-and-delete theory of ellipsis (Jayaseelan 1990; Merchant 2001; 2004; Gengel 2008; Toosarvandani 2008; Boone 2014; Weir 2015; see Lipták & Güneş 2022 for an overview). According to this theory, items in an elliptic clause move higher in the hierarchical structure of their clause (often to the left periphery in the CP domain). When ellipsis applies (to VP or TP), the items (including the foci) that are outside of the domains of ellipsis are left pronounced and become the remnants of ellipsis. In such accounts, the area that is elided is always structurally lower than the focused remnant (and to the right of it after linearization), e.g., in TP ellipsis the remnant is assumed to be somewhere higher in the CP domain. Therefore, this area is the post-focal area in the pre-elliptic paraphrases, e.g., in TP ellipsis the focused item, being in CP will be linearized before the contents of the TP, which is given. Move-and-delete approaches therefore predict that the domain of ellipsis is the syntactic constituent that would be linearized in the post-focal area if pronounced. This is schematized in (15):

Since post-focal deaccentuation is more prominent in intonation languages, with a view restricted to this particular typological class of languages, one may conclude that both SRD and WRD can successfully predict a correspondence between the domain of ellipsis and the domain of post-focal deaccentuation.

Given the wider range of data, in which domains other than TP or VP may be subject to ellipsis (non-constituent ellipsis e.g., see Bruening 2015), and the fact that not all remnants can be moved to the left periphery in the pre-elliptic clauses (e.g., Ott & Struckmeier 2016) scholars also posited alternative ways of predicting ellipsis domains. One of the more prominent alternative approaches is the in-situ approach to ellipsis, i.e., scattered deletion approaches. In-situ ellipsis approaches pursue the idea that remnants remain in their base-position (or non-ellipsis related moved position) and do not go through any ellipsis-related movement. If focused items remain in situ, and if these items are the only remnants in the elliptic clause, then any item that would be linearized on either side of the focus marked items in the pre-elliptic clause would be deleted in its elliptic version. This type of approach makes certain predictions that conflict with the predictions of the move-and-delete approaches. Such that ellipsis does not necessarily target a single syntactic constituent and domains that are hierarchically higher than the remnant in the syntactic structure may be deleted. This is schematized in (16) in which the ellipsis domain (ellipsis dom.) is scattered around the focused remnants:

When both stronger and weaker versions of RD are considered, WRD would be better suited even in languages such as German and English, simply because the areas that would precede the foci in pre-elliptic clauses are not necessarily deaccented in these languages. For this reason, WRD accounts would be better suited to the premises of scattered deletion accounts, if they opt for a phonological reduction strategy.

The literature that links deaccentuation and givenness is built on the speaker’s behaviour. Based on production experiments, depending on whether there are prosodic properties of deaccentuation on the items that are (linguistically) given, the scholars decide whether a language systematically utilises deaccentuation to mark givenness. Such conclusions often disregard the availability of optionality in using deaccentuation. It may well be the case that, for instance, in Icelandic, givenness related deaccentuation (with low-flat intonation) is available to the speakers, yet they often produce linguistically given items with a degree of accentuation. Optional availability of post-focal (de)accentuation has been reported for a number of genetically related languages such as German (Kügler & Féry 2017) and Swedish (Myrberg & Riad 2015) The optional availability of deaccentuation effects the way we can restrict the RD hypothesis. As such, one may state that the availability of deaccentuation of a given domain is the key condition on the prosodic licensing of ellipsis, such that if a domain is deaccentable, then it may be elidable. For the case of Icelandic elliptic clauses, then, one may speculate that although in the majority of the pre-elliptic clauses speakers do not deaccent givenness, all elliptic clauses are systematically deaccented, exhibiting low-level flat intonation at some point in derivation (simply because it is available in their grammar). This way of reformulating RD is, however, not free of issues. First of all, since one cannot measure the prosodic properties of the ellipsis domain, the argument that this area is always produced with deaccentuation (unlike their pre-elliptic counterparts) is unfalsifiable. Additionally, this idea is based on the assumption that complete deaccentuation is available in the prosodic grammar of Icelandic as a (seldomly used) marker of givenness, yet this has not been tested or attested before.

A perception study that compares the acceptability of low-flat intonation and lowered accentuation (as reported in this study) as markers of givenness is necessary as a next step. One way of carrying out such a study would involve an offline perception and decision task, in which participants are asked to decide what question (a broad focus triggering question vs. narrow focus triggering question) is suitable for the audially presented answer tokens. Stimuli would involve the recordings elicited for two conditions in the current experiment as well as recordings in which the post-new accents on the post-new condition are destroyed, and the post-new area is completely flattened (the latter can be artificially generated as a resynthesized speech stimuli or produced naturally by a trained native speaker). As a next step, participants would be asked to decide on the naturalness of those tokens for the given context which can be done via forced choice tasks (in which accented and deaccented post-new recordings may be contrasted directly) as well as Likert Scale assessment task which would also allow to observe gradience in judgments.