Voice mismatch and contrast in French Right-Node Raising

Abstract

Right-Node Raising is generally considered to impose stricter identity conditions than other kinds of ellipsis, such as VP ellipsis, according to Hankamer & Sag 1976 and Hardt 1993. In this paper, we investigate voice mismatch in French Right-Node Raising (RNR) through a corpus study and two experiments. We show that RNR with voice mismatch can be found in a written corpus (frTenTen 2012) and that many examples involve coordination of a reflexive active and a short passive form. We suggest this is because semantic contrast (here, between self and external agent) plays a role according to Hartmann (2000) and Abeillé and Mouret (2010). We ran two acceptability judgement experiments to test voice mismatch and semantic contrast. We did not find any penalty for voice mismatch with VP ellipsis but an interaction with semantic contrast. We also found an effect of semantic contrast when coordinating an active and a passive VP without participle ellipsis. We conclude that voice mismatch is acceptable with RNR and propose a Head-driven Phrase-Structure Grammar (HPSG) analysis, following Chaves (2014) and Shiraïshi et al. (2019).

1. Introduction

The possibility of voice mismatch has been a hot topic in the study of VP ellipsis since it undermines the deletion under identity approach, or structural approach to ellipsis (Merchant 2013). On the other hand, it is expected under semantics-based approaches to elliptical constructions (Ginzburg & Sag 2000 a.o.). Other kinds of ellipsis (gapping, sluicing) have been shown to resist voice mismatch, but Right-Node Raising (RNR) has not been considered in this respect. Different analyses have been proposed for RNR (Ross 1967, Kayne 1994, Chaves 2014) but none of them predict the existence of such mismatch.

1.1. Voice mismatch and ellipsis

Voice mismatch has been studied for VP ellipsis and is attested in English (Hardt 1993, Kehler 2000, Kertz 2013). Both active elided VPs with a passive antecedent (Examples (1a, c)) and the reverse (Example (1b)) (a passive elided VP with an active antecedent) are observed.

Merchant (2013) tries to reconcile the possibility of voice mismatch with a deletion under identity approach: he posits a VoiceP node and assumes that VP ellipsis takes place below VoiceP: so in this approach, there is no mismatch. (Figure 1, Figure 2).

Figure 1 Passive structure from Merchant (2013)

Figure 2 Active structure from Merchant (2013)

This analysis makes two predictions: (i) VP ellipsis with voice mismatch should be as acceptable as with match and (ii) voice mismatch is only possible with small (VP) ellipsis but not with large (clausal) ellipsis.

Most of the time, there is a penalty associated with mismatch (Kim & Runner 2018). Different authors have proposed factors that favor or disfavor voice mismatch with VP ellipsis. According to Kehler (2000), asymmetric discourse relations between the source and the target such as Cause-Effect (Example (3a)) make voice mismatches more acceptable than symmetric discourse relations such as Resemblance (Example (3b)).

Kertz (2013) suggests that topic continuity plays an important role in VP ellipsis. In a series of experimental studies, she shows there is a penalty with voice mismatch when the subjects are contrastive topics as in Example (4a). In Example (4b), on the other hand, the subject of the sentence with voice mismatch is not a contrastive topic (the agent is the same as in the source sentence) and the mismatch penalty is reduced.

She also suggests that the penalty for lack of topic continuity is independent of ellipsis.

From a different perspective, Arregui et al. (2006) assume that the grammar of ellipsis requires a syntactically matching antecedent for the elided constituent. Ellipsis is ungrammatical when no syntactically matching antecedent is available, despite the actual occurrence of such examples in naturally occurring speech. In case of ungrammaticality, the parser may however “recycle” a parsed constituent and thus give the illusion of grammaticality. In a series of experiments, they found a greater penalty for a passive elided VP with an active antecedent (Examples (1b), (5b)) than for an active elided VP with a passive antecedent (Examples (1c), (5a)).

The Recycling Hypothesis relies on earlier findings showing that a passive, as a non canonical construction, is more easily misrecalled as an active (Mehler 1963, Thuilier et al. 2020) than the other way around (an active being misrecalled as a passive). A passive-active mismatch may thus create an illusion of grammaticality because the speaker misrecalls the antecedent as active.

Poppels & Kehler (2019) tested the Recycling Hypothesis in a series of experiments where they found a general passive penalty: VP ellipsis with a passive antecedent is judged worse than with an active antecedent, even with voice match.

They also challenge the Recycling Hypothesis by testing cataphoric VP ellipsis: if the ellipsis site precedes the source, there is no antecedent to be misrecalled.

Poppels & Kehler (2019) found a mismatch penalty (Examples (7c), (7d)) and a passive first penalty (Examples (7b), (7d)) but no interaction between the two. This contradicts the Recycling Hypothesis that would predict that the mismatch in the passive first condition would be better since the passive could be misrecalled as active. They conclude that their results suggest a more general Passive Ellipsis Clause Penalty (PECP), accounting for the contrast between Examples (5-a) and (5-b), between Examples (6a) and (6b), and between Examples (7c) and (7d). To see whether PECP applies to other ellipsis, they tested sluicing and gapping in a pilot study: only sluicing led to a small non-significant passive penalty so that they concluded that PECP may be unique to VPE.

If Merchant 2013 is right, “larger” ellipsis should not allow for voice mismatch. Indeed, sluicing Example (8b) and gapping Example (8d) do not exhibit such acceptable mismatches.

But contrary to Merchant (2013)’s predictions, in a corpus study, Miller (2014), as well as Kim & Runner (2022), found attested examples of voice mismatch with pseudogapping (Example (9a)), that should not be possible under a raising remnant analysis (Kuno 1981) (NP, Noun Phrase, PP, Prepositional Phrase; VP, Verb Phrase):

We can conclude that structural identity-based approaches may either accept (Merchant 2013) or reject (Arregui et al. 2006) voice mismatch, and have a hard time predicting which cases of voice mismatch are acceptable and which are not. It also remains to be tested whether voice mismatch is acceptable for other types of ellipsis than VP ellipsis and pseudogapping.

In this paper, we focus on voice mismatch in RNR (Ross 1967, Chaves 2014) for French, which has been shown to accept some syntactic mismatches (Shiraishi et al. 2019).

1.2. RNR and syntactic mismatch

Right-Node Raising is often set apart from other elliptical constructions since the elided clause precedes the full clause.

Several syntactic analyses have been proposed for RNR. It has been analyzed in terms of rightward movement (Ross 1967, Hankamer 1971, Postal 1974, Gazdar 1981, Steedman 1996, Sabbagh 2007, Sabbagh 2014) (Example (10a)), multidominance (McCawley 1988, Goodall 1987, Moltmann 1992, Wilder 1999, Gracanin-Yuksek 2007, Bachrach & Katzir 2009) (Example (10b)) (Figure 3), and deletion under identity (Wexler & Culicover 1980, Kayne 1994, Hartmann 2000, Yatabe 2001, Beavers & Sag 2004, Yatabe 2012, Chaves 2014, Shiraishi et al. 2019) (Example (10c)).

Figure 3 Simplified RNR analysis with multidominance

Syntactic mismatches between elided and peripheral material cast doubt on raising and multidominance theories. Under their most natural implementations, all of these analyses predict that voice mismatch should not be possible.

Chaves (2014) proposes that the missing and peripheral strings must have the same morphophonology and that the latter must be prosodically independent. Shiraishi et al. (2019) propose a revised deletion analysis based on lexeme identity instead of morphosyntactic identity, in order to capture polarity, preposition, and verb form mismatch. Verb form mismatch has been observed by Shiraishi et al. (2019) (Example (11a)) as well as polarity (Example (11b)) (Kayne 1994) or preposition mismatches (Example (11c)) (Abeillé et al. 2016):

Some examples of voice mismatches can also be found in English, but it is not easy to tell RNR apart from cataphoric VP ellipsis.

Other languages such as German may have voice mismatch in RNR too (without having VP ellipsis) (examples from B. Crysmann, pc).

Some examples of voice mismatches can be found for French as well (Abeillé et al. 2016).

In Examples (14) and (15), the missing past participle in the first conjunct corresponds to the passive participle in the second conjunct. In French, as in English, past participle and passive participle are syncretic forms. But unlike English, French does not have VP ellipsis with auxiliaries (Emonds 1987; Abeillé & Godard 2002), be they followed by a past or a passive participle (avoir, être). ²

1.3. RNR and semantic contrast

According to Hartmann (2000) and Ha (2008), RNR needs to have a semantic contrast (contrasting values inside a set of alternatives) between the predicates in the sequences (typically conjuncts) sharing the peripheral sequence (e.g., detests and likes in Example (17a) and has and have not in Example (17b)) (see Rooth 1992 for a formal definition of semantic contrast).

As pointed out by Chaves, following Kentner et al. (2008), this semantic contrast does not always lead to typical contrastive focus intonation, even if it may, as noted by the capital letters in Example (17a). In an English corpus study (Penn Treebank), Bîlbîie (2013) found that most RNR examples involved polarity, modality, aspect, or tense contrast (Examples (18), see also, Example (11a) above with tense contrast).

We call this the “contrast condition on RNR”, and, to the best of our knowledge, it has not been tested experimentally. In the absence of such contrast, syntactically well-formed RNR are less acceptable, as shown by the intuitive difference between Examples (19a) and (19b): ³

The unacceptability of Example (19b) shows that subject contrast is not enough, and in that case, VPE (or stripping) is preferred, as in Example (19c).

For French, Abeillé & Mouret (2010) conducted a study on written (newspaper) and spoken (radio) corpora. They found that most RNR examples displayed polarity, tense, or modality contrast between the two verbal conjuncts, as in the following examples (see also Example (15) above which displays past/present tense contrast.

They also observe a tendency towards topic continuity, with VP coordination (with the same subject) rather than clausal coordination, in naturally occurring examples of RNR, although it is not always observed in the linguistic literature, Example (19a).

Given the lack of empirical data on voice mismatch and semantic contrast in RNR, we investigate French RNR in comparison with previous studies on voice mismatch in VP ellipsis. Among other differences, RNR differs from VP ellipsis since: (i) almost anything can be elided (not just a VP or a predicate phrase), (ii) the ellipsis site precedes the full clause, and (iii) RNR must be intrasentential, unlike VPE. In what follows, we limit ourselves to (VP) RNR in coordinated clauses.

In Section 2 of this paper, we present the results of a corpus study on voice mismatch in French RNR. In Section 3, we report on the results of an acceptability judgement experiment on VP RNR, showing no mismatch penalty but a possible effect of semantic contrast. In Section 4, we present a second experiment which confirms a semantic contrast effect in coordinated clauses independently of VP RNR. In the last section, we sketch a Head-driven Phrase-Structure Grammar (HPSG) analysis based on lexeme identity, following Shiraishi et al. 2019.

2. A corpus study on voice mismatch in French RNR

We first conducted an informal search on the Internet, with conjoined active and passive auxiliaries. Short passives are easier to find than long passives (with a “par” (by) phrase). RNR involving an active verb with an NP complement and a long passive verb as in Example (21) are not very natural (Abeillé et al. 2016).

We mostly found examples of voice mismatch with reflexive active verbs, as in Example (22) from a newspaper article:

We then did a more systematic search on the web corpus frTenTen 2012 (about 10 billion words, Baroni et al. 2009) for sequences of coordinated relative clauses with different voice auxiliaries and for sequences with coordination of different voice auxiliaries. We searched for the following sequences :

qui se sont et/ou qui ont été (“who se(reflexive) are and/or who have been’), qui ont été et/ou qui se sont (“who have been and/or who se(reflexive) are”), qui s’est et/ou qui a été (“who se(reflexive) is and/or who has been”), qui a été et/ou qui s’est (“who has been and/or who se(reflexive) is”), qu’on a et/ou qui a/ont été (”that one has and/or that has/have been”), se sont et/ou ont été (“se(reflexive) are and/or have been”), ont été et/ou se sont (“have been and/or se(reflexive) are”), s’est et/ou a été (“se(reflexive) is and/or has been”), a été et/ou s’est (“has been and/or se(reflexive) is”).

French reflexive verbs come in different varieties (Melis 1990, Barque & Candito 2019). We only kept those which display an argumental reflexive. We thus left aside cases without a non reflexive counterpart, be they intrinsic reflexives (s’évanouir “faint”), mediopassives, or anticausatives, as in Example (23).

The results are provided in Table 1. We do not report the strings which returned no hits.

Table 1 RNR with voice mismatch in frTenTen12 corpus.

We found 27 examples of RNR with voice mismatch in frTenTen 2012. Both active-passive (Example (24a)) and passive-active (Example (24b)) orders are attested but the latter are more frequent (21 out of 27). ⁴ This is difficult to compare with previous findings on voice mismatch with VP ellipsis (Section 1.1): Arregui et al. (2006) found that passive-active mismatch was easier than active-passive mismatch, and proposed a processing explanation, but, in the case of VPE, the full antecedent precedes the elliptical VP, and can be misrecalled, while in RNR, the ellipsis site precedes the peripheral material. As mentioned before, Poppels & Kehler (2019) proposed a more general passive penalty for VP ellipsis, but in our data, the first elided predicate is actually more often the passive one.

All examples that we found involve a reflexive active and a passive. We did not find any examples with a non reflexive active and a passive. Reflexives have been argued to bear some similarities with passives in French. They use the same être (be) auxiliary as passives, and the participle obeys the same (subject) agreement pattern (non reflexive transitive verbs have the avoir (have) auxiliary and the participle does not agree with the subject) (Abeillé et al. 1998, Abeillé & Godard 2002):

From a derivational perspective, some authors treat French se verbs as intransitive unaccusatives, like passives (Wehrli 1986): se absorbs accusative case or se prevents the verb from assigning case to its object, which raises to subject position (see also Grimshaw 1982 for an analysis of Romance reflexives as intransitive verbs, inside a non derivational framework). In this approach, there is less syntactic mismatch between a reflexive active and a passive, since both have an empty object position, than between a transitive active and a passive.

However, this unaccusative approach of passives is controversial. Sells et al. (1987) show that reflexive predicates vary cross-linguistically in the mapping between the lexicon, syntax, and semantics, in particular, as being closed or open predicates. Labelle (2008) argues that the French se verb is an open predicate, since it can be transitive and have another reflexive marker (Example ((26)): she proposes that se does not reduce the argument structure of the verb and that the DP that surfaces in subject position is the external argument. We follow this proposal. For more on our syntactic analysis, which does not rely on syntactic movement, see Section 5.2 below.

From a semantic perspective, in our corpus examples (as in Example (24)), the reflexive and the (agentless) passive verbs obey the semantic contrast condition on RNR seen above (Section 1.3), since one involves a reflexive predicate (self agent) and the other the same predicate with an external (unknown or implicit) agent. In an example of a non reflexive active followed by a passive (Example (27a)), on the other hand, the two conjuncts do not exhibit a simple semantic contrast, since one has an unknown patient and the other an unknown agent.

Furthermore, from a discourse perspective, reflexives may also be argued to share some properties with passives, since in our examples, they also keep topic continuity (same subject) better than non reflexive actives, a condition independently observed by (Kertz 2013) for English VPE. To conclude, we suggest that the tendency found in the corpus for examples with a (intransitive) reflexive active and a passive (and not with a transitive active and a passive) may not come from syntax (assuming a reflexive verb has an empty object position like a passive) but rather from semantics, since reflexives/short passives allow a contrast between self and external agent, and also from discourse since they assure topic continuity.

3. Experiment 1: Testing RNR with voice mismatch

Our corpus data lead us to suppose that RNR with voice mismatch can be acceptable. We conducted two acceptability judgement experiments in order to test both the effect of voice mismatch and the effect of semantic contrast on French RNR. We tested whether voice match and semantic contrast play a role and whether there is an interaction between the two factors. To the best of our knowledge, semantic contrast has not been tested experimentally so far (see Section 1.3 above).

3.1. Materials

The materials for Experiment 1 were inspired by the examples from the corpus study, with two conjoined relative clauses and a shared participle. We constructed 20 target items with four conditions: $ \pm $ match, $ \pm $ contrast (Example (28)). ⁵ All conditions have topic continuity with the same subject. In the mismatch conditions, we always have active-passive order, in order to avoid a penalty predicted by Arregui et al. (2006)’s Recycling Hypothesis and Poppels & Kehler (2019)’s Passive Ellipsis Penalty. In the conditions with semantic contrast, the thematic roles of the antecedent of the relative clauses are contrasted with an agent in the first clause and a patient in the second clause. In the mismatch condition, the agent of the predicate is thus the subject of the (reflexive) active, while it is unspecified in the passive (Example (28a)). In the match condition, the agent is the subject of the reflexive active and the indefinite “on” of the non reflexive active (Example (28c)). In the conditions without contrast, the antecedent is always a patient: the agent is indefinite (on) or unspecified in the mismatch condition (Example (28b)), and is unspecified in the match condition (Example (28d)). Notice that in this last condition (no mismatch, no agent contrast), the auxiliary past tenses are as similar as possible (imperfect and perfect).

20 target items were created, each of them in the four conditions described above. They were distributed across four lists using a Latin Square design, such that each participant saw five items in each condition but never the same item in more than one condition. Target items and 40 unrelated fillers were presented in a random order which was different for each participant. The full set of items can be found in the OSF-project: RNR-Materials (https://osf.io/2pdnj/). Fillers were French garden-path sentences based on materials from (Paape et al. 2018). These fillers had the advantage of spanning a wide range of acceptability (from difficult garden-paths with a mean acceptability of 4.9 to easy non-garden-paths with a mean acceptability of 7.7, on a 0–10 scale).

3.2. Procedure

We conducted the online experiment on IbexFarm (Drummond 2010), using a server at the University Paris Cité for data storage. Before starting the experiment, subjects provided information about their age, native language, and gender and gave their written consent to participate in the experiment. For each experimental trial, participants read a sentence and judged its acceptability on a scale from 0 (not at all acceptable) to 10 (fully acceptable). ⁶ The duration of the experiment was 15 minutes on average. Each experimental session started with three practice items.

48 participants (22 female, mean age: 30 years), recruited on Prolific, with the constraints that they currently live in France and only spoke French when growing up, judged the acceptability of the sentences on a scale of 0 to 10 and answered simple comprehension questions. Two participants were excluded from the analysis for answering less than 80% of the questions correctly. All participants received £2.20 via the Prolific platform. ⁷

3.3. Results

Given that we obtained acceptability judgements from our participants, which do not fulfill requirements for linear mixed models, we analyzed the results with cumulative link mixed models using the clmm function in the ordinal package in R (Christensen 2019). Figures in this paper show means and confidence intervals for convenience since most readers are familiar with this kind of data presentation, although models for data analysis are based on the raw ordinal data resulting from the acceptability judgements. Independent variables were voice mismatch and semantic contrast. They were centered around zero for statistical analyses. Participants and items were included as random variables. Only intercept models could be run due to convergence failure of more complex models (model equation: clmm(judgment = contrast $ \ast $ match $ + $ (1 $ \mid $ participant)+(1 $ \mid $ item))). The results are given in Figure 4. ⁸

Figure 4 (Color online) Results of Experiment 1 on VP RNR

While acceptability judgements across conditions are all very high (7–8 out of 10), we did not find any significant effect of voice mismatch. We did, however, find a significant interaction of Match and Contrast, with Contrast leading to higher acceptability ratings in the mismatch condition. No other effects were significant (interaction effect: $ \beta $ = $ -.75, st. error $ = $ .24,z $ = $ -3.114,p $ < $ .002 $ ).

In the match condition, the contrast condition (Example (28c)) involves two active auxiliaries (qui se sont ou qu’on a), whereas the no contrast condition (Example (28d)) involves two passive auxiliaries (qui étaient ou qui ont été), but there was no significant difference there. In the mismatch condition, the contrast condition (Example (28a)) involves an active reflexive and a passive (qui se sont ou qui ont été), while the no contrast condition (Example (28b)) involves a non reflexive active and a passive (qu’on a ou qui ont été), and we find a significant difference in favor of the former ( $ \beta $ = $ .50, st. error $ = $ .25,z $ = $ 2.034,p $ < $ .05 $ ).

Unlike the voice mismatch penalty observed in previous experiments on English, VP ellipsis (Arregui et al. 2006, Kertz 2013, Poppels & Kehler 2019 a.o.), we did not find a mismatch penalty: the target items with voice mismatch are as acceptable as the items with voice match. As predicted by Hartmann (2000), Ha (2008), we found a penalty for items without semantic contrast but only in the case of voice mismatch. This does, however, not necessarily mean that the contrast effects depend on mismatch (see Experiment 2 below). The lack of an effect in the match condition may also be due to the minimal difference between the two past tense forms in the +match $ - $ contrast condition (Example (28d)), which can be interpreted as aspect contrast between imperfect and perfect.

One may attribute the high acceptability of RNR with voice mismatch to the syncretism between active and passive participles. However, Shiraishi et al. (2019) found that RNR with verb form mismatch without syncretism is highly acceptable in French.

One may also consider the VoiceP hypothesis to play a role for French. If we follow Merchant (2013)’s proposal, there is no syntactic mismatch between active and passive. The structure for active-passive would involve the same VP as passive-passive:

Figure 5 Active structure with VoiceP

Figure 6 Passive structure with VoiceP

What about reflexives? If they are similar to passives (syntactically intransitives, with an object raised to subject position (Wehrli 1986)), our (Example (28a)) condition (reflexive and passive) would be a case of match and our (Example (28c)) condition (reflexive and active) would be a case of mismatch. However, we did not find a significant difference between the two (p $ > $ .40). On the other hand, Labelle (2008) has argued that reflexive is a specific voice in French, with its own VoiceP, different from (non reflexive) active and passive:

Figure 7 Reflexive structure with VoiceP

Under Labelle’s view, there would be voice mismatch also in the (Example (28c)) (reflexive-active) condition. But the reflexive-active condition (with contrast) was rated higher than the match (passive-passive) (Example (28d)) condition (without contrast).

We conclude that the effect of structural identity (with voice mismatch under some analyses, without mismatch under a VoiceP analysis) is not sufficient to explain our data, and that semantic contrast plays a role.

As for semantic contrast (Hartmann 2000, Ha 2008), we found an interaction with voice mismatch in Experiment 1, so that the effect of contrast was significant in the mismatch condition. ⁹ Since we expect this effect to be independent from voice mismatch, we ran a second experiment without voice mismatch.

4. Experiment 2: Contrast effects without voice mismatch

Experiment 2 was designed to test whether semantic contrast plays a role independently of voice mismatch. We therefore used materials similar to Experiment 1 but with the participle repeated and not elided. Most of the time, a final phrase was shared between the two conjuncts (and, hence, could be argued to be elided in the first conjunct). In Example (31), it is the PP à Rome (”to Rome”). We applied the same 2x2 design ( $ \pm $ match, $ \pm $ contrast) as in Experiment 1.

20 target items were created, each in the four conditions described above. Target items and 16 unrelated fillers were distributed across four lists using a Latin Square design and presented in a random order which was different for each participant.

The filler items came from an unrelated experiment, on the attachment of relative clauses to coordinated noun phrases as in Example (32) (Hedier et al. 2021).

4.1. Procedure

The procedure in this experiment was nearly identical to Experiment 1. The only difference was that participants judged sentence acceptability on a scale from 1 (not at all acceptable) to 5 (fully acceptable), since this scale was used by the unrelated experiment which provided the filler items. ¹⁰ The duration of the experiment was 10 minutes on average.

Data from 54 French native speakers (32 female), recruited on the RISC web site (http://www.risc.cnrs.fr/) who volunteered to participate in the online experiment on Ibex (Drummond 2010) were analyzed in this experiment. The results from one participant who was not a native speaker of French were excluded from the analyses. The age of the participants ranged from 18 to 69, with a mean age of 31 years.

4.2. Results

We analyzed the results with cumulative link mixed models mixed effects regression models using the clmm function in the ordinal package in R (Christensen 2019). As before, independent variables were voice mismatch and semantic contrast. They were centered around zero for statistical analyses. Participants and items were included as random variables. Convergence failures made it impossible to include random slopes for the fixed factors. The results are given in Figure 8.

Figure 8 (Color online) Results of Experiment 2 on PP RNR

Even without VP RNR, we could establish a significant effect of semantic contrast with an advantage for the +contrast conditions (contrast: $ \beta $ = $ -0.4744, st. error $ = $ .1367,z $ = $ -3.472,p $ < $ .001 $ ). We also found an advantage for the mismatch condition (match: $ \beta $ = $ 0.3546, st. error $ = $ 0.1312,z $ = $ 2.701,p $ < $ .01 $ ) but no interaction between the two factors.

4.3. Discussion

The +contrast conditions received higher acceptability ratings than the $ - $ contrast conditions even without participle ellipsis. This shows that the contrast effect shown in Experiment 1 on VP RNR is more general. This confirms Hartmann (2000), Ha (2008), and Abeillé & Mouret (2010)’s contrast condition on RNR proposed for English, German, and French. It is also possible that a semantic contrast between two coordinate predicates is an independent requirement of verbal/clausal coordination.

The advantage for the mismatch condition may be the consequence of the fact that the items were based on naturalistic items with mismatch from the corpus study, while match items were adapted from them. It may be possible that factors not included in this study make the mismatch condition more natural. This would be the topic of future studies.

5. An HPSG analysis

Following Chaves (2014), Shiraishi et al. (2019), we show how an HPSG analysis can account for our data.

5.1. Previous analyses of RNR

As noted earlier (see Section 1.2), several analyses have been proposed for RNR. Even though a movement analysis may be reconciled with mismatch, various arguments have been proposed against movement analyses (for overall discussion, see Beavers & Sag 2004, Bachrach & Katzir 2009 a.o.), namely, RNR can target more than one constituent (Abbott 1976), as shown in Example (33a); it can target non-constituents, as shown in Example (33b) below; and it does not obey locality constraints otherwise observed for wh-movement or extraposition (Levine 1985), as shown in Example (33c).

Syntactic mismatches cast doubt on multidominance approaches, pace (Citko 2018). Such approaches are at pains to explain why in case of conflict, it is enough for the shared element to meet the requirements of the final conjunct and to ignore those of the first conjunct. For all these reasons, we do not adopt a movement-based analysis nor a multidominance analysis. This means that we consider Right-Node Raising a misnomer and that Peripheral Ellipsis would be more appropriate. However, we keep the traditional name and acronym here.

5.2. An analysis for French passives and reflexives

In what follows, we rely on previous work on French using HPSG (Pollard & Sag 1994, Müller et al. 2021). Following Bresnan (1982), passives are analyzed in HPSG via a valence changing lexical rule (Flickinger 1987, Müller 2001). Assuming a verb has its syntactic arguments in an argument structure list (ARG-ST), the first one corresponding to the subject, the passive lexical rule gives rise to a passive lexeme with a different argument structure, while keeping the semantics (and the thematic roles) the same. In French, the by-phrase can be introduced by par or de, depending on the verb agentivity (Koenig 1999, Zribi-Hertz & Abeillé 2021), and it is analyzed as a complement (see also DaCunha & Abeillé 2021, Angelopoulos et al. 2021).

Transitive verbs select the perfect auxiliary avoir, via an aux feature (Abeillé & Godard 2002). Reflexive verbs are derived through another lexical rule that changes the auxiliary selection to être and types their complement as a non-canonical anaphoric affix (Abeillé et al. 1998), following the affixal analysis of French pronominal clitics (Miller 1992, Miller & A. Sag. 1997).

In the perfect tense, the reflexive affix is realized on the auxiliary as any other clitic.

In order to account for clitic climbing (Example (37)), Abeillé & Godard (1994, 2002) analyze French auxiliaries as complex predicates: they analyze the perfect auxiliaries as taking the past participle as a complement, and inheriting its arguments as subject and complements. ¹¹ They also analyze the passive construction as a complex predicate, with the copula taking the passive participle as its complement, since it allows for clitic climbing as well (Example (38)).

5.3. An HPSG analysis for RNR

Following Shiraishi et al. (2019), we assume that RNR does not require phonological identity (Example (11a)) but lexeme identity (unless pun or zeugma). The following example is unfelicitous because it involves two different homophonous lexemes: the baseball bat (elided) and the animal bat (peripheral):

To account for Example (11b), they consider that positive and negative polarity sensitive determiner (any/some) share the same lexeme, and introduce the Lexical Identifier (LID) feature to capture lexeme identity.

We assume the same LID value for passive and past participles which are derived from the same lexeme (Sadler & Spencer 2001, Bonami & Webelhuth 2013). The feature structure in Example (40) illustrates a simplified lexical entry for the past participle mobilisé. The feature structure in Example (41) gives a simplified lexical entry for the past participle with reflexive mobilisé. The feature structure in Example (42) shows a simplified lexical entry for the passive participle mobilisé.

Following Chaves (2014), we formalize a RNR rule (43) which targets Morphophonology (MP) and does not add further syntactic constraints. It basically says that in a construction with two sublists with identical lexemes, and usually identical phonology, the first sublist can be deleted, leaving the meaning unchanged. We leave for future work the addition of a Contrast relation in the semantics (inside synsem), since we conjecture it could be more general and associated with clausal coordination. ¹²

The MP feature of the Daughter phrase is divided into five sublists, which must obey prosodic constraints, which we ignore here. The left-hand sublists, $ {\mathrm{L}}_1 $ and $ {\mathrm{L}}_2 $ (respectively, in black and in grey in (43)), correspond to any non final conjunct, and the right-hand sublists, $ {\mathrm{R}}_1 $ (in blue), $ {\mathrm{R}}_2 $ (in red), and $ {\mathrm{R}}_3 $ (in black), to the final conjunct. The first sublist $ {\mathrm{L}}_1 $ is kept. The deleted sublist $ {\mathrm{L}}_2 $ (in grey) must comprise elements with the same LID as $ {\mathrm{R}}_2 $ (the shared peripheral element(s), in red), and by default (noted /) the same phonology. Note that the elements in $ {\mathrm{L}}_2 $ are not preserved in the MP feature of the mother phrase (resulting in their not being pronounced). Thus, the form of the peripheral elements $ {\mathrm{R}}_2 $ (in red) is always that required by the final conjunct. The $ {\mathrm{R}}_1 $ (in blue) is the sublist before the shared elements and may comprise a coordinating conjunction. The extra $ {\mathrm{R}}_3 $ sublist accounts for Right-Node Wrapping ¹³ and can be empty.

Figure 9 provides the analysis for RNR with voice mismatch.

Figure 9 Analysis of French RNR with voice mismatch

6. Conclusions

In this paper, we have investigated voice mismatch in French Right-Node Raising through a corpus study and two acceptability judgement experiments, and provided an HPSG formalization. RNR is often thought to be special because it is backwards and less syntactically constrained than other kinds of ellipsis (Chaves 2014). We show that RNR allows for voice mismatch like VP ellipsis and pseudogapping. We found that RNR with voice mismatch is attested in written French. In our corpus study, most RNR with voice mismatch involve coordination of reflexive active and passive forms. This shows the importance of semantic contrast (here, between self and non self agent) in RNR, and also of topic continuity as in VP ellipsis (Kertz 2013). The first acceptability judgement experiment revealed that RNR with voice mismatch is as acceptable as without mismatch, and that (agent) contrast may play a role, in particular, for voice mismatch. The importance of semantic contrast is confirmed by the second experiment, without VP RNR.

This may be difficult to account for by a raising or multidominance analysis of RNR. We show that it is compatible with a lexeme deletion-based approach. Following Chaves (2014) and Shiraishi et al. (2019), we present an HPSG deletion-based analysis of RNR based on lexeme identity.