1.1 Two structural analyses of P-omission under sluicing

Merchant (2001) derives sluices from verbal counterparts and P-omission under sluicing from P-stranding in regular wh-questions. Thus English, which has two options for wh-questions, allows for sluices with (3a) and without (3b) a preposition.

He claims that non-P-stranding languages ban P-omission under sluicing, which he calls the Preposition Stranding Generalization. However, many non-P-stranding languages have been shown to allow for it, including Indonesian (Fortin 2007; Sato 2011), Brazilian Portuguese (Almeida & Yoshida 2007), Spanish (Rodrigues et al. 2009), Polish (Sag & Nykiel 2011; Nykiel 2013), Russian (Philippova 2014), Bulgarian (Abels 2017), Greek and German (Molimpakis 2019), Mauritian creole (Abeillé & Hassamal 2019) and Saudi Arabic (Alshaalan & Abels 2020).

Following Szczegielniak (2008) and van Craenenbroeck (2004), Rodrigues et al. suggest that, in non-P-stranding languages, the possibility of P-omission in sluices comes from the deletion of a cleft structure, with the wh-remnant functioning as the pivot of the cleft, as in Spanish (4).

Since clefts are used to focus an element in a set, they suggest that P-omission improves when the remnant belongs to such a set, as is the case with a nominal correlate (una chica ‘a girl’) and an anaphoric remnant (cuál ‘which’) (see (4)), compared to a pronominal correlate (alguien ‘someone’) and a non-anaphoric remnant (quién ‘who’) as in (5).

1.2 Alternative analyses of P-omission under sluicing

Under direct interpretation approaches to ellipsis (Ginzburg & Sag 2000; Culicover & Jackendoff 2005; Sag & Nykiel 2011), an interrogative sluice is not derived from a full interrogative counterpart: a wh-fragment recovers its propositional meaning from an antecedent available in the context, and it inherits its syntactic and semantic properties from a contextual correlate.

This is why, when the correlate is an animate NP, the remnant must be an animate NP, not a PP (6).

If the antecedent clause includes a PP, two correlates are in fact available; the whole PP or the NP inside it (7), hence two possible corresponding remnants (an NP or a PP), and the possibility of P-omission (7b).

Under such theories, any language may allow P-omission in sluices, and non-syntactic factors may play a role. Nykiel (2013), working on Polish, a non-P-stranding language, found that P-omission is acceptable in sluices, while the cleft source proposed by Szczegielniak (2008) was as degraded as ungrammatical controls. She also found that P-omission is favored by nominal correlates and remnants (8a), while pronominal correlates (8b) favor PP remnants.

Based on these results, Nykiel suggested a processing-based theory. Interpreting a sluice comes with a processing cost since an antecedent has to be found. As in anaphora processing, some antecedents are more accessible than others, and some cues are better than others (Lewis & Vasishth 2005; Martin & McElree 2011; Parker 2019). Nominal correlates are more accessible than pronominal ones due to their richer semantic and morphosyntactic content (Hofmeister et al. 2007).

Prepositions also act as cues to retrieve the PP correlate (preventing who to be misinterpreted as referring to Paul in (7b), for instance), hence P-omission in sluices comes with a processing cost (Nykiel & Hawkins 2020). Philippova (2014) found that preposition length favors P-omission in Russian, longer prepositions offering a better cue for ellipsis resolution, and Nykiel & Kim (2022b) found a similar effect for Polish.

According to Nykiel & Hawkins (2020)’s theory, two general principles are at play: Minimize Forms (MiF) (Hawkins 2004), favoring ellipsis and P-omission, and Minimize Domains (MiD), which may disfavor P-omission. According to MiF, forms can be reduced (e.g., via ellipsis) when the necessary information can be easily retrieved from the context, with full noun phrases being more accessible than pronouns. For sluices, MiF predicts that P-omission is more likely to occur with nominal correlates, which are easily accessible, and also with nominal remnants, which are better cues than pronominal ones.

According to MiD, on the other hand, processing efficiency increases when semantically and syntactically dependent elements are processed in the smallest possible processing domain. In a PP sluice, P is typically processed within the PP domain when P is independent from the verb as in (9a). However, if the P is dependent on the verb as in (9b), it is interpreted in the VP domain. This domain extension creates an additional processing cost for sluices as P in the remnant must go back to re-establish its dependency relation with the V for interpretation. Therefore, MiD predicts that V-dependent Ps are more easily omitted in a sluice.

In such a theory, the acceptability of P-omission under sluicing is flexible, and favored by contentful remnants and correlates, and also by more dependent prepositions.

1.3 The case of French

As noted by Dagnac (2018: 8), “French sluicing has been poorly studied so far”. According to Merchant (2001), P-omission is not possible in French sluices (10a), and (10b) is the only option. According to him, this is explained by the impossibility to strand the preposition as shown in (10c).

Admittedly, Merchant only asked three French speakers (linguist colleagues), who also consider P-omission as more acceptable with the preposition avec ‘with’ as in (11).

One may thus hypothesize that the badness of (10a) comes from à being a weak non-stand-alone preposition, like de ‘of’ (Abeillé & Kupfermann 2021), while avec can stand alone as shown in example (12).

Rodrigues et al. (2009) suggest that P-omission comes from a cleft source in French, as in Spanish. They assign a question mark to (13a) and consider P-omission to be ameliorated with a nominal correlate (une des filles ‘one of the girls’), which defines a set, and an anaphoric remnant (laquelle ‘which.f.sg’) (13b).

They also consider that French clefts have a “weak exhaustivity requirement”, like Brazilian Portuguese and unlike Spanish (Destruel 2012), so that P-omission should be possible with an additive sluice (14):

They thus suggest that French sluices come in two varieties: with a cleft source when P-omission is involved, and with regular wh-fronting in other cases¹. They do not provide examples of P-omission with other prepositions than avec.

Dagnac (2018) finds both (10a) and (13a) ‘perfectly acceptable’, and Gotowski (2022) reports that “some speakers from France find (10a) perfectly acceptable” (p.13). Following Abeillé & Hassamal (2019), they both challenge the cleft analysis for quoi ‘what’ sluices, since this pronoun cannot be fronted (Obenauer 1976; Abeillé & Godard 2021):

As noted by Obenauer (1976) and Abeillé & Godard (2021), it is not possible to extract bare quoi, which must be in situ, unlike the weak form que ‘what’, which must be extracted as in Qu’est-ce? ‘What is it?’ (Abeillé & Godard 2021).²

Dagnac (2018) provides an example of P-omission with quoi (17a), and so do Abeillé & Hassamal (2019) (17b).

While Gotowski does not explicitly consider such cases, her movement analysis (+PF insertion of quoi) implies that P-omission should not be possible in a quoi sluice.

In a corpus of contemporary texts (Frantext after 1980), Smirnova & Abeillé (2021) found two instances of P-omission with qui ça (‘who that’) and de (‘of’), which is a weak preposition like à (18). However, these examples are reprise questions (i.e. clarification requests) with a discourse particle ça ‘that’, which may be more flexible than genuine information seeking questions.

Direct interpretation approaches (Ginzburg & Sag 2000; Culicover & Jackendoff 2005; Sag & Nykiel 2011) predict that P-omission should be possible in French sluices. Nykiel’s cognitive hypothesis, along with the cue-based retrieval approach, also makes predictions about French. A remnant with a preposition carries more information about the correlate than without, and a nominal remnant carries more information than a pronominal one. Hence, for ellipsis resolution, a preposition is less required with a nominal remnant and more so with a pronominal remnant. French sluices can thus be thought of as following the information hierarchy in Figure 1, with more information on the right favoring P-omission: a nominal correlate (e.g. a friend) is more informative than a pronominal one (e.g. someone) (hence more P-omission), itself more informative than an implicit correlate. Quel (‘which’) + Noun being more contentful than lequel (‘which.m.sg’), itself more contentful than qui/quoi (‘who/what’), would also trigger more P-omission.

Figure 1: Information Hierarchy for French sluices.

Similarly, longer and more contentful prepositions may offer a better cue for ellipsis resolution and favor P-omission, compared to monosyllabic ones like à ‘to’ and de ‘of’ (Nykiel & Kim 2022b). On the other hand, taking into account Verb-Prep dependency, Nykiel & Hawkins’s (2020) MiD-based theory predicts that more dependent, less informative prepositions (like à ‘to’ and de ‘of’) have more chance to be omitted than more independent, more contentful ones, such as avec ‘with’ or pour ‘for’.

In this paper, we present two experiments testing the acceptability of P-omission in French interrogative sluices and a (written) corpus study, to examine whether P-omission naturally occurs in French, and which factors have an effect. We also look at the plausibility of a cleft-source explanation for our results. Subsequently, we offer a direct interpretation analysis of French interrogative sluices, based on Ginzburg & Sag (2000), Sag & Nykiel (2011), Ginzburg (2012), Ginzburg & Miller (2018) and Nykiel & Kim (2022a).

2.1 Experiment 1: qui sluices

The first experiment tests French interrrogatives with the pronoun qui (‘who’). 40 native speakers of French were recruited via Prolific and were paid 1.8GP for a 15-minute experiment. Among the information we collected from the participants, the only selection criterion was being an adult (monolingual) native speaker from France.

2.2 Results of Experiment 1

The results in Figure 2³ indicate that the –Prep –Ellipsis condition elicited the lowest acceptability ratings (mean=2.07), confirming its ungrammaticality (e.g. *Qui as-tu parlé? ‘*Who did you talk?’). In contrast, the –Prep+Ellipsis condition received a much higher rating (mean=3.12). The +Prep conditions do not show any difference between –Ellipsis (mean=3.88) and +Ellipsis (mean=3.86) conditions.

Figure 2: Experiment 1: Acceptability results: Remnant qui.

The data were analyzed with a Bayesian ordinal cumulative mixed-effects model with a probit link function (Carpenter et al. 2017) using the brms package (Bürkner 2017) for Bayesian regression models in R⁴. The factors Preposition and Ellipsis were dummy-coded, with +Prep and –Ellipsis serving as reference levels, respectively. Participants and items were included as random factors. Eight sampling chains were run, each consisting of 9000 iterations with a warm-up period of 4500 iterations. We report the parameters’ expected values from the posterior distribution and their corresponding 95% credible intervals (CrIs) for the mean values and differences between them.⁵ We adopted the default priors provided by the brms package for the analysis. The analysis showed a strong evidence for an effect of Prep ($\widehat{\beta }$ = –2.53, 95% CrI = [–3.12, –1.98], P($\widehat{\beta }$ < 0) = 1), suggesting an overall preference for +Prep, and a weak evidence for an overall preference for –Ellipsis ($\widehat{\beta }$ = –0.24, 95% CrI = [–0.73, 0.25], P($\widehat{\beta }$ < 0) = 0.83). It also presents strong evidence in support of an interaction of Preposition and Ellipsis ($\widehat{\beta }$ = 1.67, 95% CrI = [1.14, 2.22], P($\widehat{\beta }$ > 0) = 1), where P-omission is more acceptable under +Ellipsis, compared to –Ellipsis items. See Appendix 1 for the model.

We also looked at possible differences between the ratings of items with à ‘to’ and items with de ‘of’. Results (see Figure 3) show that in the –Prep+Ellipsis condition, the omission of de was rated higher (Mean=3.39) than that of à (Mean=2.85).

Figure 3: Experiment 1: Acceptability results: à vs. de.

The results of the Bayesian ordinal model (with à as the reference level) show an estimated probability of 0.56 with a 95% CrI of [–0.40,0.46]. This suggests no evidence for a difference between à and de. The model, on the other hand, showed a weak evidence for a preference for de under the +Ellipsis condition, compared to à ($\widehat{\beta }$ = 0.27, 95% CrI = [–0.15, 0.71], P($\widehat{\beta }$ > 0) = 0.89), suggesting that the omission of de is more acceptable than the omission of à in sluices.⁶ This is not predicted by any theory claiming that P-omission is ungrammatical in French sluicing.

2.3 Experiment 2: quel+ noun

In order to test Rodrigues et al. (2009) and Nykiel (2017)’s hypotheses, we conducted a second experiment with a nominal remnant. The design was the same as the first one. We recruited 48 French native speakers using Prolific and each participant received a compensation of 1.8GP.⁷

2.4 Results of Experiment 2

The results (Figure 4) show that the –Prep–Ellipsis condition received the lowest ratings (mean=2.16). Unlike in Experiment 1, the ratings of +Prep+Ellipsis (mean=4.72) are now higher than those of +Prep–Ellipsis (mean=4.55) but the difference is quite small. Additionally, the –Prep+Ellipsis sentences now have acceptability ratings (mean=4.47) comparable to those of the +Prep–Ellipsis sentences.

Figure 4: Experiment 2: Acceptability results: Remnant quel + noun.

As in Experiment 1, the data were fitted to a Bayesian ordinal cumulative mixed-effects model with a probit link function, with the same independent variables and the same random factors. +Prep and –Ellipsis were used as reference levels. Eight sampling chains were run, each consisting of 9000 iterations with a warm-up period of 4500 iterations. Results show a strong evidence for an overall preference of the +Prep condition ($\widehat{\beta }$ = –2.83, 95% CrI = [–3.44, –2.24], P($\widehat{\beta }$ < 0) = 1), as in Experiment 1. They also show that +Ellipsis is preferred ($\widehat{\beta }$ = 0.89, 95% CrI = [0.39, 1.46], P($\widehat{\beta }$ > 0) = 0.99), unlike in Experiment 1. We also have strong evidence in support of an interaction, where P-omission is more acceptable in +Ellipsis than in –Ellipsis items ($\widehat{\beta }$ = 2.35, 95% CrI = [1.57, 3.15], P($\widehat{\beta }$ > 0) = 1). The estimate of the interaction in Experiment 2 ($\widehat{\beta }$ =2.35) is higher than that in Experiment 1 ($\widehat{\beta }$ =1.67), indicating a stronger effect of the interaction in Experiment 2 (See Appendix 2 for the model).

While it is not easy to compare two experiments, with different participants and different fillers, P-omission under ellipsis seems to be more acceptable in Experiment 2 (with a nominal remnant) than in Experiment 1 (with a pronominal remnant).⁹

We compared the ratings of the items with à and the items with de in this experiment (see Figure 5). The data were fitted to a Bayesian model and, similarly to experiment 1, results did not show an evidence for a difference between the two prepositions ($\widehat{\beta }$ = 0.15, 95% CrI = [–0.51, 0.81], P($\widehat{\beta }$ > 0) = 0.69), and no interaction with Ellipsis, unlike experiment 1.

Figure 5: Experiment 2: Acceptability results: à vs. de.

2.5 Discussion of the experimental part

The main result of both experiments is that French does permit the omission of prepositions à and de under sluicing, albeit not to the same extent as when the preposition is included. The findings also highlight the distinction between sluicing and verbal interrogatives, as they show mandatory P-inclusion in the latter, in contrast to its optionality in the former. They thus contradict Merchant’s generalization. Furthermore, the acceptability of P-omission in elliptical sentences was influenced by the informativeness of the remnant, as seen in the higher acceptability rate found with nominal (quel ‘which’ + noun) remnants, compared to pronominal (qui ‘who’) remnants. Our results thus support Nykiel’s hypothesis that P-omission in sluicing is affected by processing-based factors.

The overall preference for Ellipsis was stronger in Experiment 2 than in Experiment 1, especially in the +Prep condition. We consider it to be due to the informativeness of the nominal remnant in Experiment 2 (see Nykiel 2017 for English).¹⁰

3.1 Corpus extraction

We extracted data from the contemporary sub-corpus of the Frantext database, which comprises over 1001 texts published after 1980 for approximately 62.1 million words.¹¹ We extracted sluices with pronominal remnants (auquel, duquel, lequel, qui, quoi, respectively ‘to which’, ‘of which’, ‘which.m.sg’, ‘who’, ‘what’), using two search queries: “interrogative pronoun + ?”, for matrix sluices, and “verb + (adv) + interrogative pronoun + punctuation”, for embedded sluices. For sluices with nominal remnants, we used the queries: “quel (‘which’) + noun + ?”, for matrix sluices, and “verb + (adv) + quel + noun + punctuation”, for embedded sluices. We filtered out the occurrences without PP antecedents, and the ones that were not sluices (e.g., n’importe qui/quoi/(le)quel ‘no matter who/what/which’ + noun as indefinite NPs). We obtained a data set of 250 sluices, as shown in Table (5).

3.2 Corpus annotation

We manually annotated the following factors:

• the preposition of the correlate and of the remnant (if any)

• the type of preposition (weak/strong, stand-alone/not-stand-alone) (Merchant 2001)

• predicate-preposition dependency (Nykiel & Hawkins 2020)

• whether the correlate is implicit, nominal, or pronominal (Nykiel 2017)

• whether the sluice is direct or reprise (Ginzburg & Sag 2000)

• whether the remnant is pronominal or nominal and whether it is anaphoric or not (Nykiel 2017, Rodrigues et al. 2009)

• whether the sluice is root, coordinate, or embedded

Table 6 summarizes the sluices that we found in the Frantext subcorpus, distinguishing whether they are PPs or NPs, and whether they appeared as matrix (root or coordinate) or embedded phrases. Overall, we found more P-omission (62%) than P-inclusion (38%).

The following are examples of sluices with PP correlates and NP remnants from the corpus. Notice that both (25a) and (25b) involve the omission of the weak preposition à ‘to’, and that the sluice in (25a) means à quelle amie (dois-je) demander? (‘which friend (should I) ask?’), with modal enrichment (Anand et al. 2021).

3.3 Corpus analysis

We consider the effect of the correlate and remnant, the effect of the preposition and verb-preposition dependency, the effect of matrix/embedding, then the effect of direct/reprise use¹².

3.4 General corpus model

We fit our corpus data in a Generalized Linear Mixed Model (GLMM) (McCulloch & Neuhaus 2005). We used the glmer function from the lme4 package in R, specifying a binomial family, and using the Nelder_Mead optimizer within the glmerControl function. In our best model, the fixed factors are P-omission (reference level P-inclusion) and two predictors: pronominal correlates versus nominal correlates (the latter being the reference level¹⁸), and anaphoric remnants versus non-anaphoric remnants (the latter being the reference level). Author names were included as a random factor. The model in Table 21 highlights that, among the predictors, only the anaphoric remnant and its interaction with pronominal correlates are significant¹⁹. Anaphoric remnants favor P-omission and disfavor it with pronominal correlates. We note, however, that only two examples with these criteria were found in the dataset, cf. Table 14 (section 3.3.3).

The model achieved approximately 98.24% accuracy, 98.68% precision and 99.33% recall, outperforming simpler models that included only correlate or remnant as predictors, which showed significant effects for each factor. Models with more predictors (embedding, stand-alone P, direct/reprise, etc.) did not show other significant effects. Attempts to compute the interactions between all the variables resulted in convergence issues.

3.5 Discussion of the corpus study

Our corpus study shows that P-omission in French interrogative sluices is highly preferred. Among the factors we tested, only correlate and remnant types had significant effects. As predicted by Rodrigues et al. (2009), anaphoric (lequel) remnants favored P-omission, and so did nominal (quel+noun) remnants, as predicted by Nykiel.

The dispreference for pronominal correlates and non-anaphoric remnants that we found may be a general property of discourse cohesion and anaphora resolution. Pronominal correlates have weaker memory representations compared to nominal ones, making them harder to retrieve and necessitating more complex anaphors as cues (Hofmeister et al. 2007, Parker 2019). Sluices with pronominal correlates (quelqu’un) and non-anaphoric remnants (Qui?) may increase processing difficulty, which can explain their lower frequency in our corpus, independently of P-omission, as MiF predicts (Nykiel & Hawkins 2020).

Contrary to Merchant (2001)’s contrast for French, the type of preposition did not have an effect in our data.

In Table 22, we compare our results with those from Nykiel (2017), on (spoken) English (Corpus of Contemporary American English, Switchboard and Santa Barbara corpus) and Polish (National Corpus of Polish).

Nykiel’s hypothesis that in a non-P-stranding language like Polish, P-omission is possible but not preferred, unlike in a P-stranding language like English, is not confirmed for French, which shows a higher P-omission rate than English.

Modality may play a role since her corpora were spoken while ours is written. The corpus samples may play a role as well. In Polish and English, she extracted sluices with bare pronominal (who, what) and nominal remnants (which/what/whose+noun) but not which or which one, which would parallel our lequel sluices. Leaving aside lequel remnants, we found a 72.7% P-omission rate in our French data, which is still higher than what she found for English.

4.1 Issues for a P-stranding analysis

Our experiments show that P-omission is acceptable in French sluices with à and de, unlike in full interrogatives. Our corpus study shows that P-omission is possible with various prepositions and is preferred to P-inclusion, despite French being a non-P-stranding language. This contradicts Merchant’s hypothesis that P-omission would be restricted to prepositions like avec ‘with’ or other stand-alone prepositions (with a null pronoun), such as pour ‘for’ and contre ‘against’. In our corpus data, the P-omission rate with de ‘of’, à ‘to’ and en ‘in’ (87.4%), was not significantly lower than with strong prepositions (89.8%) or stand-alone ones (90.5%).

Since Merchant’s illusory contrast between à and avec was based on the intuition of three linguist colleagues, we may also highlight the importance of quantitative data and controlled experiments in linguistics (Gibson & Fedorenko 2013).

Given our corpus data, we are in a better position to understand why Merchant’s informants may have been mislead: the example he used was with a pronominal correlate (quelqu’un) and a bare wh-remnant (the non-anaphoric pronominal qui), which we found to be the least frequent in our corpus study. Only 8/170 of the analyzed data were with pronominal correlates and pronominal remnants, out of which only one example is with P-omission (42). This also highlights the failure of the one-example-one-speaker intuitionist method, and the importance of taking into account naturally occurring data.

French thus adds to the list of non-P-stranding languages, allowing for P-omission under sluicing.²¹

4.2 Issues for a cleft analysis

Rodrigues et al. (2009) suggest that in non-P-stranding languages, on top of a simple derivation with a PP remnant, an alternative derivation with a cleft is available for NP remnants. Given the more complex derivation for sluices with P-omission, they would expect that they would be less frequent than prepositional ones, with a PP correlate, contrary to our corpus findings. We also found examples of P-omission with the strong pronoun quoi ‘what’, which are not compatible with a cleft source, as noted by Dagnac (2018), Abeillé and Hassamal (2019) and Gotowski (2022).²²

The P-omission rate of quoi (4/12=33.3%) is lower than that of lequel (98.1%), but higher than that of qui (2/7= 28.6%).

Similar findings challenging a cleft source have been reported in other non-P-stranding languages, such as Polish (Nykiel 2013). Molimpakis (2019) experimentally showed that P-omission is possible in Greek and German sluices, but not in clefts. Alshaalan and Abels (2020) presented similar results for Saudi Arabic.

4.3 Further issues for syntactic reconstruction

Further issues arise with syntactic reconstruction theories. Any analysis based on wh-movement predicts that sluices should be sensitive to locality constraints. In our corpus, we found examples which violate island constraints, such as the adjunct island. These instances occur with a correlate inside a relative clause as in (44), and inside an adjunct PP as in (45).

Some correlates also occur in a subject island, as in (46).

While various proposals have tried to explain island insensitivity under deletion theories (Merchant 2001, a.o.), it seems fair to say that direct interpretation approaches predict such insensitivity (if there is no syntactic structure at the ellipsis site).

In addition, we found instances of sluices without a verbal antecedent in the corpus, as in (47) and (48). In such cases, reconstruction of a verbal structure is challenging.

4.4 Processing-based factors

Our corpus study showed that (written) French sluices (with an explicit PP correlate) have a very high rate of P-omission (88.2%), driven by nominal correlates and anaphoric remnants: quel + noun (87.5%) and lequel (98.1%), which are also more frequent than the non-anaphoric remnants qui/quoi (33.3% P-omission rate). This confirms the effect of processing-based factors.

In our experiments, with nominal correlates, P-omission was found acceptable, and (anaphoric) nominal remnants favored P-omission (Exp.2), while (non-anaphoric) pronominal ones disfavored it (Exp.1). This aligns with cue-based retrieval mechanisms, given that more informative nominal remnants are better cues than pronominal ones.

In (spoken) English, Nykiel (2017) found that contentful (nominal) correlates exhibited a greater tendency to allow for P-omission (i.e., NP remnants). Conversely, non-contentful correlates (non-anaphoric pronouns), were predominantly followed by prepositional remnants. In our corpus, the effect of the correlate type was even stronger: with 20% P-omission rate with pronominal correlates and 92.5% P-omission rate with nominal ones. This goes in line with Nykiel (2017) and Nykiel and Hawkins (2020)’s theory since nominal correlates are more accessible in memory than pronominal ones.

4.5 An HPSG analysis

Our results go in-line with context-based direct interpretation approaches to ellipsis (Ginzburg & Sag 2000; Sag & Nykiel 2011, Ginzburg 2012). According to these analyses, there is no syntactic structure at the ellipsis site. Sluicing, then, yields a structure in which a clause is reduced to a wh-phrase, which derives its meaning from the context. In this section, we sketch an analysis based on Head-driven Phrase Structure Grammar (HPSG) (Pollard & Sag 1994; Müller et al. 2021), following Sag & Nykiel (2011). In this framework, linguistic signs can be lexical or phrasal, and their description is represented with a feature structure including syn (for syntax), sem (for semantics) and ctxt (for context) attributes. Well-formed signs inherit their properties from a hierarchy of types. An example of a hierarchy of phrases is in Figure 6, where phrasal types are distinguished according to the syntactic function of their immediate constituents (head, subject, complement, etc.) and to their clause type (declarative, interrogative, etc.).

Figure 6: Clausal types hierarchy.

We analyze sluices as clausal fragments (head-only phrases) with a propositional meaning (sem) computed from the context. We use Ginzburg & Fernandez (2010)’s Dialogue Game Board (DGB), which records information that emerges through dialogical interaction, keeping track inter alia of the interlocutor roles (speaker, addressee) and discourse topic(s), and includes a Maximal Question under Discussion (max-qud), updated throughout the interaction, which serves as the antecedent of the sluice. We rely on a sluice-phrase type, which inherits from a more general head-fragment-phrase type. As suggested by Ginzburg & Sag (2000) and Sag & Nykiel (2011), the content of the antecedent clause (max-qud) is analyzed as a polar question, i.e. lambda abstraction over an empty set of parameters, with a propositional content (p(j)). The sem of the sluice is a wh-question, with a non-empty set of parameters, and the same propositional content (p(j)).²³

The remnant is coindexed with a correlate, which is a salient utterance (sal-utt) in the dgb, with which it shares its syntactic properties (cat for syntactic category).²⁴

Applying this rule to example (19), repeated in (50) for convenience, the sluice retrieves its propositional content (p(j)) from the max-qud in the antecedent clause; the remnant, the only daughter of the S node, is coindexed (index j) with a contextual correlate (sal-utt) and matches its category (cat 1): it is thus either a PP as in (50a) (see Figure 7) or a NP as in (50b) (see Figure 8). The rels feature specifies the semantic relation expressed by the noun in the remnant.

In both examples, the antecedent clause is speaker A’s utterance: I talked to a friend. So the max-qud is a polar question (whether speaker A talked to a friend), and the sluice inherits the same propositional content to denote a wh-question (which friend A talked to). In both Figures 7 and 8, the sal-utt is the correlate, either the PP à un ami or the NP un ami, respectively. It shares its category and index with the remnant’s node: a PP (À quel ami) or a NP (Quel ami).

Figure 7: Analysis of a PP sluice with an explicit PP correlate.

Figure 8: Analysis of an NP sluice with an explicit PP correlate.

Appendix 1. Materials of experiment 1

Appendix 2. Materials of experiment 2