Argument clauses are special in that they allow wh-extraction under certain conditions. The examples in (332) show that the extracted wh-phrase can be e.g. an argument of the embedded verb or an adjunct. The traces (ti) refer to the original position of the moved wh-phrases in the embedded clauses.

The fact that wh-phrases can be extracted from argument clauses is rather special, since this is categorically excluded from adjunct clauses. For instance, the examples in (333) show that conditional clauses are strong (i.e. absolute) islands for wh-movement; they block wh-extraction of both arguments and adjuncts.

There are good reasons to assume that the wh-phrases in (332) are not moved into their target position in one fell swoop, but that this involves an additional movement step via the initial position of the embedded clause; cf. Section 11.3 and Chomsky (1973/1977). This is indicated by the structures in (334), where the traces ti and t'i refer to the positions that the moved phrase has occupied during the derivation; the CP/TP structure of clauses assumed here is discussed in Section 9.1.

The main reason for assuming that the wh-phrases are moved via the initial position of the embedded clause is that this immediately explains why examples like (335b&c) are unacceptable; since the clause-initial position of the embedded clause is already occupied by the subject pronoun wiewho, wh-extraction of the object/adjunct must take place in one fell swoop, which is not allowed. Note that the indices are used to indicate that ti is the trace of wie, while tj is the trace of the wh-phrase in the main-clause initial position. Note also that (335c) is acceptable if the adverb wanneerwhen is construed as a modifier of the matrix predicate, but the intended reading here is the one indicated by the traces, i.e. as a modifier of the embedded predicate dit boek kopenbuy this book.

Although more can be said about the restrictions on wh-movement (cf. Section 11.3.1), the above is sufficient for the main topic of this subsection: which verbs can function as bridge verbs, i.e. allow wh-extraction from their argument clauses? For example, while verbs of saying/thinking usually allow wh-extraction from their complement clause, verbs of manner of speech such as fluisterento whisper usually do not; this observation is attributed to an unpublished paper by Janet Dean Fodor in Erteschik-Shir (2006/2017) from 1967.

The examples in (337) show that irrealis verbs expressing a hope or a wish constitute another set of verbs that readily license wh-extraction in Dutch; cf. Haider (2010:108) for the same observation for those varieties of German that allow wh-extraction from embedded declarative dass-clauses.

Factive verbs such as betreurento regret, on the other hand, seem to systematically block wh-extraction from their complement clause, as examples like (338b-d) are generally considered unacceptable; cf. Hoeksema (2006:147).

Examples like (338b&c), in which an argument is extracted, are degraded, but are often considered better than examples such as (338d), in which an adjunct is extracted. This so-called argument-adjunct asymmetry is often attributed to the referential status of arguments; cf. Rizzi (1990). That this may be relevant is suggested by the fact that argument extraction becomes even better when the argument is D-linked, i.e. linked to the discourse in the sense that the addressee is asked to select an entity or set of entities from a presupposed list. Although there is considerable variation in acceptability judgments on examples of this type, some speakers even seem to find them perfectly acceptable; cf. Zwart (2011:209) for cases of object extraction. If acceptable, the examples in (339) show that factive clauses can be taken to be weak (selective) and not strong (absolute) islands for wh-extraction.

All in all, the list of bridge verbs seems to be relatively small. Hoeksema (2006) collected a sample of such verbs selected from written sources published after 1780. The complete collection consists of 963 tokens and 110 types. Most of the types have a very low frequency: nearly fifty types occur only once. If we restrict ourselves to types occurring at least six times, we get the result in Table (340). Bridge verbs are not only relevant for wh-questions but also for relative clauses and topicalization constructions.

begrijpen ‘to understand’	18	#	verzekeren ‘to assure’	8
beweren ‘to claim’	9		vinden ‘to consider/think’	34
denken ‘to think’	318		voelen ‘to feel/think’	9
geloven ‘to believe’	29		vrezen ‘to fear’	10
hopen ‘to hope’	37		wensen ‘to wish’	17
horen ‘to hear’	7		weten ‘to know’	73	#
menen ‘to suppose’	62		willen ‘to want’	119
oordelen ‘to judge’	7		willen hebben ‘would like’	6
rekenen (meaning unclear)	6	#	zeggen ‘to say’	59
vermoeden ‘to suspect’	15		zich voorstellen ‘to imagine’	8
vertrouwen ‘to trust’	6		zien ‘to see’	18
verwachten ‘to expect’	13

Since Hoeksema does not give his list of attested examples, we searched the internet with the string [Wat V[±past] je dat] what do/did you V that ...? in order to check whether the verbs in Table (340) occur in the relevant type of wh-question. The three types for which we could not find such examples are marked by a number sign; their use may be restricted to relative clause or topicalization constructions; cf. example (331b), which was taken from Hoeksema (2006). This leaves us with no more than twenty verbs that are regularly used as bridge verbs in wh-questions, and there is in fact only one verb, denkento think, that is really frequent (>300 tokens). Another relatively common bridge verb is the irrealis verb willento want (>100), but all other verbs are relatively rare. Other corpus-based research has shown a similar preference for the verb denken and, to a lesser extent, willen; cf. Verhagen (2005:119ff) and Schippers (2012:105).

Wh-extraction has given rise to two main lines of research, which Erteschik-Shir (2006/2017) refers to as the structural and the semantic approach, respectively. We will start with arguments in favor of the structural approach, according to which bridge verbs have some special syntactic property that makes their complement clauses transparent to wh-movement. For example, Erteschik-Shir mentions that verbs of manner of speech such as fluisterento whisper differ from verbs of speaking and thinking in that they can occur without a propositional clause, and suggests on the basis of this that the syntactic status of embedded clauses that co-occur with verbs of manner of speech is different from those that co-occur with verbs of speaking and thinking.

Further support for the structural approach comes from languages like English and German. In English, the set of bridge verbs seems to more or less coincide with the set of verbs that allow that-deletion in embedded declarative clauses, an observation again attributed to Janet Dean Fodor in Erteschik-Shir (2006/2017). Verbs of speaking/thinking, for example, allow that-deletion, while factive verbs such as to regret do not (although it is not too difficult to find cases on the internet). Since Dutch does not allow dat-deletion in embedded declarative clauses, we cannot provide similar Dutch data.

For German, Haider (1985:55) claims that most bridge verbs trigger embedded verb-second in the German varieties that have it (i.e. the northern varieties as well as standard German), although irrealis verbs such as mögento like are an exception to this general rule; cf. Haider (2010:124, fn. 25). The examples in (343) show that wh-extraction even requires the embedded finite verb to be in second position in those varieties that do not allow wh-extraction from embedded declarative dass-clauses, i.e. in dialects that do not allow a set-up as in (343a), wh-extraction obligatorily triggers verb-second as in (343b). Again, there is no similar evidence for Dutch, which does not have this form of embedded verb-second.

In short, arguments in favor of the structural approach to wh-extraction emphasize that bridge verbs obligatorily take a complement clause and that, at least in some languages, such verbs can affect the form of these clauses by licensing complementizer deletion or embedded verb-second. The semantic approach, originating in the seminal work of Erteschik-Shir (1973), emphasizes the effect of information structure on wh-extraction. The generalization is that wh-extraction is only possible from clauses that are focused, i.e. that express the new information of the clause. This immediately accounts for the fact that wh-extraction is normally not possible from factive clauses, as these present propositions whose truth is presupposed by the speaker; cf. the discussion in Section 5.1.2.3. However, it may also explain the contrast between the primeless and primed examples in (344b&c). Note in passing that we are dealing with weak islands in the primed examples; this is clear from the fact that they exhibit the argument-adjunct asymmetry discussed in Subsection I. Note further that the intended reading of the two (c)-examples is the one in which the adverbial wanneerwhen modifies the embedded clause; matrix scope is not intended here.

Erteschik-Shir (1973:95ff) shows that adding a negation to the matrix clause has the effect of defocusing the embedded clause; whereas example (344a) without negation can easily be used to introduce the proposition expressed by the embedded clause into the domain of discourse, example (344b) with negation is most naturally interpreted as a denial of the presupposed truth of the embedded proposition. This means that (344b), but not (344b'), is consistent with the generalization that wh-extraction requires the embedded clause to be part of the focus of the clause.

Since adding information to the matrix clause generally makes the embedded clause more likely to be defocused, the generalization predicts that this may also have a degrading effect on wh-extraction. This may explain the contrast between the examples in (336), repeated here as (345). The verb fluisterento whisper is more informative than the verb zeggento say because it contains a manner component: Jan expressed his assertion in a low voice. In fact, adding a manner adverb such as zachtjessoftly in (345c) seems to have a similar degrading effect on wh-extraction, a fact that seems to have escaped attention in the literature.

Erteschik-Shir’s generalization is also consistent with the findings in Verhagen (2005:124ff): on the basis of the corpus research mentioned at the end of Subsection I, it is claimed that attested cases of wh-extraction deviate only minimally from what is assumed to be the basic template, given here as (346). It is further claimed that “invented examples of wh-extractions are judged worse to the degree that they deviate more from the [...] pattern [in (346)]”.

The evidence reviewed above suggests to us that each of the two approaches has something different to contribute to the description of wh-extraction. The structural approaches may be correct in claiming that wh-extraction is subject to certain structural conditions, e.g. that the embedded clause is a complement of the verb in the matrix clause. The semantic approaches, on the other hand, may be correct in postulating additional conditions on the use of wh-extraction constructions, e.g. that the embedded clause is the focus of attention and thus constitutes the new information of the clause, which implies that the semantic contribution of the matrix clause should be limited. A review of various other syntactic and semantic approaches to island phenomena can be found in Szabolcsi (2006) and Szabolcsi & Lohndal (2017).