3.1 The VERBMOBIL corpus

As data source we used the transliteration files of all German dialogues in the VERBMOBIL corpus, recorded during spoken conversations between two native speakers of German. Participants were over 1000 adult native speakers of German recruited from regions around Munich, Bonn and Kiel. The conversations were recorded between 1993 and 2000 as part of a large computational-linguistic project on spoken language translation. During simulated phone conversations, the speakers discussed arrangements for a joint trip to a meeting. Seated in different soundproof rooms, they could hear each other via an audio connection. In most cases, they could also see each other through a window between the rooms.⁴

The dialogue contributions had previously been transliterated in detail, as part of the VERBMOBIL project. From the transcripts of the dialogues, we extracted all turns containing one or more tokens of an adverbial clause that is preceded by the main clause, and is introduced by the causal connectors weil, da, or denn (i.e., we included all “trailing” causal clauses and disregarded “leading” ones).⁵ We included da and denn because they also express causality but do not allow any verb placement freedom: da always introduces a VF clause, denn always a V2 clause. We found 721 usable clauses introduced with weil. Of these, 385 have VF word order (53%), and 336 occur with V2 (47%). There are 160 sentences with a denn, and 104 with a da clause. These numbers (Figure 2) do not include a small number of sentences that we had deemed unusable: Some clauses could not be classified unambiguously as V2 or VF because, apart from the subject, all nonverbal constituents were located in the endfield (e.g., as a complement or adverbial clause), or because they were interrupted and aborted before a potential V2 position of the verb. The resulting collection consisting of 985 usable sentences was subjected to the analyses we report below.⁶

Figure 2

Distribution of four classes of causal clauses in the VERBMOBIL corpus.

To convey an impression of similarities and differences between the extracted clauses, we list one sample sentence from each of the four classes; the four explananda are very similar in that they all designate reasons why a proposed time slot does not suit the speaker (compare (4a–d); connectors and finite verbs in the causal clauses are highlighted in bold font).

The transcriptions do not only contain the words said by the speakers but also tags denoting syntactic and/or prosodic features. Given the importance of the level of integration of explanandum and explanans clauses in the weil-V2 literature (see Section 2), we focused on two types of tags immediately preceding, during, or immediately following the connector: (1) those marking syntactic/prosodic clause boundaries, and (2) those denoting filled or unfilled speech pauses. The former consist of commas, full stops, and question marks; the latter mark disfluencies: hesitations, repairs, and editing terms. We have treated full stops and question marks as signs of relatively low levels of integration. Commas not accompanied by signs of disfluency were taken to reflect a relatively high level of integration. (The annotators had been instructed to insert a comma at the boundary between two clauses.) Combinations of a comma and a disfluency tag were assumed to represent intermediate integration. See Section 3.3 for details, and Appendix A.1 for annotated examples.⁷

3.2 Defining and recognizing explanation types

Our attempts at classifying the causal clauses into the types distinguished in the literature (propositional explanations, epistemic explanations, and speech act justifications) turned out to be harder than anticipated. Here are the definitions we tried out.

With respect to epistemic causality, we assume it involves Forward Causal Inference (FCI) or Backward Causal Inference (BCI; also known as abduction⁸). Both types of inference include a causal chain (CC),

Causation(Effect:Prop2, Cause:Prop1),

where Prop1 is a proposition denoting some observed event or state-of-affairs causing another event or state-of-affairs denoted by proposition Prop2.

FCI can now be characterized informally as follows. In FCI, the speaker claims Prop2 to be true (justified) based on a two-part argument consisting of (1) evidence that Prop1 is true, and (2) knowledge of the plausibility/likelihood of the event denoted by Prop1 being the cause of the event denoted by Prop2:

Justification(Claim:Prop2,
Arg:[Prop1 & Probable(Causation(Effect:Prop2, Cause:Prop1))])

For example, based on knowledge that Hans was speeding (Prop1), and that speeding drivers are often ticketed (Prop1 & (Causation(Effect:Prop2, Cause:Prop1))), one can rightly claim that Hans got a ticket because/for he was speeding.

In BCI (abduction), the roles of evidence and claim played by the two propositions are reversed, but not those of cause and effect:

Justification(Claim:Prop1,
Arg:[Prop2 & Probable(Causation(Effect:Prop2, Cause:Prop1))])

A person saying Hans was speeding because/for he got a ticket deploys this inference pattern to justify his/her claim about the likelihood of Hans having been speeding. From evidence that Hans was ticketed (Prop2), the speaker reasons backward—using the same CC as in the previous example—and makes the claim that Hans was speeding (Prop1).

Interrogative and imperative speech act justifications (see examples (3b) and (3c)) are similar to BCI, the only difference being that “Claim:Prop1” is replaced by “Question:Prop1” and “Command:Prop1”, respectively. Propositional explanations are simpler than any of the foregoing types: They consist only of a causal chain, whereas a CC is a proper part of any explanation based on logical inference. Another way of characterizing the difference between propositional (CC) and epistemic (FCI, BCI) explanations is as follows. In a propositional explanation, the set of properties attributed to the event that is described by the explanandum, receives an additional member (the property denoted by the explanans proposition: its cause); i.e., the explanandum event is “modified” by the explanans. In an epistemic explanation, the inference act justifying the conclusion/claim expressed in the explanandum proposition receives an additional specification: its arguments. Hence, the explanans does not “modify” the event denoted by the explanandum but the inference act yielding the explanandum proposition.

When inspecting the corpus sentences, we did not find weil, denn, or da clauses unambiguously classifiable as epistemic (as opposed to propositional explanations) by the above definitions.⁹ Virtually all the sentences express some form of psychological (motivational), social or institutional causality, but it was not possible to reliably distinguish cases expressing “mere” causality (CC) from cases where causality functions within FCI or BCI. For instance, it is hard to determine whether a person who produces the two clauses in Hans got a ticket because he was speeding in a prosodically more or less integrated fashion, asserts a causal relationship (CC), or lays a claim based on causal inference (FCI). Corpus sentences such as those in (4) and in Appendix A.1 testify to the problem. Hence, we will be unable to present numerical data pertaining to explanation types—except for 26 speech act justifications, which are easily recognizable by their explananda (interrogative main clauses tagged with a question mark: 22 cases with weil-V2; 2 with weil-VF; 2 with denn; 0 with da).¹⁰ This means the corpus is unsuited to study the effect of epistemicity, in particular BCI, on the choice of V2 vs. VF word order in weil clauses.

However, the corpus contains similar proportions of weil-VF and weil-V2 clauses. In combination with the annotations added to weil, da and denn sentences, this enables us to explore verb placement determinants other than explanation type (if any). Identifying such factors thus becomes the main goal of the present study. To this purpose, we focus on corpus data tapping into how the speakers delivered the transitions between explanandum and explanans clause, and how they shaped the explanans clauses themselves.

3.3 Main quantitative results

After initial explorations of the final collection of sentences, we decided to rate the explanandum-to-explanans transitions for prosodic integration of its delivery according to the following criterion. Transitions marked only by a comma were assigned the highest level of integration. If such a transition was additionally tagged with one or more disfluency markers, its integration level was coded as intermediate. The lowest level of integration was assigned to transitions tagged with a full stop or a question mark. In view of the sparsity of the latter type of transitions, we did not subdivide them into cases with vs. without disfluency tags. Figure 3 shows how the three levels of prosodic integration are distributed over clause types.

Figure 3

Distribution of integration levels over the four types of explanans clauses. (Percentages based on the raw scores in Table 1.)

Initial inspection of the sentence collection suggested that constituent order within the explanans clause is correlated with the syntactic shape and position of the subject NP of the clause.¹¹ In the large majority of VF clauses, the conjunction is immediately followed by a personal-pronoun subject (ich ‘I’, du ‘you’, er ‘he’, sie ‘she’, es ‘it’; wir ‘we’, ihr ‘you’, Sie ‘you’, sie ‘they’), whereas this is the case in only a minority of the V2 clauses. Hence, we also report counts of personal-pronoun subjects and the position of these subjects in the clauses. If a clause contains a personal-pronoun (PPro) subject, we distinguish between placement immediately after the conjunction (at what we informally call the Wackernagel position; label: PPro+W+), or at a later position (PPro+W–); subject NPs with any other type of head word (including demonstrative or indefinite pronouns) are coded as PPro–. See Figure 4 for the distribution over clause types. Note that, in V2 clauses, subjects carrying the “PPro+W–” tag actually occur in the midfield, due to subject-verb inversion. In VF clauses, hardly anything can be placed in between the conjunction and the personal-pronoun subject; we found only two¹² cases (one being weil an anderen Sonntagen ich keine Zeit habe ‘because on other Sundays I have no time’). This suggests that, if speakers—for information-structural reasons¹³ or otherwise—wish to put a non-subject constituent before a personal-pronoun subject, they better build a V2 clause. In what follows, we treat the occurrence of PPro+W+ subjects in explanans clauses as a factor promoting VF word order because these subjects facilitate word order decisions in VF clauses (see Section 4.2).

Figure 4

Distribution of subject NP types over explanans clause types. For abbreviations, see text. The percentages are based on the raw scores in Table 1.

Table 1 shows how many tokens of the various combinations of properties we found in the final sentence collection. The core of these data is presented graphically in Figure 5.

Figure 5

Graphical representation of the numbers in Table 1 expressed as percentages of the totals in the lower panel of the Table. Top: V2 clauses. Bottom: VF clauses. Abbreviations are explained in the text.

Also noteworthy is the distribution of disfluencies adjacent to the conjunction. Before the conjunction (i.e., at the end of the explanandum), the four clause types attract the same number of disfluencies (around 12%; see Figure 6). After the conjunction, the V2 explanantia are accompanied by somewhat more disfluencies than their VF counterparts. That the delivery of weil-V2 clauses tends to proceed less fluently than that of weil-VF clauses is also reflected by decelerated pronunciation of the conjunction itself (marked, in the transcriptions, by a delay tag before the final consonant of weil): These slow-downs occurred 12 times in total, but most of them (10) appeared in clauses that developed into weil-V2s.

Figure 6

Breakdown of disfluencies according to their position immediately before, in, or after the conjunction. Note: The percentage of in/after-conjunction disfluencies for weil-V2 clauses (21%) includes 17 cases where the annotators inserted a comma tag after weil. As these commas are more likely to represent a disfluency than a syntactic transition, we included them in the graph. Without them, the percentage drops to 16. For denn, there are 4 similar cases; without them, the percentage drops to 29.

Here are summaries and brief discussions of the results reported above:

• (#i) A substantial proportion (46.6%) of weil clauses adopts V2 word order (Section 3.1). This percentage resembles the one obtained by Dittmar & Bressem (2005) in a corpus of conversations recorded in Berlin around 1995, but deviates substantially from the data reported by Freywald (2010) for spoken corpora recorded between 1955 and 1974: only 8.3%. We will not go into the reasons for this cross-corpus variability, which may be due to differences between participating speakers, regions, speech situations, and/or to historical developments (see, e.g., Freywald (2010) for data and references to the literature).

• (#ii) However, no weil-V2 clauses we registered could be classified unambiguously as expressing an epistemic explanation in the abduction-based definition put forward in Section 3.2 (BCI).

• (#iii) In the majority of cases speakers accomplish the transition from the leading explanandum clause to the trailing explanans clause smoothly, without overt disfluencies (see Figure 3, and panel “Higher level of prosodic integration” of Figure 5). For VF clauses, this is the expected pattern. But from the weil-V2 literature one might get the impression that, more often then not, clauses are delivered with speech pauses before or after the conjunction. However, this is not borne out by the data in our corpus, where 64% of the weil-V2s is delivered without disfluencies (215 out of 336; Table 1). The corresponding proportion of denn-V2s is a little lower but still the majority (59 percent: 94 out of 160 cases). This data pattern amounts to a partial dissociation between prosodic and syntactic integration insofar as a relatively high level of prosodic integration appears to go together with a relatively low level of syntactic integration (V2 < VF). We also observed some instances of the opposite mismatch: cases of intermediate or lower prosodic integration marked by a relatively high level of syntactic integration (19% weil- and 11% da-VF cases in Figure 5, middle and right columns; examples from the literature are (3a) and (3b) mentioned in Section 2). The possibility of rather high levels of prosodic integration demonstrated by weil- and denn-V2s, and low such levels by weil- and da-VFs, received little attention in the literature, which focused on epistemic explanations (abductions) and speech act justifications, i.e., on weil-V2 structures where prosodic integration levels indeed tend to be low.

• (#iv) Immediately before the conjunction, disfluencies are relatively rare in V2 as well as in VF clauses. Immediately after the conjunction (and even during the pronunciation of the conjunction), disfluencies increase in V2 but not in VF clauses (Figure 6).

• (#v) In the large majority of VF clauses, the subject role is fulfilled by an NP consisting of a personal pronoun occupying the so-called Wackernagel position, i.e., adjacent to the conjunction. The corresponding percentages in V2 clauses constitute the minority (Figure 4).

• (#vi) The two types of VF clauses behave remarkably similarly with respect to the distribution of higher, intermediate and lower levels of integration, and the position and word class of the head word of subject NPs. The data pattern for the two V2 clause types resemble one another as well, but differ considerably from the VF pattern.

In the next section, we develop a production-based hypothesis about the origin of V2 word order in weil clauses based on these data. We seek this origin in special properties of the lexicalization and grammatical encoding processes in the speaker who is about to produce a weil-V2 clause.

4.1 Some syntactic and pragmatic properties of weil, denn, and da

It is generally recognized that denn is a paratactic rather than a coordinating conjunction (Höhle 1986; Pasch et al. 2003; Freywald 2009; Breindl 2009; Reis 2013: 225). Important arguments are the following. Unlike coordinating conjunctions such as und ‘and’ and aber ‘but’, denn imposes V2 word order on the clause it introduces (5a). Denn explanans clauses cannot get conjoined with VF explananda: In (5b), the denn clause can only be interpreted as paratactically conjoined with the erwarte ‘expect’ clause. Furthermore, denn does not license coordinate ellipsis (Antomo & Steinbach 2010: 27; e.g. (5c)); and two or more denn clauses cannot be coordinated (Breindl 2009: 280; compare examples (5e/f)). Finally, as is frequently noted in the literature, denn clauses cannot be within the scope of negation and quantifiers in the explanandum. (This also applies to the English causal connector for: The sentence John didn’t leave for he was ill entails that John stayed; it cannot mean that John left, although for a reason other than being ill.)

These properties of denn can be summarized in a simple rule: Denn is obligatorily followed by a syntactically entirely independent (unintegrated, autonomous) sentence. No special constraints seem to apply to the grammatical shape of this sentence, which can be of any type (declarative, interrogative, or imperative; simple, complex, or compound). Hence, after denn has delegated to an independent sentence the task of grammatically encoding the explanans content, this task can be carried out in the same manner as the task of grammatically encoding the content of an arbitrary subsequent sentence not preceded by denn (that is, only the usual constraints between consecutive sentences in connected discourse hold). But, although the denn explanans is syntactically independent of its explanandum, the rhetorical/discourse relation expressed by the connector still holds between the connected propositions.¹⁴

This excursion into denn enables us to make the crucial point that the mentioned denn properties hold for weil-V2 as well. The reader can verify this for her/himself by substituting weil for denn in examples (5b/c/f). It follows that, if weil precedes a V2 construction, this construction is best viewed as a grammatically fully independent sentence as well. (This obviously can only be true for weil explanantia that, like denn clauses, assert a proposition beyond the scope of explanandum constituents.)

If the explanans following the connector denn or weil is indeed encoded as a new sentence rather than as a clause licensed by denn or weil, the conceptualization process engendering the explanandum proposition must have terminated prior to embarking on the explanans—similar to the course of events during the production of any two consecutive sentences in a discourse. The resulting proposition is “frozen” and will not be modified anymore (except in the course of self-corrections due to “changes of mind”). We will refer to this “wrap-up” as closure of the explanandum. Closure of a proposition immunizes it against amendments. However, a closed proposition can still be embedded wholesale into an overarching proposition. Specifically, an explanandum proposition can become an argument in a causal chain or in an inference act (see Section 3.2). Importantly, closure of an explanandum proposition does not necessarily lead to closure of the syntactic structure encoding this proposition; this structure can still be extended, e.g. with a subordinate clause encoding the cause of an overarching (superordinate) causal chain.

If denn/weil-V2 explanantia indeed constitute new sentences and their production is preceded by pre-explanans closure of the explanandum proposition, then this implies that the explanandum and explanans contents are conceptualized in two steps, i.e., incrementally,¹⁵ each step engendering a proposition possessing (assertional) illocutionary force. However, not all explanandum-cum-explanante sentences are produced in this manner. Consider the examples in (6).¹⁶ Sentence (6a), with weil-VF, can be interpreted as resulting from integrative (i.e., non-incremental) conceptualization, yielding a single proposition with assertional illocutionary force: Some drivers in the domain of discourse owe their tickets to speeding. Within this assertion, the explanans expresses new information, whereas the explanandum encodes conceptual content that is presupposed (given, old). A speaker who intends to convey the explanans proposition as new information, will foreground (i.e., assert) it. Propositions that are already shared with the interlocutor—technically called presuppositions—, will remain backgrounded because they already belong to the interlocutor’s knowledge base. Mentioning a presupposed proposition serves as a signal to the addressee that the proposition should be activated, i.e., retrieved it from long-term memory and placed in her/his working memory (Dryer 1996).

Weil-V2 example (6b), on the other hand, expresses two propositions/assertions brought about by an incremental, two-step conceptualization process. This mode of conceptualization tends to be reflected prosodically by two separate sentential intonation contours, each spanning one clause (but see Section 4.3). Integrative conceptualization is more likely to yield a single sentential intonation contour spanning both clauses.

We now briefly note some properties of da, the third causal connector. Da cannot be within the scope of a negation element in the explanandum (see Sentence (7)) and often introduces epistemic explanations (Pasch 1983, 1997; Pasch et al. 2003: 397–399; see also Stede 2008: 224–225 and Bisiada 2013: 14). In this respect, da resembles denn. However, whereas denn introduces an explanans in the form of a main clause with foregrounded information, da expresses old/presupposed, hence backgrounded information. Da indeed does not start up a new sentence but forms a subordinate clause within the same sentence as the explanandum.

In contrast, weil is the only connector compatible with integrative as well as incremental conceptualization, and capable of introducing clauses with backgrounded as well as foregrounded information. Hence, weil spans a range of usage options that includes those of both denn and da. This suggests the prediction that weil is more frequent than denn and da. Table 2 verifies this for spoken language, implying that weil is a jack-of-all-trades: the most widely usable and most easily accessible lexical choice as causal connector.

Finally, consider Figure 7 for an overview of what we discussed in the present section. The left column assumes that conceptualization mode (incremental vs. integrative) correlates with closure point (pre- vs. post-explanans) and level of syntactic integration. The middle column shows the possible combinations of fore-/backgrounding of explanandum or explanans. Underlying this column are two—presumably uncontroversial—combinatorial assumptions: Incremental conceptualization always foregrounds the (leading) explanandum; and in case of integrative conceptualization, the explanans is always expressed by a subordinate clause irrespective of whether it is fore- or backgrounded. The right column summarizes the lexico-syntactic repercussions: eligible connectors and word orders in explanans clauses (to be detailed in Section 4.2).

Figure 7

Some important relations between the theoretical notions discussed in Section 4.1.

Figure 7 also reveals that epistemicity and propositionality of the speaker’s communicative intention are not allotted a role as (co)determinant of the shape of the explanans clause. We indeed view explanation type as an emergent property reducible to other aspects of the production process.

4.2 Weil-V2 as a performance phenomenon: Miscoordination between sentence production modules

The account we now propose for weil-V2 may be viewed as, in essence, a combination of the information summarized in Figure 7 and Table 2. The decision tree in Figure 8 highlights the crucial points. The situation denoted at the top in layer A of the figure is that of a speaker who is about to produce either an entire explanandum-cum-explanante sentence (integrative conceptualization) or to add an explanation to an already planned or produced explanandum (incremental conceptualization). The two actions listed in this layer can proceed in parallel (synchronously): activation of the causal connectors in the mental lexicon, and initiation of the conceptualization process that will ultimately engender the proposition(s) to be grammatically encoded. The three connectors compete with one another for insertion in the utterance-under-construction, with weil initially being the most likely winner due to its high frequency. This may lead to early insertion of weil into the sentence plan. The decision to close the explanandum before or after the explanans (layer D) depends on the planning process: incremental vs. integrative conceptualization and, in the latter case, on the presence of grammatical encoding problems incurred by the explanans (layer C). While the proposition(s) leading to (explanandum and) explanans is/are crystallizing out, their newness or givenness can already affect the syntactic (and prosodic) shape of the utterance, in particular foregrounding (assertion) vs. backgrounding (presupposition) of the explanans: layer E. In parallel, the activation levels of the connectors get modulated: Foregrounding the explanans boosts the activation level of denn, backgrounding of the explanans boosts da. The (already high) activation level of weil is not affected due to its neutrality regarding newness or givenness of the clause it introduces (see below for detail).

Figure 8

Summary of the proposed speaker-centered account of weil-V2 and related phenomena in the form of a decision tree. Nodes indicate speaker decisions or properties of current conceptual or syntactic structure. Edges below them indicate consequences of decisions or properties. In the text we use the labels A–G to refer to layers of nodes; and nodes within layers are counted from left to right.

What happens when the speaker selects the backgrounding option for the explanans (E2)? This option boosts the activation of da-VF which, if able to beat weil, will introduce the explanans clause; otherwise weil will do (G4 vs. G5). In both cases, the default VF word order will manifest itself.

If no closure has been applied to the explanandum proposition (D2), weil-VF is appropriate (G6); if the explanandum was closed, denn-V2 is licensed (G2). Weil-VF is also licensed in this case (G3), giving rise to an adverbial (subordinate) explanans clause uttered as a kind of afterthought. Reis (2013) treats this option as “unintegrated” weil-VF (e.g. in the context of examples such as in (3)).

The miscoordination underlying the ungrammatical weil-V2 alternative is depicted in node G1. This choice comes about as a consequence of early—in fact premature—insertion of weil, based on its high initial activation before the other connectors have undergone the influence of explanans fore- or backgrounding, and of pre- or post-explanans closure of the explanandum.

This account is not yet complete, though, because it fails to explain why the speaker does not always select the “unintegrated weil-VF” option (G3). Before addressing this issue, we provide some more details regarding lexical selection of connectors.

Denn and da are more specific connectors compared to weil. This affects the competition between the three connectors when the grammatical encoder needs one and sends a query to the mental lexicon. In line with much psycholinguistic evidence on the lexicalization process (“word finding”), we assume that the three connectors compete with one another for selection, and that the outcome is determined (partly probabilistically) by the activation levels of the competitors. The current activation depends not only on usage frequency (causing weil to be more highly activated than denn and da), but also on goodness of fit between the query addressed to the mental lexicon and the specific properties associated with candidate lexical items. Lexical items that “match” the query more closely (better fit) receive more activation than lexical items that match the query more poorly. For instance, if the query lists the property “foregrounding”, then extra activation will be gained by denn but not by weil (which is neutral with respect to prominence level). Specificity thus can raise the activation level of a matching item. In case of a denn vs. weil contest, if the speaker’s communicative intention includes the wish to foreground the explanans proposition, denn will get a head start which may enable it to beat the otherwise more frequent, hence more highly activated, connector weil. Likewise, the even less frequent da has a chance to beat weil if the query mentions backgrounding. Finally, also in line with mainstream psycholinguistic assumptions, the winning item sends inhibition (i.e., negative activation) to the competitors, suppressing them to below-threshold activation levels. As a consequence, when denn and da are defeated, they will be temporarily inaccessible.

The combined effect of frequency-based and specificity-based activation is the following: If denn or da, in spite of the competition from weil, do emerge as winner of the connector competition, they will introduce a clause whose pragmatic properties match the properties associated with them in the mental lexicon. In particular, denn will not be forced into combination with a proposition that is to be backgrounded or with a propositional explanation, neither will da have to introduce a foregrounded clause. On the other hand, the fact that weil is much more frequent than its contestants and easily beats them, readily leads to premature selection—premature in the sense of being uninformed by properties of the upcoming explanans proposition—, and thus running the risk of violating its obligatory VF property. Due to the temporarily subthreshold activation levels of the suppressed denn and da connectors, these more appropriate competitors cannot present themselves as substitutes. Indeed, we did not find a single V2 explanans where weil was replaced by denn.

We now turn to the question why speakers often select the ungrammatical V2 option of launching a new sentence (main clause) to accommodate the explanans proposition instead of the well-formed “unintegrated weil-VF” alternative. Our answer is based on data in our corpus that signal grammatical encoding problems while planning the explanans (layer C). One of our observations concerns the higher incidence of disfluencies in V2 than in VF explanantia. We attribute these disfluencies to sentence planning problems of any kind—lexical, syntactic, or morphological. Due to such problems, the grammatical encoder may need to mount extra processing capacity and working memory space. One remedy is to close the explanandum proposition and to allocate all available processing capacity to explanans proposition. This “focus (of attention) shift” will lead to rapid decay of the conceptual and syntactic structure built for the explanandum, and to availability of free processing capacity and space for the explanans.¹⁷ This strategy predicts that, other things being equal, explanans clauses that recruit less processing capacity are more likely to emerge with VF, and vice-versa.

We surmise that this relation underlies the notably higher proportions of non-personal-pronoun subject (PPro–) NPs in weil- and denn-V2 explanantia compared to their VF counterparts (see Section 3.3, especially Figure 4). Unlike PPro+ subjects, PPro– subjects do not have a privileged placement option such as the Wackernagel position; hence, PPro– subjects have to compete with other clause constituents for a suitable position more frequently than PPro+ subjects. This linearization factor, which increases the processing capacity load to be recruited for PPro– clauses compared to PPro+ explanans clauses, thus may have yielded a higher proportion of explanandum closures. A second factor promoting VF word—or rather, preempting explanandum closures that would give rise to V2—is that personal-pronoun NPs are easy to process referentially (the referents, internal representations of speaker and addressee, are easily accessible) and lexically (high-frequent, single-word NPs); hence, clauses containing one or more will mobilize less processing capacity than clauses containing full NPs instead. This reduces the need to gain extra working memory capacity by releasing a closed explanandum at the transition to the explanans: Thus, the grammatical encoder can afford to keep the explanandum structure active while processing the relatively simple explanans.

Another condition conducive to efficient management of restrictive cognitive processing capacity arises when the speaker produces a main clause without yet being aware that it will turn into the explanandum of a causative sentence. For example, consider a speaker who, having delivered a verbal description of some event or state of affairs, realizes that this description needs an explanation or justification, and starts planning the explanans clause during or after the delivery of the event description, hence after closure of the explanandum. The speaker’s awareness that an explanans is desirable can also come from gestural or other reactions from dialog partners (Eisenberg 2006: 4–5; Diessel & Hetterle 2011: 210). In such cases, the explanandum and explanans clauses are planned incrementally rather than in overlapping time intervals. Consequently, the explanandum is planned, encoded, and closed as a separate sentence; and when the explanans proposition becomes available, it is often too late to interact with the explanandum proposition, with denn/weil-V2 as only viable option. Table 1 indeed shows that when prosodic integration decreases (going from top to bottom through the table), the percentage of V2 clauses increases substantially: 44% in the top, 62% in the middle, and 86% in the lower panel.¹⁸ We attribute this rise to increasing likelihood for the explanandum to be closed and cleared from working memory due to increasing asynchrony of the two propositions.

In sum, weil-V2 can be accounted for as a performance phenomenon induced by miscoordination between strategies seeking optimal utilization of the processing capacity of the conceptualization, lexicalization, and grammatical encoding mechanisms.

However, the story does not end here.

4.3 An indirect consequence of failing coordination between sentence production modules

The weil-V2 engendering miscoordination scenario may have an indirect effect: If it unwinds frequently enough in speakers, it may induce a lexical acquisition/learning process in hearers. Since hearers are speakers themselves, their mental lexicon may develop an association between weil and the cognitive V2 scenario consisting of explanandum closure, focus shift toward the explanans, and launching a new sentence. Depending on the frequency of weil-V2 perception in the language community, this association can lead to increased weil-V2 production by speakers in spite of VF remaining the licensed word order. That is, their weil-V2 usage can get triggered by the association instead of by cognitive processing factors of the type to which the V2 engendering scenario presumably owes its existence.

This means we should reckon with the possibility that a non-negligible proportion of the weil-V2 clauses in our corpus is due to association-based facilitation of the construction rather than to cognitive load. Indirect support for this suggestion comes from the relatively low correlation we observed between syntactic and prosodic integration (see data summaries at the end of Section 3.3). A relatively high level of prosodic integration is demonstrated not only by the syntactically highly integrated weil-VFs but also by the syntactically unintegrated weil- and denn-V2s. The proportions of “higher level of prosodic integration” observed with weil- and denn-V2 clauses are substantial (weil 64%, denn 59%), although clearly smaller than the very large proportions for VF explanantia (weil 80%, da 89%).¹⁹

We account for the high level of prosodic integration observed for da-VF clauses by postulating an association between da and the following production scenario underlying its use: The intention to communicate an explanans proposition as given/presupposed is always preceded by closure of the foregrounded and asserted explanandum proposition. The original motivation for speakers to apply closure in this speech situation may be similar to the one underlying the production of an explanandum-cum-explanante sentence consisting of two foregrounded/asserted propositions: the fact that the conceptualization process delivers these propositions incrementally, i.e., sequentially. However, presumably due to high usage frequency in the language community, the association between da and explanandum closure has become strong enough to get elicited even when the two propositions are conceptualized more or less synchronously. (As da always introduces a backgrounded proposition, there is no pressure towards V2 because a main clause must carry assertional illocutionary force.)

Notice that the acquisition process proposed in this section differs from another conceivable acquisition process in language learners/users, viz. that weil-V2 could be the result of assimilation of weil to denn. The acquisition process we propose is secondary to—enabled by—the miscoordination stipulated in Section 4.2. This stipulation has the additional advantage of accounting for the direction of the assimilation: from weil-VF to weil-V2 instead of vice-versa. The assimilation hypothesis cannot rule out denn assimilating to weil, with denn-VF clauses as result.

Whatever the precise nature of the association formation processes in the mental lexicon, one should expect them to affect not only sentence production and comprehension by language users directly but also, indirectly, their grammaticality/acceptability judgments of self- and other-produced utterances.

5.1 Summary of corpus data and proposed theoretical framework

We have reported our examination of a collection of about 1,000 spoken German sentences that contain a causal clause (“explanans”) introduced by one of the connectors weil, da or denn, positioned immediately after the main clause (“explanandum”). We had extracted these sentences from a corpus of spoken dialogues (VERBMOBIL) recorded between 1993 and 2000. As part of the VERBMOBIL project, the dialogs had been annotated with tags marking major prosodic and syntactic boundaries, and various types of disfluencies (speech pauses, hesitations, self-repairs). We calculated how frequently these annotations occurred in combination with verb-second or verb-final word order in the explanans clause, and with ordering patterns of the first few constituents of the explanans. The observed frequency patterns suggest that an important factor leading speakers to select V2 after weil is anticipation, or actual experience, of grammatical encoding problems early on in the weil clause. We also noticed that, in this respect, weil-V2 clauses resemble the denn-V2 clauses in the corpus. Our original conjecture was that sentences with a V2 explanans are somehow harder to plan than sentences with a VF explanans. However, a grammatical phenomenon uniquely associated with denn—i.e., not with other connectors—suggested that the V2 explanans could facilitate the processing of complex explanandum-cum-explanante sentences. (In the German linguistic literature, this denn property is known as parataxis; see citations at the onset of Section 4.1). Denn affords the speaker a “divide and conquer” strategy that splits the parallel-interactive planning of a complex sentence into the sequential planning two simpler sentences, first the explanandum, then the explanans sentence.

Could V2 after weil be another manifestation of the same divide et impera strategy? If so, the V2 clause introduced by weil should not be viewed as a clause subordinate to the preceding explanandum but as a second independent sentence featuring V2 as the default ordering pattern for main clauses. This presupposes that the explanandum proposition has undergone closure (“freezing”). In this case, V2 is the only possible word order option within the explanans clause. Otherwise (i.e., in the absence of closure), the explanandum proposition can be extended with an explanans proposition, enabling the grammatical encoder to express the latter proposition as a VF clause attached to the explanandum clause (provided the explanans is foregrounded/asserted). The latter scenario is the more likely, the simpler the planning process for the explanans. This predicts, for instance, a higher incidence of VF in clauses with personal-pronoun subjects occupying the preferred Wackernagel position. However, when grammatical encoding problems with respect to the explanans are expected or experienced, the encoder can deploy the divide and conquer strategy, and cast the explanans in the V2 mold.

Note that closure of the explanandum in case of denn is not an encoder operation specifically associated with this connector. If indeed denn is paratactically related to the explanandum, with the denn clause starting up an independent new sentence, then explanandum closure is just the standard operation the encoder executes when terminating any sentence prior to embarking on the next sentence in the discourse. By virtue of the linguistic arguments put forward in Section 4.1, we stipulate that this also holds for weil when followed by a V2 clause.

The proposed account of weil-V2 proposed in Section 4 assumes that VF is the only word order option licensed by weil. The reason it is regularly followed by a V2 clause has to do with weil’s high frequency in language use—much higher than that of denn and da—and with the fact that it does not impose pragmatic restrictions on the explanans clause. The weil clause may express a foregrounded/new/asserted proposition or a backgrounded/old/presupposed proposition. This contrasts both with denn, which introduces an assertion, and with da, which introduces a presupposition. The broad usability of weil conspires with its high frequency and high resting level of activation to make it an easily accessible and attractive lexical candidate to express causation or causal inference. However, if selected early, its VF property may clash with V2 word order in an explanans clause that is currently being encoded as an independent new sentence—due to a currently high cognitive processing load (Section 4.2) or to a strong association between weil and the closure-with-focus-shift scenario (Section 4.3); and since the early, premature selection of weil has led to suppression of the appropriate alternative denn, the latter cannot rescue the imminent violation of weil’s VF projection constraint.

The larger the number of language users in whose mental lexicon an association develops between weil and the V2 engendering cognitive scenario, the lower the threshold for weil-V2 to get activated and to affect speech situations without, or with only mild, cognitive processing load. A trend in this direction within the language community is expected to render the weil-V2 construction acceptable in spoken language use.

In the remainder of this section, we compare our account with extant theoretical frameworks, and confront it with additional empirical findings.

5.2 Some linguistic and psycholinguistic repercussions

Separate weil-V2 and weil-VF lexical entries? In the final section of her paper, Reis (2013) expresses dissatisfaction with an important theoretical implication of her analysis of weil-V2 phenomena: the requirement that the lexicon of modern German contains two lexical items for the conjunction weil, which project a VF clause and a V2 clause, respectively. In view of the fact that the meanings associated with the two weil items are identical, she would prefer “to posit only one lexical entry and to derive the distributional differences (syntactic as well as interpretive) from the regular impact of other factors”.²⁰

The proposed model preempts the necessity of postulating two weil entries in the mental lexicon of speakers of German who regularly use VF and V2.²¹ According to the model, V2 constituent order in the clause following weil is not licensed by weil but by word order rules for main clauses (see also Footnote 14). According to the model, weil-V2 emerges precisely because the explanans proposition has “escaped” from the weil-clause. An additional advantage is that no selection process between two weil entries is required.

Explanandum closure also in the hearer? There is mounting psycholinguistic and neurobiological evidence of close collaboration, if not substantial overlap, between the cognitive processing modules responsible for the grammatical aspects of sentence production and comprehension, that is, between the grammatical encoder and the grammatical decoder (Kempen 2000, 2014; Segaert et al. 2012). Hence, the comprehension system of a hearer may “know” that the production of a V2 explanans clause by the interlocutor must have been preceded by closure of the proposition expressed in the explanandum. In response, the hearer will also close his/her interpretation of the explanandum, thereby restricting the interpretation of the ensuing explanans clause to one resulting from incremental conceptualization, i.e., an epistemic one (FCI or BCI).

This reasoning predicts the results of an interesting experiment reported by Antomo & Steinbach (2010: 25ff). In a test with printed materials, they presented their participants (a group of German university students) with sentences such as (8a) and (8b). These sentences are ambiguous in that they leave open which of the two described events is the physical cause of the other one. They found that weil-VF clauses such as in (8a) elicited a large majority (72.3%) of propositional explanations (airbag inflation as cause), whereas weil-V2 clauses as in (8b) elicited only a small majority (55.5%) of epistemic explanations (airbag inflation as effect).

The authors put forward a detailed pragmatic-linguistic account of why weil-V2, in contrast with weil-VF, allows a reversal of the understood direction of causation. This account (partly going back to Blakemore & Carston 2004) postulates a comprehension system based on specific “pragmatic processing units” and “processing stages” (for a brief summary, see Section 2). We make the—presumably uncontroversial—assumption that the grammatical decoder parses the VF clause in (8a) as resulting from integrative conceptualization. This engenders the propositional explanation where airbag inflation is cause of the accident—the students’ preferred reading (72.3%). The same standard parsing strategy fails when the decoder detects V2 word order in the explanans. However, the grammatical decoder recognizes in the weil-V2 structure a pattern that the grammatical encoder generates when the communicative intention necessitates incremental conceptualization, namely, explanandum closure. Therefore, the decoder also closes the explanandum proposition and takes the relation between explanandum and explanans as one of justification (FCI or BCI) rather than CC (a causal chain; see Section 3.2). This implies an epistemic rather than a propositional explanation, with the inflated airbag as evidence/argument for the inference about the accident. However, given knowledge of the modern world, the hearer cannot decide between airbag inflation as cause or as effect of the accident (FCI versus BCI). Hence, the participants could not do much better than making a random choice (55.5%).

Verb-second in adverbial clauses with other connectors? As is well-known, V2 word order can also be projected by certain non-causal subordinating conjunctions. Examples are obwohl ‘although’ (concessive) and während ‘whereas’ (adversative), often preceded and/or followed by a short prosodic break. However, other subordinating conjunctions (temporal, conditional, instrumental, resultative) do not license V2. (To illustrate in terms of VERBMOBIL: Of the 71 obwohl tokens in the corpus, 47 preceded a V2 clause (66%), and 34 of the latter were tagged with a disfluency mark after the connector.)

It does not seem farfetched to generalize the theoretical framework we developed for causal V2 clauses to concessive and adversative clauses: Trailing obwohl and während clauses with V2 assert independent propositions, on a par with weil/denn-V2 clauses; and the main clauses preceding them are also independent speech acts. If the generalization holds, we also need to assume that concessive and adversative V2 clauses are accompanied by closure of the main clause proposition. The generalization also forces us to postulate that the other group of connectors (temporal, conditional, etc.) do not close their main clause. Stated differently, they keep the main clause proposition open, susceptible to modification by additional conceptual content (integrative conceptualization); and the result is one asserted proposition spanning the entire sentence (similar to the single asserted proposition conceived during planning an integrated causative sentence where the causal clause modifies the meaning of the main clause). We are confident that the first part of the generalization—obligatory closure of the proposition preceding obwohl and während—is correct, viewing the disfluencies that tend to accompany them.

Whether the second part holds (i.e., no closure before temporal, conditional, etc., connectors), is an open empirical issue. An indirect argument in its favor is derivable from the VERBMOBIL data concerning the connector sodass ‘so that’—the “resultative” VF connector which expresses a relation similar to causality, although in a direction opposite to that of weil-VF. (In the corpus, the connector is spelled as two contiguous words: so daß—according to old German spelling.) The corpus contains 63 exemplars of sodass, all unambiguously classifiable as VF. Remarkable, and in strong contrast with obwohl, is their high level of prosodic integration: Only 6 sodass-VF clauses were realized with a disfluency around the connector. Moreover, 46 out of the 63 sodass tokens (73%) had a personal pronoun subject in Wackernagel position (PPro+W+; the remaining 27% were all PPro–). These proportions strongly resemble those of weil/da-VF depicted in Figure 4. On the assumption that sodass clauses represent foregrounded/new information within the complex sentence, it follows that the underlying production process is similar to that depicted in the rightmost branch of Figure 8 (from layer A to node G6).

Although this scenario indeed does not include closure of the main clause preceding sodass, it leaves open the possibility for the sodass clause to be within the scope of operators in the main clause. However, we have not been able to find—or construct—any well-formed example of such a structure, suggesting that the apparent prohibition on cross-clause scope relations with sodass has another origin than pre-sodass closure of the main clause. Here are two options (both presuming that sodass denotes a causal relationship). The first one proceeds from the assumption that speakers who introduce scope relations between clauses, deploy logical reasoning as part of the conceptualization process, hence manipulate propositions. This would rule out using sodass if its lexical entry would specify that it does not connect propositions but events referred to by propositions (i.e., defining sodass as a referential rather than inferential connector; cf. the familiar distinction between cause/effect vs. argument/conclusion). However, even if true, this would merely describe the observation, not explain it. The second speculation: Sodass clauses are always foregrounded/asserted whereas the preceding main clauses are virtually always backgrounded/presupposed. This would imply integrative conceptualization and preclude the divide-and-conquer solution of splitting the complex sentence into a pair of independent sentences. Consequently, the frequency of this solution would be too low to trigger the learning process sketched in Section 4.3: associating sodass with main clause closure. However, this hypothetical account is incomplete without a reason why sodass, in contrast with the almost equally frequent obwohl, is rarely preceded by a foregrounded/asserted main clause. Hence, we have to leave unanswered the question why there is no sodass-V2 phenomenon.

Finally, we wish to note that the theoretical framework developed in this paper should be viewed as a set of hypotheses suggested by the corpus data, and needs additional empirical support from experiments that test predictions derived from it. This requires experiments manipulating the speaker’s communicative intention and the cognitive load imposed by (un)problematic explanans clauses as independent variables, and assess the tendency for speakers to use verb-second or verb-final word order as dependent variable, preferably in conjunction with other behavioral and neurobiological measurements. Empirical work of this type is likely to have impact beyond the domain of a handful of lexical items, on the study of language production generally.