One approach to an account of language variation is to explore the potential role of the processing systems in shaping grammars (e.g., Hawkins 1994; 2004; MacDonald 2013; Hawkins 2014; McDaniel et al. 2015). Hawkins’s fundamental observation is that processing difficulties in some languages correspond to ungrammaticality in other languages (Performance-Grammar Correspondence Hypothesis, Hawkins 2014: 3). This phenomenon suggests that grammars are in part determined by processing pressures. Hawkins’s work shows that many cross-linguistic patterns can be accounted for in these terms. Similarly, MacDonald (2013) proposes the Production-Distribution-Comprehension account, which holds that sentence planning pressures affect language form. The perspective taken here is that human grammars are constrained by Universal Grammar, possibly with Merge at its core (e.g., Hauser et al. 2002). But Merge alone is not enough to fully account for the evolution of language and the possible grammars that emerge. The processing systems come into play as fundamental factors in the process of externalization (Chomsky 2010). Hawkins focuses primarily on the contribution of parsing pressures in shaping grammars, whereas MacDonald and colleagues argue that production pressures are primary and that parsing principles secondarily reflect sentence-formulation processes (e.g., Gennari & MacDonald 2009; MacDonald & Thornton 2009; MacDonald 2013). Similarly, McDaniel et al. (2015) argue that the production system plays at least as important a role as comprehension. In many cases parsing and production pressures drive grammars in the same direction. For example, a fronted topic aids the parsing process by signaling the topic at the outset, and also allows the speaker to begin with the most salient part of the message. But production and comprehension pressures can also be in conflict. For example, overt morphological case marking seems to benefit the parser rather than the speaker. Wasow (2002) argues that heavy NP shift is better accounted for in terms of production than parsing, proposing that the Principle of End Weight (Quirk et al. 1972: 14.8) derives from a sentence-planning preference rather than a parsing preference. McDaniel et al. make the same claim about the that-trace effect; the that-trace structure would create a challenge for the production system (since a new planning unit begins with an empty category), whereas it would facilitate the gap-finding process for the parser (since the sequence that+VP would clearly signal the gap).

The focus of the current research is a cross-linguistic phenomenon in long-distance (LD) extraction structures that is arguably best accounted for in terms of a response to a production pressure. The structure involves a type of grammatical attraction effect where a constituent extracted from an embedded complement clause is integrated with the upper clause verb, as though it had originated in the upper clause. Crucially, these structures are grammatical rather than the result of speech errors. The extraction attraction phenomenon (henceforth, EA) manifests itself differently in different languages, depending on language-specific morphology, as discussed below. The accounts of these structures in the syntactic literature vary depending on the specifics and do not treat EA as a uniform phenomenon (though both Boeckx 2008 and den Dikken 2009 include many of the structures in discussions of successive cyclicity). This lack of uniformity is not incompatible with the account proposed here. The general idea is that the syntactic system allows for certain possibilities, and processing pressures push languages in certain directions. Differences in the languages (such as in their morphology and word order) will lead to different solutions to the same processing challenge. In order to explore EA cross-linguistically, a descriptive approach is taken that abstracts away from in-depth syntactic analyses. This account is therefore compatible with varying analyses. It can be seen as aiming for the level of broad explanatory adequacy – why languages would have these unexpected structures.

This paper is organized in the following way. Section 2 is a discussion of the processing challenges posed by LD extraction structures, with focus on the sentence planning process. Section 3 characterizes the EA structure and its manifestation in a variety of languages. Section 4 considers non-EA LD extraction structures that are compatible with this account. Section 5 is a discussion of the broader significance of EA, in particular with respect to the role of language production in shaping grammars. Section 6 is a brief summary and conclusion.

Research on language production has yielded models that include a number of levels, from conceptualization to articulation (e.g., Garrett 1980; Levelt 1989). Broadly, speaking involves generating a preverbal message, retrieving lemmas (lexical items without their phonological form), encoding a morpho-syntactic structure (including features corresponding to functional elements), encoding the phonology (retrieving the phonological forms of the lexical items and then the phonological forms corresponding to the grammatical features), and articulating a phonetic plan (determining the phonetic forms and the prosody). The parsing process would run similarly but starting from the output and working back to the message.

Within this general framework, sentence planning proceeds incrementally, and the size of the planning window varies (Ferreira & Slevc 2007; Konopka 2012). The finding most relevant to the EA structure is that the clause is a major planning unit (Boomer 1965; Ford & Holmes 1978; Beatie 1980; Butterworth 1980; Bock 1996; Clark & Wasow 1998; McDaniel et al. 2010; 2015). So, for example, in the production of a two-clause sentence, planning would initially focus on the material up to the clause boundary (probably including the top node of the embedded clause), and planning for the inside of the embedded clause would follow.

Extraction structures, even in short-distance (SD) extraction, pose some challenges for the processing systems, since the word order change requires procedures that are different from the ones used in non-extraction. Consider the sentence Who did you see?, for example. Having the wh-phrase fronted may facilitate the parsing process by signaling that it’s a question, as well as the planning process by allowing the speaker to start with the focused element (which is likely at the forefront of their mind). But it also creates the parsing challenge of the search for the gap, and the production challenge of planning a clause with non-canonical word order. In fact, speech errors have been reported that show attraction effects in SD wh-structures, where the verb inflection indicates number agreement with the fronted wh-phrase rather than with the subject, as in Which flowers are the gardener planting? (Kimball & Aissen 1971; Dillon et al. 2017).

If the clause is a major processing unit, then it’s expected that LD extraction would pose an additional challenge. For the parser, the challenge is carrying the search for the gap into a new clause (Frazier & Clifton 1989). For the speaker, the challenge is planning the filler and gap structures at different points. Consider, for example, the sentence Who do you think that Jo saw? The speaker would first plan [Who do you think [CP]] and then plan the internal structure of the CP. The challenge is that who is structurally linked to the inside of the CP that isn’t planned till later.

Various ways the equivalent of LD wh-structures are handled cross-linguistically in languages that have SD wh-movement could be seen as responses to the processing challenge. Some examples are listed in (1).

Several of these, or a combination, often occur in the same language. For example, the that-trace effect is a combination of (1c) and (1e).

Taken together, these structures provide support for a role of processing in shaping grammars. This paper focuses on EA, which is arguably the clearest evidence in support of an account specifically based on the role of the production system, since for the parser, EA appears to be at best unhelpful and at worst misleading.

The EA structure is an LD extraction structure where the extracted element integrates morphologically with the higher clause verb instead of with the verb of the clause it extracted from, behaving like an argument of the higher verb. The specifics depend on the language’s morphosyntax. Generally, the consequence is that an LD extraction sentence beginning like Who do you think ends up morphologically parallel to an SD extraction sentence like Who do you admire? I treat the EA structure as a kind of attraction effect, since the morphology suggests a local relationship that appears to be inconsistent with the syntax. However, the attraction effects reported in the production literature are speech errors, usually involving number agreement (e.g., Bock & Miller 1991), whereas the EA structure is grammatical.

In addition to the EA structure being unexpected with respect to the grammar, it also seems problematic for the parser. Instead of aiding the gap-finding process, the morphosyntax signals falsely that the gap is in the higher clause. But the EA structure is plausibly beneficial to the speaker, who isn’t trying to find the gap. The challenge for the speaker is to plan the structure, as discussed above. The EA structure allows the speaker to plan the initial wh-phrase without having planned the internal structure of the clause it’s extracted from.

Based on an overview of data discussed in articles on LD structures, the following languages appear to have some kind of EA structure:

The EA manifests itself differently depending on the language’s morphosyntax. The next sections illustrate various types of EA structures.

Although EA structures are likely to be relatively rare cross-linguistically, it is important to note that many other LD extraction structures are compatible with this account, in particular, any LD extraction structure that allows the speaker to plan the wh-phrase and the higher clause without committing to the structure of the extraction clause. This situation occurs in any language where the morphology doesn’t mark for the grammatical function of the wh-phrase, regardless of whether there are other types of agreement with the extracted phrase. English is an example of such a language, since who do you think [CP] can be planned without committing to the internal structure of the CP.4 Other examples are given in (19).

In structures like (19a) and (19b), who do you think is morphologically parallel to who do you admire, as in English, since there is either no agreement in SD or LD structures (Dutch) or the agreement is the same for SD and LD structures and for subject and object extraction (Ojibwe). The fact that agreement is also required in the extraction clause in Ojibwe is not consequential, since the speaker plans that clause with the extracted wh-phrase in mind. The structure illustrated in (19c) might seem problematic, since in this case, which chairs did you say is morphologically different from which chairs did you paint? This is because the gender agreement would occur in the SD extraction structure. However, the lack of agreement on the higher verb doesn’t create a planning challenge; it only means that the speaker is aware that it’s an LD extraction structure. The speaker still does not need to plan the internal structure of the complement clause at the outset. It is important here to clarify that planning who do you think… on a par with who do you admire does not mean that the speaker is unaware of the complement CP in the LD structure; rather, it means that the speaker does not have to commit to its internal structure. This point is discussed in more detail in Section 5.

The account is also compatible with the alternatives to LD extraction that are listed in (1), though a detailed discussion is beyond the scope of this paper. Many of these cases could be accounted for in terms of either production or parsing; in other words, the two processing systems would push in the same direction.

The question arises of the kinds of structures this account would not handle. In other words, what kinds of counterexamples would challenge it? As mentioned in Section 3.1, some case-marking languages allow for non-EA LD extraction. If this kind of structure occurred freely within languages and widely cross-linguistically, it would be problematic for the account. But, as noted, the structure is limited in the languages that allow it. In terms of wh-agreement morphology, counterexamples would be structures where wh-agreement occurred on the verb of the higher clause to indicate the grammatical function of the element extracted from a lower clause. No reports of any such examples were found.

I will first consider in more detail how the EA structure addresses the planning challenge of the LD extraction structure. The consequence of the EA structure is morphological parallelism between an SD extraction structure and the higher clause of the LD extraction structure. However, the parallelism is likely limited to the overt morphology; the syntactic structure of the SD and LD structures will differ, since the argument structure of the higher verb does not include both the wh-phrase and the embedded clause it’s extracted from. The syntactic analysis of the structure (which may vary across languages) will account for how the wh-phrase ends up being morphologically treated as though it’s the clausal argument of the verb. For example, den Dikken (2009) proposes that in Hungarian the wh-phrase extracts from the higher clause and binds a null pronoun in the embedded clause. Letsholo (2002) proposes that wh-agreement in Ikalanga results from the wh-phrase moving to the FocP through the specifier of a Polarity Phrase, accounting for why the agreement shows up on the verb of the clause where the wh-phrase is pronounced. Whichever syntactic account is correct for a given language, that structure needs to be planned by a speaker producing an LD extraction structure. But regardless of the structure, the advantage to the speaker is not having to plan the internal structure of the extraction clause at the outset. In other words, in planning a sentence like Who do you think Jo admires? the speaker can plan the first major unit, which is the part down to the CP: Who do you think CP, saving the internal structure of the CP for a later step in the process.

A further question is how the EA structure would have arisen. The general idea is that the various cross-linguistic morphosyntactic devices (listed in (1)) arise from the grammar’s response to the processing challenge posed by LD extraction. In other words, these can be seen as a kind of grammaticalization. Put in terms of a sequence of steps, the process would be something like the following: Merge allows for long-distance dependencies; but LD extraction is challenging for the processors; so the language uses morphosyntactic devices to limit LD extraction. In the specific case of EA, the sentence formulation process would be faced with morphology that would normally require early planning of an embedded clause. To counteract this, the language adopts a morphosyntactic structure that allows for the extracted element to be morphologically treated as a constituent of the higher clause.

It is important to emphasize that the EA is not in itself a syntactic account of LD extraction in a particular language with this structure. Rather, this account is compatible with a variety of syntactic analyses. The goal here is not to choose among syntactic analyses or to provide a new one, but to find an explanatory account of why languages have gone in this particular direction. The idea is that the processors play a role in shaping languages’ instantiations of UG.

From an evolutionary perspective, the claim is that morphosyntactic devices beyond the basic innovation (e.g., Merge) would have evolved in tandem with the processing systems as part of the process of externalization. The processing systems would have developed procedures to deal with the grammar (combined with extra-grammatical pressures, such as memory limitations), and the evolution of these procedures (such as clause-by-clause planning) would in turn shape the further evolution of the morphosyntax. This process would result in a UG that is more flexible than the parameters model.

Before concluding, I consider the strengths of this account. It’s important here to distinguish between the account of the EA structure specifically and the general claim that sentence planning pressures affect grammatical structure. The EA phenomenon on its own may not argue convincingly for the general claim, but it adds to a growing body of evidence in support of it. As far as the specific account, it unifies and provides an explanation for a class of cross-linguistically occurring structures that are otherwise mysterious, unexpected both syntactically and semantically, and apparently in conflict with parsing considerations. The account also provides an explanation of why languages tend to resist LD movement (illustrated in (1)), though, as mentioned earlier, an account in terms of parsing would handle many of these cases as well. The production account could be further explored by studying LD extraction structures in experiments on language production; the prediction is that speech errors and other indicators of difficulty, such as dysfluency, should occur more in LD structures that involve non-EA overt morphology (such as non-switched case-marking of a fronted wh-word) than in other types of LD structures, including EA structures. Furthermore, when both kinds of structure occur in a language (which seems to be the case in Estonian, as discussed above), the prediction is that the EA structure will be used with greater frequency.

The EA structure, which occurs in a wide variety of languages and manifests itself in various ways morphologically, is unexpected grammatically, since the extracted phrase articulates with the clause where it is pronounced instead of with the clause it is extracted from. Embracing the claim that the processing systems play a role in shaping grammars, I have argued that the EA structure points specifically to the role of the sentence production system.