12.3.1 Serial vs. Interactionist Processing Models

Psycholinguistic models of sentence processing fall, roughly, into two types. One type assumes that a syntactic analysis is initially performed on the basis of bottom-up (i.e. syntactic category) information only. According to serial or syntax-first processing models (Frazier Reference Frazier1979; Frazier and Clifton Reference Frazier and Clifton1996; Friederici Reference Friederici2002), the syntactic analysis computed initially is not influenced by higher-level information such as semantic plausibility or contextual biases. Ambiguous words or phrases are integrated into the emerging structural representation in accordance with a narrow set of structure-based minimal effort principles. Many cases of syntactic misanalysis can be attributed to a general preference for the simplest possible analysis compatible with the current input, and for integrating new incoming words or phrases into the constituent currently being processed. These two structure-based economy constraints are known as Minimal Attachment and Late Closure, respectively (Frazier Reference Frazier1979; Frazier and Clifton Reference Frazier and Clifton1996).

This hypothesized preference for the simplest syntactic analysis provides a plausible account for many types of garden-path phenomena. For garden-path ambiguities as in The horse raced past the barn fell, for example, analysing the verb raced as a main verb (7a) requires a less complex representation than analysing it as a relative clause-introducing participle (7b), an analysis that requires postulating an embedded clausal constituent which separates the head of the subject phrase (the horse) from the predicted main clause predicate.

1. 1. a. [_DP The horse] [_VP raced [_PP past the barn]]

   2. b. [_DP The horse [_RC raced [_PP past the barn]]] [_VP fell]

Another well-attested processing preference is for new incoming words or phrases to be integrated into the most recently processed constituent. This Late Closure (or local attachment) preference can, for example, account for the parser’s initial preference for analysing the postverbal DP the beer in (2a) as a direct object of the preceding verb drinking rather than as the subject of the (predicted) main clause. The alleged universality of the Late Closure principle has been called into question, however, by findings suggesting that it is not necessarily applied consistently across different languages (see, e.g., Brysbaert and Mitchell Reference Brysbaert and Mitchell1996; Carreiras and Clifton Reference Carreiras and Clifton1993; Cuetos and Mitchell Reference Cuetos and Mitchell1988; Frenck-Mestre and Pynte Reference Frenck-Mestre and Pynte1997; Papadopoulou and Clahsen Reference Papadopoulou and Clahsen2003).

Information other than structural economy constraints which can affect the likelihood of misanalysis includes probabilistic information such as verb subcategorisation biases, semantic information such as noun phrase (in-)definiteness, and referential information provided by the situational or discourse context. Experimental findings which suggest that these factors are able to affect the processing system’s initial analysis of ambiguous input (e.g. Altmann and Steedman Reference Altmann and Steedman1988; Britt Reference Britt1994; Garnsey, Pearlmutter, Myers and Lotocky Reference Garnsey, Pearlmutter, Myers and Lotocky1997; Trueswell, Tanenhaus and Kello Reference Trueswell, Tanenhaus and Kello1993; Trueswell et al. Reference Trueswell, Tanenhaus and Garnsey1994) are problematic for strictly serial (i.e. syntax-first) processing models.

Interactionist models, by contrast, assume that sentence processing is influenced by a range of simultaneously interacting linguistic (and, possibly, non-linguistic) constraints, with semantic and pragmatic interpretation cues potentially affecting the initial analysis (for review, see McRae and Matsuki Reference McRae, Matsuki and van Gompel2013). These models differ from syntax-first models in that they do not assume syntactic or category information to be in any way privileged. Under this view, the way ambiguous input is first integrated into the current representation will be influenced by lexical-semantic, pragmatic and probabilistic biases (e.g. MacDonald Reference MacDonald1994). Interactionist models are supported by the observation that garden-path effects can often be ameliorated, or even prevented, by semantic or pragmatic information. In (8) below, for instance, the sentence-initial animate DP the defendant in (8a) is a much better potential Agent argument of the verb examine than the inanimate DP the evidence in (8b). This makes comprehenders more likely to initially adopt an erroneous main verb analysis for examine in (8a) than in (8b) (Trueswell et al. Reference Trueswell, Tanenhaus and Garnsey1994 – but cf. Ferreira and Clifton Reference Ferreira and Clifton1986 for different findings).

1. 1. a. The defendant examined by the lawyer turned out to be unreliable.

   2. b. The evidence examined by the lawyer turned out to be unreliable.

Many interactionist models assume that all possible analyses are computed in parallel, ranked either according to preference or likelihood (e.g. Gibson Reference Gibson1991) or with respect to their relative activation levels (e.g. McRae, Spivey-Knowlton and Tanenhaus Reference McRae, Spivey-Knowlton and Tanenhaus1998) and continuously evaluated as processing proceeds. Analyses that prove incompatible with the current input will be discarded or their activation level will be reduced.

While there is broad agreement that structural ambiguity resolution is influenced by a variety of different types of constraints (Altmann Reference Altmann1998; Gibson and Pearlmutter Reference Gibson and Pearlmutter1998), psycholinguists still disagree as to (a) whether processing is serial or proceeds in parallel, (b) whether one or more information sources are privileged in some way and (c) the point in time at which different types of information begin to affect processing. Other controversial issues concern the separability of grammar and parser and the way grammatical knowledge interacts with least-effort processing constraints such as the hypothesized preferences for local attachment and for the smallest possible structural representations. An in-depth discussion of these issues is beyond the scope of this chapter, however.⁶

12.3.2 Recovering from Misanalysis

In sentence processing research, the term reanalysis is used to refer to the mental processes involved in correcting an erroneous first analysis of the input. Successful reanalysis during processing will help ensure that the listener or reader arrives at exactly the interpretation that the speaker or writer intended, after initially being side-tracked by ambiguous or underspecified input. The psycholinguistic use of the term thus differs from its use in historical linguistics, where reanalysis refers to the re-structuring of (ambiguous) surface strings in the course of language change (Langacker Reference Langacker and Li1977). To avoid any unnecessary confusion, in the following, I will use the term neoanalysis (Andersen Reference Andersen and Andersen2001a; Traugott and Trousdale Reference Traugott and Trousdale2013) to refer to reanalysis in the historical linguistics sense.

Serial and interactionist processing models make different assumptions about how the parser goes about recovering from initial misanalysis. From the point of view of parallel or interactionist models, corrections made in the face of incompatible or disambiguating input may, for example, involve a re-ranking of options (e.g. Gibson Reference Gibson1991) or a change to the relative activation levels of competing analyses (e.g. McRae et al. Reference McRae, Spivey-Knowlton and Tanenhaus1998). Serial models, on the other hand, typically assume that an erroneous structural representation needs to be repaired, following backtracking and error identification (see, e.g., Fodor and Inoue Reference Fodor and Inoue1994; Sturt, Pickering and Crocker Reference Sturt, Pickering and Crocker1999). Individual proposals differ with regard to the technical details of how repairs might be accomplished, but there is evidence showing that some types of error are easier to undo than others.

Consider sentences containing temporary subject/object ambiguities such as (9a,b), for example. Sentences such as (9a), which involve an adjunct clause that precedes the main clause, often give rise to stronger garden-path effects than sentences such as (9b), in which the main clause verb takes a clausal complement, following the point of disambiguation, the auxiliary was (e.g. Sturt, Pickering and Crocker Reference Sturt, Pickering and Crocker1999).

1. 1. a. While Peter was baking the cake was sold to a young woman.

   2. b. Mary reported the crime was extremely serious.

This indicates that detaching the ambiguous DP (the cake in 9a, and the crime in 9b) from the preceding verb and reanalysing it as the subject of the following predicate incurs greater processing cost in (9a) than in (9b). From the point of view of repair-based approaches to reanalysis, this observation has been argued to suggest that reanalysis which involves elevating a misanalysed constituent to a hierarchically higher position (such as the main clause subject position in 9a) is more difficult than reanalysis that does not change the original dominance relations (e.g. Pritchett Reference Pritchett1992). Parallel-interactionist models, on the other hand, would assume that a higher processing cost is associated with changing the relative ranking or activation levels of the two competing analyses in (9a) compared to (9b), possibly because a verb such as report in (9b) admits clausal complements quite readily (compare e.g. Garnsey et al. Reference Garnsey, Pearlmutter, Myers and Lotocky1997).

A wrong analysis may also be harder to undo if it can be maintained for a long time (e.g. Ferreira and Henderson Reference Ferreira and Henderson1991), as in example (7) above, and an initially highly plausible misanalysis is more difficult to correct than an initially implausible one (e.g. Pickering and Traxler Reference Pickering and Traxler1998). Contextual or discourse biases may also affect the way structural ambiguities are resolved (e.g. Altmann and Steedman Reference Altmann and Steedman1988). Our processing system is able to correct many types of initial misanalysis automatically, that is, without our conscious awareness. Some types of misanalysis, for instance that involved in the processing of difficult garden-path sentences such as (7), may require some conscious effort to be resolved or can lead to processing breakdown. Other cases of reanalysis failure may go unnoticed rather than leading to processing breakdown, with the resulting misinterpretation being maintained (Christianson, Hollingworth, Halliwell and Ferreira Reference Christianson, Hollingworth, Halliwell and Ferreira2001; Ferreira Reference Ferreira2003).

12.3.3 ‘Good Enough’ Processing

As we saw earlier, for temporarily ambiguous sentences that are strongly biased toward a specific analysis, the parser normally commits to this analysis very quickly. Psycholinguistic experiments have shown that the resulting incorrect interpretation is sometimes carried forward even after a sentence has been grammatically (or otherwise) disambiguated. Christianson et al. (Reference Christianson, Hollingworth, Halliwell and Ferreira2001), for example, found that readers would often misunderstand sentences such as (10) below, in which the postverbal NP the deer is a highly plausible direct object of the verb hunted, despite the fact that the following verb ran clearly signals that the direct object analysis is incorrect.

1. (10) While the man hunted the deer ran into the woods.

When asked whether the man hunted the deer, comprehenders would frequently respond with ‘yes’ even though sentence (10) does not in fact provide any information at all as to what the man was actually hunting. Such cases of semantic persistence could, in principle, result from failed or incomplete reanalysis (Christianson et al. Reference Christianson, Hollingworth, Halliwell and Ferreira2001), from a failure to erase incorrect representations completely from memory (Slattery, Sturt, Christianson, Yoshida and Ferreira Reference Slattery, Sturt, Christianson, Yoshida and Ferreira2013) or from comprehenders computing underspecified syntactic analyses in the first place. According to Ferreira (Reference Ferreira2003), comprehenders may sometimes compute syntactically shallow, semantics-based representations of the input that are just ‘good enough’ for recovering basic sentence meaning. Another way to account for Christianson, Ferreira and colleagues’ findings is to assume a dual-pathways model of sentence processing, according to which syntactic and semantic analyses can proceed in parallel but largely independently (e.g. Townsend and Bever Reference Townsend and Bever2001), and with the semantically driven or ‘good enough’ route to comprehension sometimes winning out over the syntactic one (for review and discussion, see Ferreira and Patson Reference Ferreira and Patson2007).

Effects of semantic persistence have also been observed in studies using time-course sensitive experimental methods such as eye-movement monitoring (Jacob and Felser Reference Jacob and Felser2015; Sturt Reference Sturt2007). Jacob and Felser had participants read sentences containing temporary subject/object ambiguities such as (11). These sentences were disambiguated twice: first by a finite auxiliary (e.g. were) in need of a subject, and later on by a verb phrase that was pragmatically inconsistent with a direct object analysis (e.g. being arranged for the party).

1. (11) While the young girl was drawing the flowers were carefully being arranged for the party.

Despite receiving two consecutive disambiguation cues, native English readers retained the incorrect direct object reading of the flowers in around one in three cases.

Taken together, these findings suggest that encountering disambiguating input does not always lead to successful structural revision or repair. Understanding how syntactic misanalysis comes about – and why it sometimes persists – in real-time language processing might also contribute to a better understanding of syntactic neoanalysis.

Recall that for the kind of garden-path sentences and attachment ambiguities considered thus far, an initial misanalysis will also give rise to an erroneous interpretation. Computing a fully detailed, accurate grammatical representation is not always necessary for recovering the intended meaning, however. For illustration, consider complements of perception verbs such as see in I saw the neighbor painting the fence. These can be shown to be three-ways structurally ambiguous (e.g. Declerck Reference Declerck1982; Felser Reference Felser1999). The different analysis options are shown, in simplified form, in (12a–c).

1. 1. a. Reduced relative clause analysis

      I saw [_DP the neighbour [_RC painting the fence]]

      “I saw the neighbour who was painting the fence”

   2. b. Participial adjunct analysis

      I saw [_DP the neighbour] [_CP painting the fence]

      “I saw the neighbour as she was painting the fence”

   3. c. Clausal complement analysis

      I saw [_AspP [_DP the neighbour] [_VP painting the fence]]

      “I saw the event of the neighbour’s painting the fence”

Example (12a) shows the reduced relative clause analysis, with painting the fence functioning as a non-finite clausal modifier of the DP the neighbour. In (12b), painting the fence functions as an adverbial rather than as a postnominal modifier. Whilst in (12a) and (12b), the neighbour is the direct object of see, the complement of see in (10c) is a non-finite clausal constituent (an Aspect Phrase, in Felser’s Reference Felser1999 analysis), with the neighbour functioning as the embedded clause’s subject. As the paraphrases in (12a–c) indicate, the three structural options each correspond to slightly different semantic interpretations. Nothing crucial hinges on the specific choice of analysis, though, as all three options essentially describe the same thing being witnessed: the neighbor in the process of painting the fence. Comprehenders who compute a different analysis from the one intended by the speaker or writer would thus not miss out on any key information. Provided that contextual information does not somehow disambiguate the sentence, this gives rise to two possibilities: comprehenders will either compute one of the three possible analyses in (12a–c) at random (or probabilistically, based on their relative occurrence frequencies) or compute a syntactically underspecified analysis.

There is processing evidence to suggest that comprehenders do indeed sometimes compute underspecified grammatical representations. Swets, Desmet, Clifton and Ferreira (Reference Swets, Desmet, Clifton and Ferreira2008), for example, report reading-time evidence showing that readers do not necessarily commit to a specific attachment choice for ambiguous relative clauses in sentences such as (13a), unless they expect to be asked detailed questions about the sentence’s content. That is, comprehenders will understand that somebody has burned herself but may not immediately decide whether the person in question was the schoolgirl (13b) or the schoolgirl’s sister (13c).

1. 1. a. The sister of the schoolgirl who burned herself was usually very careful.

   2. b. The sister of [_DP the schoolgirl who burned herself]

   3. c. [_DP The sister [_PP of the schoolgirl] who burned herself]

Evidence for morphosyntactic underspecification comes from a reading-time study by Pickering, McElree, Frisson, Chen and Traxler (Reference Pickering, McElree, Frisson, Chen and Traxler2006), who found that readers may leave a verb phrase’s aspectual (telic vs. atelic) value unspecified initially when encountering aspectually ambiguous sentence fragments such as The insect hopped. Not committing to a specific analysis of ambiguous input may save comprehenders time and/or processing resources, and helps avoid the need for a potentially costly reanalysis later on (see also Sturt, Pickering, Scheepers and Crocker Reference Sturt, Pickering, Scheepers and Crocker2001).

In short, the examples in (12) and (13) illustrate cases of ambiguity where recovering the intended structural analysis is not crucial for gist comprehension, in communicative situations which involve what Denison (Chapter 13 in this volume) describes as ‘vagueness’.

12.4.1 Category Change: Conversion

Many English words belong to more than one syntactic category, with N-to-V conversion (e.g. to send an email – to email) being a particularly productive example. Denison (Reference Denison, Traugott and Trousdale2010a) distinguishes between category change with and without structural change. Examples of recategorisation that does not require any restructuring include cases of N-to-A conversion as in It was a fun event or He is a rubbish footballer. This kind of change might have been facilitated by examples like (14) in which the categorical status of fun is not unambiguously determined by the syntactic context, and given that nouns and adjectives show some overlap in their defining features (see, e.g., Denison Reference Denison2013, Chapter 13 in this volume).

1. 1. a. The party was [_{A / N} fun]

   2. b. Last night’s film was [_{A / N} rubbish]

From a processing perspective, there are several theoretical possibilities as to how comprehenders may deal with contextually underspecified words such as fun or rubbish in bridging contexts such as (14a, b). Normally the processor would retrieve the word’s categorical features from its lexical entry and then create a corresponding syntactic head node. Miscategorisations may occur if the processor determines an incoming word’s category label on the basis of the syntactic context (and/or probabilistic biases) alone, either failing to check or simply overriding the word’s categorial specification. Note that the interpretive consequences of choosing one contextually licensed analysis over the other seem to be fairly minimal here. An alternative possibility would be for the processor to leave the post-copular node underspecified. Whilst in cases such as (14) above, computing an underspecified or ‘good enough’ analysis seems perfectly sufficient for comprehension, underspecification at the phrase structure level would not by itself account for lexical recategorisation.

During incremental language processing, the immediate assignment of category labels to new incoming words is thought to be a prerequisite for efficient syntactic structure-building. Rapid syntactic categorisation aids processing in at least two ways: (a) by allowing for new words to be integrated into the emerging structural representation immediately, thus helping to speed up processing, and (b) by enabling the parser to make predictions about the constituents yet to come, or which are required to complete the sentence in a grammatically appropriate way. With these two points in mind, let us consider the following (largely speculative) scenario in relation to (14) above. Having encountered the auxiliary or copula be, the processor may predict that be is most likely going to be followed by an adjective here, thus generating an adjectival node in anticipation. The absence of a determiner before fun or rubbish is fully consistent with this prediction, making the anticipated analysis more likely to be maintained. If the word’s lexically specified categorical features then turn out not to match after all, the processor has two choices: either to correct the analysis by analyzing fun and rubbish as nouns instead, or to ‘attach anyway’ (to borrow a term from Fodor and Inoue Reference Fodor, Inoue, Vincenzi and Lombardo2000), that is, to force these words into the predicted adjectival slot. As a result of such a forced match, the word’s categorial features may be altered or overwritten, ultimately leading to the creation of a homophonic lexical entry with different categorial specifications.

Let us briefly switch perspective from language comprehension to language production. This is thought to involve, among other things, the creation of a syntactic frame for the intended utterance, with lexical items that match in terms of their categorical specification then being inserted into appropriate syntactic slots (e.g. Bock and Levelt Reference Bock, Levelt and Gernsbacher1994). Speakers who have successfully recategorised a particular noun as an adjective should then be able to slot this word into adjectival positions other than contextually underspecified ones as in (14), as well as into nominal positions. The comprehension-based scenario outlined above does not preclude the possibility that innovative recategorisations may originate in individual speakers’ or writers’ minds, of course, as witnessed by the fact that conversion is a popular linguistic device in poetry.⁷ Consider the following examples (15a–c) from Shakespeare (cited by Crystal Reference Crystal2004: 332–33) for illustration:

1. 1. a. Thank me no thankings, and proud me no prouds

      (Romeo and Juliet, 3.5.125)

   2. b. a hand that kings have lipped (Anthony and Cleopatra, 2.5.30)

   3. c. Thou losest here, a better where to find (King Lear, 1.1.261)

Many cases of diachronic category change, including the examples in (14) above, show signs of gradience. This raises important questions about the nature and mental representation of syntactic categories, which the results from real-time sentence processing or ambiguity resolution studies cannot currently answer. For a fuller discussion, see, e.g., Denison (Reference Denison, Traugott and Trousdale2010a, Reference Denison2013, Chapter 13 in this volume) and Trousdale (Reference Trousdale, Giacalone Ramat, Mauri and Molinelli2013). Some examples of category change with accompanying structural change will be discussed below.

12.4.2 Constituent Boundary Shift

As constituent boundaries are not always indicated clearly by prosody or punctuation marks, they can easily be misplaced during listening or reading. A shift of constituent boundaries is also sometimes seen in language change. The recategorisation of the preposition for as a complementiser during the change from Middle to Modern English, for example, presumably involved constituent rebracketing as illustrated in (16), below (compare e.g. Fischer Reference Fischer, Nixon and Honey1988; Harris and Campbell Reference Harris and Campbell1995: 62).⁸

1. 1. a. It is better [_PP for me] [_CP PRO to slay myself] than to be violated thus

   2. b. It is better [_CP for me to slay myself] than to be violated thus

From a processing perspective, the bracketing in (16b) is more attractive because it yields a structurally less complex sentence representation than (16a), which contains a prepositional phrase plus an infinitival clause lacking an overt subject. Covert or ‘understood’ subjects (indicated by PRO in 16a) are, by definition, not visible or audible, but once their existence has been inferred they must be linked to a suitable antecedent for the sentence to be fully interpretable. The need for establishing this referential dependency further increases the computational cost associated with building the representation in (16a) compared to that in (16b).

Cases of grammaticalisation such as the formation of the be-going-to future, illustrated in (17), also involve constituent boundary shifts (e.g. Hopper and Traugott Reference Hopper and Traugott2003).

1. 1. a. Mary_i [_TP [_TP is [_VP going]] [_CP PRO_i to buy some bread]]

   2. b. Mary [_TP [_T is going to] [_VP buy some bread]]

Analysing going as part of a modal (17b) rather than as a lexical verb (17a) again involves structural simplifications. As in (17a) above, the purpose clause’s understood subject is inaudible, thus facilitating its omission from the syntactic representation generated by comprehenders and corresponding rebracketting. The observation that lexical verbs can become auxiliaries or modals (but not the other way around) is consistent with comprehenders’ preference for delaying the closing off of the constituent currently being processed for as long as possible. During left-to-right processing, T(ense) and/or M(odality) nodes will be generated earlier than VP nodes in English, providing a potential home for any verb-like elements that can express tense or modality. Since the processor is keen to slot new incoming words into corresponding phrase structure positions as soon as possible, and assuming that in root clauses T and/or M nodes are created automatically, it is unlikely that the integration of a modal should be delayed and the modal slotted into a verbal head position further downstream instead.⁹

Other cases of rebracketting include the formation of complex determiners involving kind of or sort of (e.g. Brems and Davidse Reference Brems and Davidse2010; Denison Reference Denison2002, Reference Denison, van Kemenade and Los2006, Chapter 13 in this volume). Here a noun and the head of a following prepositional phrase (18a) fuse and form a complex determiner with a preceding article (18b).

1. 1. a. a rare [_NP [_N kind] [_PP of bird]]

   2. b. [_DP [_D these kind of] jokes]

In (18b), the lack of number agreement between kind and the preceding determiner indicates that kind is not a head noun here. Note that during speech perception, the original constituent boundary between kind and of may not be easy for comprehenders to discern; example (18b) is in fact reminiscent of a fairly common type of ‘slip of the ear’ error involving missing or misplaced word boundaries (Bond Reference Bond1999). Complex determiner formation, whilst creating new multi-word lexical entries (or ‘prefabs’, see Denison, Chapter 13 in this volume), again yields simpler phrase structures.

Complex prepositions such as on behalf of in (19b), which are discussed in some depth by Denison (Reference Denison, Traugott and Trousdale2010a), appear to provide a similar example of fusion:

1. 1. a. [_PP on [_N behalf] [_PP of DP]]

   2. b. [_PP [_P on behalf of] DP]

Besides resulting in less complex phrase structure representations, the changes illustrated by (16)–(19) above are all consistent with some form of the Late Closure processing principle. In (16), rather than closing off the constituent headed by for after the pronoun me, the following string to slay myself is incorporated into it. In (17), going to is incorporated into the functional head expressing tense/modality, whilst in (18) and (19), kind of and behalf of fuse with a preceding determiner or prepositional head, respectively.

12.4.3 Summary

Examining some well-known examples of syntactic change from the point of view of left-to-right incremental processing served to illustrate the potential role of underspecification and structural least-effort principles in facilitating neoanalysis. The observation that syntactic change leads to structural simplifications is by no means new (see, e.g., Roberts Reference Roberts1993; Roberts and Roussou Reference Roberts and Roussou2003), but few scholars have previously considered neoanalysis from a strictly left-to-right processing perspective (but see Lightfoot Reference Lightfoot2006 for some useful discussion). By ignoring semantic and other potentially relevant factors, we obviously get a rather one-dimensional view of neoanalysis, however. The precise extent to which misanalysis during processing and neoanalysis really do resemble one another remains to be determined. There may be cases of constituent boundary shift that do not lead to structural simplifications, and cases where semantic change demonstrably precedes structural change, both of which would require us to refine or rethink the processing-based scenarios outlined above.

13.1.1 Participants

This chapter concerns the role of incomplete knowledge in linguistic change, particularly on the part of an addressee/reader (AD/R). I will argue that the speaker/writer (SP/W) may sometimes lack complete knowledge, too. In addition, of course, the linguist rarely has complete knowledge either. What is not known may have a bearing on appropriate theories of language. I begin with a brief look at the difference between the (un)knowns of historical linguists and psycholinguists as observer-participants in acts of linguistic communication.

In historical linguistics, where written data is still the norm, a linguist R is a little different from the original readership envisaged by W: potentially more knowledgeable in some ways, whether by virtue of hindsight or specialist knowledge, but then again perhaps ignorant of cultural and pragmatic facts obvious to a contemporary reader. For recent audio broadcast or telephone data, the historical linguist is much like a normal AD, except that she or he can listen repeatedly to a sound clip. With spoken conversational data, non-verbal cues are usually lacking. In broad terms, however, the historical linguist is simulating being AD/R of an individual act of linguistic communication and – in corpus linguistics – generalising across many such acts.

A psycholinguist, on the other hand, does not resemble AD/R but is rather an intimate, external observer of the SP/W–AD/R dyad. Linguistic situations are contrived experimentally in order to get data which might not occur in normal life, or at least not with sufficient control or sufficient frequency for good statistics. Psycholinguists also look in detail at the processes involved in AD/R’s reception and sometimes SP’s (rarely W’s) production.

13.1.2 Ambiguity vs. Vagueness

I distinguish two kinds of incomplete knowledge. Ambiguity is where AD/R (and linguist) cannot be sure which of two or more linguistic possibilities was intended by SP/W, and something hangs on the choice. Vagueness is where a linguistic analysis is in some relevant respect underdetermined, at least for AD/R (perhaps for SP/W, too), but no further information is needed for interpretation. The distinction is a familiar one in lexical semantics, and, later in this chapter, I will extend it to other linguistic domains such as syntactic structure and lexeme boundaries.

What differentiates ambiguity from vagueness is whether or not SP/W could have made a choice, and furthermore, whether such a choice would have mattered. Since neither the existence nor the importance of a choice is a wholly clear-cut notion, there could in principle be some middle ground between ambiguity and vagueness; one such case is noted in Section 13.2.1.

Lexical ambiguity is easy to illustrate. It can be the outcome of divergence between senses of a polysemous word, or it may be due to accidental homonymy, as with bank ‘financial institution’ vs. ‘ground bordering a river’ (see here OED s.vv. bank n.¹ and n.³). The word bank in the second clause of (1), which directly follows a sentence mentioning the River Nile, is in principle ambiguous:

1. (1) When Ethiopia sought World Bank financing for this dam more than 20 years ago, the U.S. leaned on the bank to say no. (2013, via WebCorp)

Of course, in practice there is no real doubt in this case. Although lexical ambiguity can be striking and may sometimes be implicated in semantic change, it is not a prerequisite, whereas vagueness often is, as we will see in Section 13.2.

Vagueness is a trickier concept. It is difficult to restrict in a systematic way what is underdetermined in a given instance of vagueness. Traugott and Trousdale (Reference Traugott and Trousdale2013: 199) use a more practical definition of vagueness in lexical semantics: ‘blended, simultaneously present subcases of a more general meaning’, which can be extended to other domains.

With ambiguity there is a risk of choosing the ‘wrong’ reading. A mistake may subsequently become apparent to AD/R; for example, ‘garden-path’ sentences suggest a reading that comes to a dead end as more of the utterance is processed, in principle forcing AD/R to backtrack and explore an alternative analysis. A priori, vagueness should be less costly than ambiguity, as there is no need for AD/R to backtrack and try again. However, there is some evidence from psycholinguistics that, in practice, AD/Rs are reluctant to backtrack and may stick with the wrong reading, even at the cost of contradiction (Felser, Chapter 12 in this volume). With vagueness, there is no ‘wrong’ reading (though of course an AD can choose to ask SP to be more specific).

In subsequent sections we look at how ambiguity and vagueness play out diachronically in a number of different linguistic domains. I will argue that, in many cases, greater weight should be given to vagueness than to ambiguity as a driver of change, and consequently that tolerance of vagueness is necessary for realistic linguistic analyses.

13.2.1 Semantic Vagueness and Ambiguity

The etymology of holiday in British English is transparent. The OED senses that are relevant for our purposes are shown in Table 13.1. The difference between senses 1 and 2a can be stated in terms of inherent and inferential features (Lipka Reference Lipka and Fisiak1985), with sense 1 inherently (i.e. semantically) about religion, and ‘no work’ an inferential feature (= contextual pragmatic inference). In sense 2a the ‘no work’ component has become inherent (has been semanticised) and the ‘religious festival’ component has been lost, while in sense 2b the ‘duration of one day’ component has been lost as well (Leech Reference Leech1981). The dictionary’s comment on sense 2a, ‘[i]n early use not separable from sense 1’, implies that in late Middle English there were many bridging contexts where the defining characteristic of the semantics could be either the religious festival or the time off work, or both: the features were pragmatically equivalent. It is difficult in such a case to distinguish between ambiguity and vagueness.

Table 13.1Some senses of holiday

Consider now passenger. We concentrate on the two senses surviving into Modern English, both probably borrowed from Anglo-Norman in the Middle English period. One is ‘[a] person who passes by or through a place; a traveller, esp. a traveller on foot’. It is marked by OED as obsolete and chiefly Scottish in later use (s.v., 3a, a1450-1886). In fact it survives into the twentieth century, as (2) attests:

1. 1. a. But other passengers were approaching Lincoln meanwhile by other roads on foot. (1919 Woolf, Night & Day)

   2. b. … his quick walk along the streets and in and out of traffic and foot-passengers (1919 Woolf, Night & Day)

The other is what the OED calls ‘now the usual sense’, defined as ‘[a] person in or on a conveyance other than its driver, pilot or crew’ (s.v., 4, 1511–).¹ Both senses refer to travelers other than those driving or directing some form of conveyance (horse, carriage, bicycle, boat, etc.). In early use it was possible to have vague contexts which did not require the senses to be distinguished:

1. (3) ‘It is the Watermen that cals for passengers to goe Westward now.’ (1595, OED s.v. ho, int.¹)

There are 15 instances of passenger(s) in PPCMBE (948,895 words, 1700–1914), of which 14 concern people aboard ships, a context which virtually guarantees the ‘conveyee’ sense (though cf. (3)). The sole exception, from the play She Stoops to Conquer, is

1. 1. a. As you say, we passengers are to be taxed to pay all these fineries. (1773, PPCMBE)

The context is (so the speaker believes) an inn to which he and his companion in a post-chaise have been directed, so either sense of passenger will fit, though contextually ‘passing traveller’ (sense 3a) is more likely, and indeed the companion responds:

1. 1. b. Travellers, George, must pay in all places …

The ‘passing traveller’ (increasingly, ‘traveller on foot’) and ‘conveyee’ senses move far enough apart to permit at least theoretical ambiguity:

1. (5) The police believe they have a clue to the man who placed the nitro-glycerine which last night blew up a Euclid Avenue car and injured four passengers. (1899, COHA)

Four people in the car or four passers-by? The difference matters. The ‘conveyee’ sense is a relationship word, implicating the existence of a conveyance. Conversely, the word passenger(s) in the context of a conveyance (ship, train, car, etc.) tends to implicate the ‘conveyee’ sense, as in (5), and even more strongly when used with the definite article. The ‘passing traveller’ sense increasingly needs contextual support.

The OED, as updated in 2005, has the following note against sense 4 of passenger:

One could speculate that extralinguistic factors – the enormous growth first of public (paid-for) transport and later also of private cars – might have strengthened the ‘conveyee’ sense. Can its dominance be linked to the loss of ‘passing traveller (on foot)’? – a sense whose obsolescence by 1919 is suggested by the more explicit compound foot-passenger in (2)b. Perhaps potential ambiguity hastened the loss. Compare how the ‘homosexual’ sense of gay has made earlier senses almost unusable.

13.2.2 Hidden Ambiguity

Ambiguity need not involve senses known to both SP/W and AD/R – a choice easily resolved if noticed. My definition of ambiguity encompasses the situation where AD/R encounters an unfamiliar term and in effect makes a guess as to the meaning. AD/R’s approximation to the original sense may later enter more general usage. Strictly speaking, there is no ambiguity within the individual grammars of either SP/W or AD/R, since their patterns of usage barely overlap: ambiguity here is essentially a contact phenomenon, at least at first.

One obvious context is when a term from a technical jargon is recruited for more general use. Thus parameter is a term with exact meanings in various fields. Its sense in mathematics, illustrated in (6) below, is probably too subtle for most non-mathematicians:

In more general discourse parameter has widened to ‘[a]ny distinguishing or defining characteristic or feature, esp. one that may be measured or quantified; an element or aspect of something’, see (7), and for many speakers it has become merely a high-flown alternative to boundary or limit, as in (8) (both developments recorded in OED s.v., 8), the latter no doubt with some contamination from perimeter.

1. (6) This technique determines a new solution with one extra parameter α. (BNC, B2K 727)

1. (7) It is their responsibility to decide what new equipments [sic] are needed, and to specify their performance and other critical parameters. (BNC, ABA 205)

1. (8) This sense of collaborative work [in a drama lesson] – within agreed parameters – is a key to good control and good drama. (BNC, HYA 944)

Vagueness may play a part in the spread of the innovation in examples which blur the distinction between a technical and a more general sense:

1. (9) The precise scale you can’t quantify, but I can give you the parameters. (BNC, A44 371)

Other technical terms which have entered general usage with rather different meanings include at fault (originally of hounds which have lost the scent), crisis (moment when illness turns towards either recovery or death), quantum leap (the smallest possible change in energy level of an electron) and psychological terms like schizophrenic.

13.2.3 Pragmatic Vagueness

The closely related seventeenth-century Latin borrowings discriminate v. and discrimination n. exemplify the development of new senses in interaction between pragmatics and semantics. They have by and large kept in step with each other through later semantic changes, which I label as stages A–D in Table 13.2. The next two columns of the table have a synopsis of the relevant senses in the respective OED entries, together with the date range over which each numbered sense is recorded. (Only sense 3 of the noun is marked as obsolete.) The last column notes whether there is an element of evaluation in the connotation.

Table 13.2Some senses of discriminate/ion

Some examples follow:

1. (10) an instance of the judgment of the ancient Sculptors in their nice discrimination of character. (1821–2, CLMET3.0) [stage A]

1. (11) Man would cease to be Man: on the one side he would lose his discrimination from God, and on the other from Nature. (1867, OED) [stage A′]

1. (12) and ridiculing the public for their want of taste and discrimination in not admiring it. (1838, CLMET3.0 [stage B]

1. (13) They do not discriminate against ships belonging to the other states. (1786, OED) [stage C]

1. 1. a. More American women than ever before are filing charges that they’re being discriminated against in employment. (1970, COHA) [stage D]

   2. b. The discrimination of Arab-Americans since 9/11 has been vastly overlooked in society. (2005, COCA) [stage D, usage not yet in OED]

The earliest stage for both verb and noun, A, illustrated in (10), carries no value judgment, nor the stage A′ of (11), which is merely a statal result use of discrimination – an easy kind of sense extension for action nouns, found also in commitment, extension, finding and many others.

In the eighteenth century the focus with the noun switches to the discriminator, stage B: discrimination regarded as a human faculty. This is a metonymy, or arguably an inference/implicature: the act or fact of discrimination implies the existence of a human mind endowed with the ability to make fine distinctions. Possession of that ability need not evoke a subjective evaluation, but one can easily arise through inference/implicature, in principle either positive (e.g. the faculty of a good judge) or negative (e.g. of a fusspot). Evidently it was positive connotations that stuck, perhaps through contextual association and repeated collocation with particular items, for example premodifiers like nice or coordination with taste or delicacy; cf. (12).

A separate semantic development, stage C, focuses on the discriminatee, at first with a specialised use in trade and economics for the favoring or disfavoring of particular products or countries, as in (13). In this factual, technical usage there need be little or no subjective speaker evaluation – the sense is vague in that respect. However, an extension to human discriminatees in the nineteenth century facilitates the addition to both verb and noun of an inference/implicature ‘unjust’ or ‘unjustified’, a negative subjective evaluation, perhaps because (economic) discrimination against is much commoner than discrimination in favour of, both in practice (imposition of tariffs, or import bans, for example) and in language use. The addition of an ‘unjust’ element is favored culturally or politically through changes in mainstream social attitudes, and linguistically through repeated premodification by racial or sex or by coordination with prejudice. If ‘unjust’ becomes part of the semantics, we have reached stage D, as in (14). Collocation with immediately following against is a convenient test to distinguish stages C and D from A, A′ and B. We can see the growth of discriminat* against in COHA (400 million words, 1810–2009):

Figure 13.2 shows the growth in frequency of such premodifiers as race or sex,² essentially from the 1920s onward after sporadic attestations beginning in 1885.

Figure 13.1

Discriminat* against in COHA.

Figure 13.2

Selected premodifiers of discriminat* in COHA.

Although nearly all the senses mentioned here remain available, stage D is probably dominant in present-day use of both noun and verb, with older meanings needing more contextual support. Notice by contrast that the adjective discriminating seems to retain as its default a positive connotation similar to stage B. Other factors must determine the direction of semantic change and whether it takes place at all; pragmatic vagueness is merely a prerequisite.

A more recent example is sketchy, a denominal adjective used transparently from 1805 of writings or drawings in the sense ‘giving only a sketch or outline’ (my précis of OED s.v., 1 and 2). By 1878 a colloquial sense ‘[o]f a light, flimsy, unsubstantial or imperfect nature’ (s.v., 3) is recorded, which I take to be a metaphorical extension: different domain, senses related by similarity. A further shift turns ‘flimsy, imperfect’ into ‘unsafe, disreputable, dishonest’, presumably when applied to such concrete or abstract referents as buildings or legality. Dubiousness and illegality are not yet recognised for sketchy by the OED but are current in the US (OxfordDictionaries.com: ‘dishonest or disreputable’; Merriam-Webster: ‘likely to be bad or dangerous’; Urban Dictionary: ‘gives off a bad feeling; unsafe’), even allowing a human referent:

1. (15) it pays to be suspicious of large, seemingly useless gifts from one’s sworn enemy. And that includes your aunt’s sketchy second husband. (2011 COCA, Caprice Crane, With a little luck: a novel)

These latest shifts in semantics are enabled by the source sense being pragmatically vague as to evaluation.

In the examples discussed it is mostly vagueness that promotes the change. Ambiguity can play a part in the aftermath, but may be more centrally involved in language contact.

No attempt has been made to offer a complete typology of semantic change.

13.3.1 Stepwise Adjective < Noun

Noun and adjective are distinct word classes in English, but crucially they have some distributional properties in common. Both can occur as premodifier of a head noun, thus real adj. in (17) and inflation n. in (18):

1. (17) Gold is real money and paper is pretend money. (1974, OED)

2. (18) That’s why inflation money is false purchasing power. (1946, WebCorp)

The word class membership of these two items is securely known from behavior elsewhere. (More problematic is pretend in (17), which in other slots is more familiar as a verb.)

Nouns and adjectives also share the possibility of occurring as predicative complement, thus beautiful adj. and fakes n. in (19), and fake adj. in (20). (The latter must be an adjective, since fake does not occur as a mass noun, and only mass nouns and plurals can form a grammatical NP without determiner.)

1. (19) … a third of the pictures are beautiful, but I think two-thirds of the pictures are fakes. (BNC, EBX 1777)

1. (20) His gentleness was fake, …. (BNC, BP7 1362)

The premodifier and sometimes the predicative complement contexts can leave the word class of their filler underdetermined if that word exists both as noun and adjective. Consider expert in these examples, uncontroversially an adjective in (21) and a noun in (22):

1. (21) Naihe from Ka’u on the Big Island was so expert a surfer that his fellow chiefs grew jealous …. (BNC, ASV 62)

1. (22) An expert’s decision is usually final and binding. (BNC, J6Y 2079)

Note that the adjective goes back to Middle English, whereas the noun use can only be traced back to 1825, probably by re-borrowing from French (OED s.v.): I am not citing this as an example of historical change, merely as a clear-cut case of synchronic homonymy for which Present-Day English (PDE) speakers need lexical entries both as noun and adjective. What then is the word class of expert in (23)?

1. (23) You could do it yourself or get expert help. (BNC, A0 G 1488)

AD/R cannot know whether expert is noun or adjective here. This is vagueness, not ambiguity, since the choice makes no difference to interpretation and no difference to constituent structure. Psycholinguists might invoke the concept of ‘good enough’ processing here (Felser, Chapter 12 in this volume).

I would argue further that the SP/W of (23) need not have decided between two lexical entries for expert, although such a claim would pose difficulties for some models of language production. In the same vein, it seems to me arbitrary for a linguist to privilege one classification or the other by assigning either Adj or N to expert in (23) (and concomitantly AP or NP to its projection). If a decision is demanded by a linguistic theory which takes unique word class to be universal and basic, maybe the theory should be questioned.

Be that as it may, the kind of analytic vagueness seen in (23) is crucial in the histories of words (unlike expert) that start out as nouns but which come to show adjectival behavior as well. The premodifier and predicative slots are what we might call syntactic bridging contexts. When such a word appears there, AD/R may be prompted to regard it as an adjective even if SP/W only has a lexical entry as noun, especially with a referent that is semantically gradable.³ I consider some evidence of stepwise progression toward full adjective behavior, with examples taken from the list of words in (16).

13.3.2 Partial Word Class Change

If such words as those listed in (16) simply developed a new adjective use by conversion, and if word classes were Aristotelian categories, then all the adjectival properties mentioned in Section 13.3.1 should arrive simultaneously. But they do not: acquisition of adjective behavior is generally stepwise rather than all at once. We see this both in corpus data and in informant testing, with speakers considering some but not all adjectival properties acceptable. In Denison (Reference Denison, Traugott and Trousdale2010a: 110–11) I cited some examples of the word rubbish that were indeterminately noun or adjective, plus others that should have been impossible for a noun. What was striking was the acceptability judgments the data evoked in a group of students. The more obviously adjectival contexts (comparative and superlative forms, modification by very) scored poorly, averaging in the range 1.95–2.55 out of a maximum 5, as against 4.70–4.75 for the vague ones. This was not unexpected for a recent innovation. Yet the use of rubbish to postmodify an indefinite pronoun scored 4.6 – in other words, was found to be almost wholly acceptable. This remains to be confirmed by a proper psycholinguistic study, but I have little doubt that the result is robust and is largely independent of prescriptivist influence.

For another, similar case, the following attested examples support the categorisation of fun as adjective, but with differing acceptability:

1. (40) Doing something fun like redecorating your room … is really interesting biz for a teen who loves being busy. (1951, OED s.v. teen n.2)

1. (41) And they are so fun to eat! (1979, COHA)

1. (42) Walking and looking is boring. Touching is funner. (1990, COCA)

Most speakers I have consulted accept the postmodification of (40) but reject the comparative of (42), while the acceptability of so fun in (41) (as opposed to such fun) is intermediate and more or less inversely correlated with age. Denison (Reference Denison2013) argues that differential acceptability of new patterns and the stepwise nature of change point to grammatical variation within a population, and that corpus techniques which rely on pooled data can only give limited insight; furthermore, a writer’s date of birth may be as pertinent as date of publication. Informant testing with psycholinguistic techniques would be a better way of getting a snapshot of change in progress and teasing out the diachronic significance of word class vagueness. But of course the view of word classes espoused here has synchronic consequences for linguistic theory too.

13.4.1 Sort of/Kind of/Type of

Phrases consisting of the so-called SKT nouns with of are highly frequent strings: sort/kind/type + of occurs 46,064 times in the BNC at a rate of 468.54 per million words, equivalent to being among the 200 most frequent word tokens in the corpus.

The broad outlines of the history of SKT were sketched in Denison (Reference Denison2002), though subsequent work has refined certain parts of the picture and more data are available, while the fullest published examination of their syntax in PDE – within the NP at least – has been by Keizer (Reference Keizer2007), and of their diachronic semantics by Brems and Davidse (Reference Brems and Davidse2010).⁵ Among the uses which have been identified are

1. (44) … the Canadians had one sort of sovereign, and the British had another sort. (BNC, A69 1471) [referential/binominal]

1. (45) There was a kind of inevitability about the whole proposal which appalled Alexei. (BNC, G17 1172) [qualifying]

1. (46) It kind of built his confidence with each successive flask. (BNC, A14 937) [adverbial]

1. (47) … but she should keep those sort of remarks to herself. (BNC, CDY 1447) [postdeterminer/complex determiner]

1. (48) They are used to all sorts of emergencies, … (BNC A2X 404) [quantifying]

1. (49) It was a grim sort of place, … (BNC, A2J 4) [descriptive modifier]

Now, even Huddleston and Pullum (Reference Huddleston and Pullum2002: 621) recognise reanalysis in the adverbial (46) type, where of can hardly be called a preposition any more. As shown in Denison (Reference Denison2002), modified by Brems and Davidse (Reference Brems and Davidse2010), it is possible to reconstruct the developmental paths of the various idiomatic uses of sort of and kind of, and in each case there are intermediate examples which are either vague or ambiguous. Here I give just three, the last involving not just the prefab kind of but also the newer prefab kind of thing, which has been developing into a discourse particle:

1. (50) ‘Don’t worry, a bit of body-popping won’t kill me.’

   ‘What on earth is body-popping?’

   ‘It’s a sort of dancing. Why, what did you think it was?’ (BNC, A0 F 2702–5)

1. (51) I Answered, That Religion being a design to recover and save Mankind, was to be so opened as to awaken and work upon all sorts of people, and generally men of a simplicity of Mind, were those that were the fittest Objects. (PPCEME, Burnet 1680)

1. (52) It created this mushroom kind of thing that people stared at. (BNC, AB5 1270)

In (50), sort can be head in the binominal construction (‘one variety of dancing’), or dancing is head in the qualifying construction (‘something roughly answering to dancing’). The semantic difference is slight but identifiable, so this is probably best viewed as ambiguity. Example (51) is taken from Brems and Davidse (Reference Brems and Davidse2010: 188), where it is suggested that a binominal use of sort quantified by all offers an invited inference of a great number, which can then be semanticised in later examples as the quantifying meaning of all sorts of. In (52), either mushroom is a modifier and kind of thing is nominal, or mushroom closes the NP, in which case kind of thing becomes an adverbial hedge. Structurally this must be a case of ambiguity, though pragmatically it makes little difference.

13.4.2 Piece of Work

A famous line in Hamlet is probably indirectly responsible for a modern application of the phrase piece of work to human referents:

1. (53) What a piece of worke is a man! (1623, Hamlet First Folio II.ii)

The definition the OED gives (s.v. piece n. P4, c.) is:

In the BNC, 12 out of 190 examples of the string piece of work have a human referent. One is the Shakespearean quotation of (53) in modern spelling, one is (54), and the remaining 10 are all preceded by the adjective nasty, as in (55):

1. (54) ‘You’re some piece of work, Mrs Sutherland, you know that?’ (BNC, FPF 1142)

1. 1. a. You’d best steer clear of him, Manderley, he’s a nasty piece of work. (BNC, HJC 1764)

   2. b. He was also a member of the Mafia, and he was up to his eyeballs in drugs. Altogether a very nasty piece of work. (BNC, GV6 3045)

1. (56) It was a nasty piece of work, done with thoroughly malicious intent. (BNC, K4W 477)

(Example (56) is the only time nasty piece of work is used of a non-human referent.)

Clearly, then, we have a prefab, somewhat lexically and semantically restricted, which plays its part in production and reception. However, the precise extent of the prefab is vague. It might, for example, be any of:

1. 1. a. piece of work

   2. b. a piece of work

   3. c. a ([optional intensifier]) [pejorative adjective] piece of work

   4. d. a (…) nasty piece of work

Any decision is going to be somewhat arbitrary. Such vagueness as to the boundaries and the fixedness of an idiom, though no hindrance to communication, is problematic for purely algorithmic theories of language use where a specific lexical item has to be mapped against an interpretive component.

13.4.3 The Process of Chunking

What we see here is prefabs of different degrees of schematicity, whose parts may be inflected or interrupted (Bybee Reference Bybee, Hoffmann and Trousdale2013: 54) – and interrupting elements are freely chosen in some cases, but in others tend to come from a restricted set (as in Section 13.4.2). This supports a usage-based model of language involving emergent grammar, where both prefab and word-by-word analyses play a part, but to different degrees, and the proportions change over time. That, however, is a statistical overview. What of the individual utterance? As soon as we turn our conscious attention to the wording, we are pushed toward recognising syntactic ambiguity, as however potentially present the two (or more) analyses may be, it is hard to conceive of them as simultaneously active (cf. non-linguistic perceptual ambiguities like Rubin’s vase or the Boring figure). What actually happens in the routine, unthinking speech situation is a question for psycholinguists to resolve. Hilpert (Chapter 3 in this volume: Section 3.6) notes that ‘the linguistic competence of speakers must include probabilistic knowledge of variation’. Speakers are certainly (unconsciously) aware of alternative analyses of a potential prefab, and a speculative extension of Hilpert’s observation would suggest that the relative strengths of the variant analyses of a given string might also form part of speakers’ competence.

Structural ambiguity is not incompatible with semantic or pragmatic vagueness. Such vagueness can contribute to the ill-definedness of chunking in a given utterance and promote the gradual formation of prefabs.

13.5.1 Prepositional Passive by Reanalysis

The prepositional passive begins to appear around 1200 (Denison Reference Denison1985, 124–7, Reference Denison1993: 140–43). Fischer and van der Wurff (Reference Fischer, van der Wurff, Hogg and Denison2006: 196–7) explain its advent as follows. A general change in word order culminating in Middle English turned English from a V-2/V-final language into what is sometimes called a V-3 or SVO language, after which a PP collocated with an intransitive verb would almost always follow the verb directly, with verb and preposition adjacent.⁶ This routine adjacency enabled a reanalysis to take place:

1. (58) V [_PP P NP] > [_V V + P] NP

Thus, for example, probably through frequent euphemistic use, lie by ‘lie beside’ can develop the meaning ‘have sexual intercourse with’, (59). The semantic change and concomitant chunking of verb and preposition allows reanalysis (58) to go ahead, weakening or destroying the unity of the PP. Now the string lie by behaves as if the following NP is complement not of the preposition alone but of the new transitive composite verb, and so a passive can be formed (60).

1. (59) Vulcanus..foond thee lyggyng by his wyf allas. (c1385, MED)

   Vulcan … found you ‘lying by’ his wife alas

1. (60) Þis maiden..feled al so bi her þi Þat sche was yleyen bi. (c1330(?a1300), MED)⁷

   this maiden … felt also by her thigh that she had-been ‘lain by’

A similar process explains the early advent of passives of such prepositional verbs as fare with ‘deal with, treat’, send after ‘summon’, speak of ‘mention’ and so on.

I have made a case for the new construction to have spread at first by lexical diffusion among closely related verb-preposition combinations (Denison Reference Denison1993: 141). More generally, Fischer and van der Wurff (Reference Fischer, van der Wurff, Hogg and Denison2006: 197) observe that the semantics of prepositional verb combinations has to allow an ‘object interpretation’ of the complement NP. Typical semantic roles for such NPs are Stimulus or Cause with experiential verbs and Patient or Goal with others (Denison Reference Denison1993: 140–1). Dreschler (Reference Dreschler2015: 120–25) advocates syntactic reanalysis but makes a case for its acceptability passing almost unnoticed via minimal alteration of several existing structures.

13.5.2 Prepositional > Phrasal Verb by Reanalysis

In one context the polysemous prepositional verb run over undergoes an interesting semantic and syntactic change in late Modern English. There are numerous examples in the Old Bailey Corpus (14 million words, 1720–1913) where the literal sense of intransitive run involving rapid movement of a person or (part of) a vehicle collocates with an over-phrase to indicate path (61), and for quite a few of these the context is a (potential) collision (62):

1. (61) … we pursue’d him over the Downs, and towards a Wooden Bridge at the Bottom of the Downs …. He ran over the Bridge. (The Old Bailey Corpus 1720–1913 (version 1.0, 2013–06-04), 1739, t17390718-10)

1. 1. a. it was bent, and look’d as if a Coach-wheel had ran over it. (OBC, 1732, t17320114-38)

   2. b. I saw a Hansom cab standing at No. 21; I saw it go past our door; the prisoner was driving it – it was going very fast – it knocked the men down and ran over them. (OBC, 1878, t-18780311–341)

Now consider a collision example like (62) made passive – also frequent:

1. (63) the man was run over on the legs […] it was through the furious driving – he was knocked down by either the bar or the van, I could not take my oath which. (OBC, 1861, t-18610408–325)

1. 1. a. a young woman with a child in her arms endeavour’d to stop the horses; I called to her to let them go, as I saw she would be run over else. (OBC, 1770, t17700711-39)

   2. b. the prisoner came up and told me I was not to go too fast, for if I did I should get run over – she told me to wait till all the carriages and horses were gone by. (OBC, 1849, t-18490226–699)

There is a potential structural and semantic ambiguity in (64) (but not (63)): either over is a preposition referring to the trajectory of the moving vehicle across and above an obstacle, or it is a resultative adverbial particle describing the trajectory of the victim out of upright position, part of a transitive phrasal verb run over ‘injure with a vehicle’, like knock down. In general, Patient or Goal is a typical semantic role for the subject of a prepositional passive (Section 13.5.1), and that association would encourage the innovative reading of (64), where the Patient role is even more marked. Note that another potential source of ambiguity does not seem to be much in evidence in the period of reanalysis, namely active VPs with a non-pronominal NP: active run over NP could in principle be ambiguous between prepositional and phrasal verb, but I have found no convincing examples in OBC, only in later corpora.

After reanalysis of the passive to the phrasal verb structure, new patterns become possible:

1. 1. a. She […] got on her bike and roared off. My father tried to stop her by standing in her way, so she ran him over and broke his leg quite badly. (BNC, HWC 2340)

   2. b. Someone’s going to run a little child over soon because the lollipop lady is busy asking drivers to move on. (BNC, K55 9072)

Since over follows the object NP in (65), it must be an adverbial particle rather than a preposition.

The period of transition seems to be largely in the second half of the twentieth century. I could not find any clear examples of active run over in the required sense as a phrasal verb in COHA until 1949, 1955 and later.⁸ Interestingly, the linguistic change is observed around then by Wood (Reference Wood1955–6: 22; see also Parker Reference Parker1976: 451–2). However, the OED has managed to trace the pattern as far back as 1860 (s.v. run v. PV1, 1). A similar ambivalence between prepositional and phrasal verb analysis, but often with unclear semantic difference or directionality of change, is seen in verbs like pass, read in combination with path particles like by, over, through and (less often) about, round.

13.5.3 Reformulation without Reanalysis

Is reanalysis important in language change? Indeed, is it real? To some extent the questions are terminological, but only to some extent. If most reanalysis scenarios can be repackaged convincingly as examples, say, of analogy, there are real consequences. Directionality and context of change may be better explained. SP/W becomes as important as AD/R in such changes – or more important, even. And change is not so limited to the life stage of language acquisition; see here Lieven’s argument (Chapter 14 in this volume) against child errors as a source of change, and also López Couso (Chapter 15 in this volume).

Consider the examples discussed above. For the prepositional passive, two crucial ingredients of the new structure are already in place before it appears, namely passive verb participles and stranded prepositions (the latter in relative and infinitive clauses, for instance). Furthermore, some of the actual lexical verbs that appear in the prepositional passive early on are (more) often used transitively and therefore already have conventional passives: do, let, send, set, tell, perhaps tend. Therefore, the new kind of passive could have been formed analogically rather than – or as well as – by reanalysis in the active. Its structure is vague, related as it is at one and the same time to its equivalent in the active voice, to other kinds of stranding pattern and to other kinds of passive; cf. here the notion of ‘serial relationship’ (Quirk Reference Quirk1965), and now also Dreschler’s (Reference Dreschler2015) rather different take on multiple resemblances to existing structures.

Turning to the replacement of the prepositional verb run over by a phrasal verb, the new structure is well established in the language for other phrasal verbs (including many with over) long before the change in question. In the context of vehicle collisions with pedestrians, it is not clear how much of a semantic difference there is between the two passive structures, so that there is at least a case to be made for semantic vagueness rather than ambiguity in the construction as a whole (though not for the semantics and syntax of over).

It looks as if the apparently reanalysed structures, or something closely similar, had a prior existence with other exponents or in other contexts. Such extension of an existing structure to new material shows innovation ‘sneaking in’ where least salient. In my case studies there is a clear semantic or pragmatic component. If this is a precursor of the structural change, then the structural change (or changes, if analysed in terms of micro-steps) would seem to arise from vagueness rather than ambiguity.