The Retreat from Locative Overgeneralisation Errors

1

The retreat from locative overgeneralisation errors:

A novel verb grammaticality judgment study

Amy Bidgood

Ben Ambridge

Julian M. Pine

Caroline F. Rowland

University of Liverpool

Address for correspondence: Amy Bidgood, Department of Psychological Sciences, Institute of Psychology, Health & Society, University of Liverpool, Eleanor Rathbone Building, Bedford St South, Liverpool, L69 7ZA, UK. Email: .

Abstract

Whilst some locative verbs alternate between the ground- and figure-locative constructions (e.g. Lisa sprayed the flowers with water/Lisa sprayed water onto the flowers), others are restricted to one construction or the other (e.g. *Lisa filled water into the cup/*Lisa poured the cup with water). The present study investigated two proposals for how learners (aged 5-6, 9-10 and adults) acquire this restriction, using a novel-verb-learning grammaticality-judgment paradigm. In support of the semantic verb class hypothesis,participants in all age groups used the semantic properties of novel verbs to determine the locative constructions (ground/figure/both) in which they could and could not appear. In support of the frequency hypothesis, participants' tolerance of overgeneralisation errors decreased with each increasing level of verb frequency (novel/low/high). We conclude by outlining a possible integrated accountof the roles of semantics and frequency in the retreat from overgeneralisation.

The retreat from locative overgeneralisation errors:

A novel verb grammaticality judgment study

As adults, we have the capacity for limitlesscreativity in language production: we are able to produce a potentially infinite number of utterances that have never been produced before. To reach this stage, children must acquire the grammar of the ambient language by forming generalisations about that language from the input. However, children must also learn to restrict these generalisations in order to avoid producing ungrammatical utterances (e.g. *I don’t want it because I spilled it of orange juice [= I spilled orange juice onto my toast], Bowerman, 1981).

Pinker (1984, 1989) listed various grammatical constructions that have two alternating forms. The locative construction, for example, alternates between the ground- (or container-) locative, as inThe farmer loaded the wagon with hay, and the figure- (or contents-) locative, as in The farmer loaded hay into the wagon. In the first sentence, the wagon is most affected, as it is changing state from empty to full. In the second sentence, it is the hay that is most affected, as it is being moved to a specific location; the wagon may or may not end up full. Pinker (1989: 79) described this change in how the event is construed as a “gestalt shift”.

When children hear verbs used in both the ground- and figure-locative constructions (load, spray, stuff, etc.), they may create a generalisation that any verb used in one of these constructions can also be used in the other, and this works well for some verbs.A child hearingYou splashed me with water, a ground-locative construction, might generalise to the figure-locative construction to produce the grammatical utterance,You splashed water onto me. However, some English verbs, such as fill and cover, can only be used in the ground-locative construction (ground-only verbs) and generalising these verbs to the figure-locative construction would produce an ungrammatical utterance, such as *We filled toys into the box. Conversely, some verbs, such as pour and spill, can only be used in the figure-locative construction (figure-only verbs). Generalising these verbs to the ground-locative construction would similarly produce overgeneralisation errors, such as *Daddy poured my cup with juice.

One factor thatcould contribute to theretreat from overgeneralisation errors is parental feedback: so-called ‘negative evidence’. It is undoubtedly the case that some parents do provide feedback on errors that their children make, either through direct correction (e.g. C: *I filled mud intothe hole, M: No, say“I filled the hole with mud”) orimplicitly, via rephrasing (e.g. M: That’s right, you filled the hole with mud), facial expressions, misunderstandings, requests for clarification, etc.Whilst evidence suggests that such feedback is indeed helpful (e.g. Chouinard & Clark, 2003), children are unlikely toreceive sufficient feedback of this type to account entirely for their retreat from overgeneralisation errors, particularly for low-frequency verbs.Furthermore, some examples of parent-child interactionssuggest that such feedback may have only a limited effect on children’s language production (for reviews, see e.g. Marcus, 1993; Pinker, 1989: 9-14).

The current paper investigates the ability of two mechanisms to constitute a solution to the ‘no negative evidence’ problem (Bowerman, 1988) and therefore to explain the retreat from overgeneralisation with locative constructions.The first of these is Pinker’s (1989) semantic verb class hypothesis: while evidence exists in support of this account, previous studies have primarily focussed on errors involving the transitive-causative and dative constructions, which, for reasons outlined in the following section, do not constitute as strong a test of the hypothesis. The second mechanism is that of statistical learning, such as entrenchment (Braine & Brooks, 1995) and preemption (Goldberg, 1995) hypotheses. Again, the locative alternation is a particularly good test of these, as detailed below.

The semantic verb class hypothesis

Pinker’s (1989) semantic verb class hypothesis aims to explain overgeneralisation errors, and children’s retreat from making such errors, through their developing knowledge of verb semantics. Initially, children group verbs into broad conflation classes based on their semantics. Linking rules are then used to spell out the argument structure of sentences using these verbs. For example, Pinker (1989: 79) states that the ‘thematic core’ of figure-locative verbs (e.g. pour)is “X causes Y to move into/onto Z”. In this argument structure, Y (the theme) is most affected by the event. In contrast, the thematic core of ground-locative verbs (e.g. fill) is “X causes Z to change state by means of moving Y into/onto it”. Here, Z (the theme in this new structure) is most affected by the event and, indeed, must have changed state in some definable way once the action has been completed.

At some point, children realise that some verbs, such as spray and load, occur in both of these broad classes. Pinker assumed that there are separate entries in the lexicon for each sense of a verb(1989: 71-72). This assumption is based on the principles of Lexical Functional Grammar (e.g. Bresnan, 1982a, b; Kaplan and Bresnan, 1982); see Pinker, 1984:17. At this point, children set up a broad-range rule that links the two entries for a verb and allows the argument structure of one to be applied to the other. The broad-range rule enables children to hypothesise that, having heard verbs such as load and spray used in both the ground-locative construction and the figure-locative construction, verbs such as splash, fill and pour may also be used in both constructions, even if these verbs have only been attestedin one. So, whilst this rule allows children to generalise from one construction to the other appropriately for some verbs (like splash), until children are able to restrict the rule to only the appropriate verbs, overgeneralisation errors will occurwith verbs such as fill and pour.

Pinker notes that the fine-grained semantic distinctions between verbs that can and cannot alternate from one locative structure to the other often seem arbitrary at first glance. For example, why can sprayalternate but not dribble? In both cases the surface on which the liquid lands is clearly affected in some way. In order to solve this problem, Pinker proposes that verbs that are very similar on grammatically-relevant semantic dimensions are grouped together in narrow conflation classes, or subclasses. Only some of these subclasses have semantics that are consistent with both senses of the locative construction and are therefore permitted to alternate. For example, the reason that spray-type verbs can appear in the ground-locative construction whilst dribble-type verbs cannot (c.f., I sprayed/*dribbled it with water) is that only the former typically denote events where the surface ends up completely covered, and hence completely affected (the claim is that I sprayed it with water is only completely felicitous as a description of an event in which a surface is completely covered with water).

Pinker (1989: 126-127) specifies 15 narrow subclasses for locatives and allocates each of 146 verbs to one subclass (with two exceptions,wrap and string, which may each be the only members of their own respective subclasses). The defining semantics of each subclass specify whether the verbs contained within it can alternate between constructions, via a narrow-range rule, althougheven alternating classeshave a bias towards one of the two constructions. Table 1 (adapted from Ambridge, Pine & Rowland, 2012, 262, based on Pinker, 1989, 126-127) detailsthe 15 subclasses.

INSERT TABLE 1 ABOUT HERE

It is only as children’s knowledge of verb semantics continues to develop that they become aware of the fine-grained distinctions between the subclasses and are able to set up the more specific narrow-range rules, which will lead to the retreat from overgeneralisationerrors. Once the necessary narrow-range rules are in place, verbs in each of the subclasses will be used in a particular construction only if another verb in its subclass has already been attested in that construction. When children learn a new verb, such as coat, they will use its semantics to assign it to the appropriate subclass, here the cover-type container-only subclass, in which “a layer completely covers a surface” (Pinker, 1989: 127) along with verbs they already know, such as cover and fill.

Since Pinker (1989), further work on semantic verb classes has been conducted by numerous researchers. Levin (1993) explored the idea of dividing verbs into classes based firstly on the alternations in which they do or do not participate, and then furthersubdividing these classes based on the verbs’ semantic properties. Rappaport Hovav and Levin (2003; see also Levin & Rappaport Hovav, 2004; Talmy, 2000) discuss the nature of argument structure alternations in terms of event structure, and classify verbs on this basis: verbs with a simple event structurecan alternate, whereas those with a complex event structure, such as result verbs, cannot.

More recently, both Levin (2010) and Boas (2008) have discussed the importance of different ‘grain sizes’ of verb classes, positing that three levels may be required to explain the behaviour of verbs in different argument structures, with the way in which verbs are classified being motivated differently from the way they were classified in earlier work. Coarse-grained classes arise from verbs’ different ontological types (e.g. verbs of manner [pour], substance [oil] or result [fill]), medium-grained subclasses are motivated by semantic differences within a particular ontological type, such as different types of manner verbs: manner of motion (swim), manner of speech (whisper), manner of contact (stroke, punch) and the finest-grained subclasses are posited to relate to agency or direct causation: it is possible to bang a drum directly, but not to *whistle a kettle, since the whistling is caused by the water boiling, expanding as it turns to steam, and forcing the steam and hot air through the whistle, not directly by the person who put the kettle on the stove.

Levin (2009) also discusses work by Atkins, Kegl and Levin(1988) on the possibility of verbs being members of multiple verb classes (although she concludes that most verbs are likely to be primarily members of a single, fine-grained class that simply allows more than one argument structure, so members of these classes are alternating verbs). However, none of these modifications change the basic prediction of the semantic verb class hypothesis (to be tested here) that children’s use and retreat from overgeneralisation errors will be predicted by their knowledge of the semantic class of the verb. It should also be noted that, in the present study, all of the verbs chosen were classified in the same way by both Pinker (1989) and Levin (1993).

The organisation of verbs into classes of the kind proposed by researchers such as Pinker and Levin is not universally accepted. For example, Fellbaum (1990; Miller & Fellbaum, 1991) proposes a different method of classifying verbs according to their semantics, taking a hierarchical approach in whichverbs are organised into a taxonomy.Other researchers have also questioned the nature of verb classes (see Bowerman & Brown, 2008:7-13, for a summary). Goldberg (1995: 125-140) reframes the motivation for fine-grained verb classes as being related to construction semantics, as opposed to whether or not verbs take part in alternations and Braine and Brooks (1995) argue that the small size and large number of subclasses, along with a number of exceptions to the rules, and the apparent high degree of semantic overlap between some subclasses, make it implausible that children would be able to acquire the correct restrictions on argument structure via Pinker’s proposed mechanism (see also Bowerman, 1988; Bowerman & Croft, 2008: 296-299; Ingham, 1992;Randall, 1990). Brinkmann (1997) raises doubt about how well English verbs fit the criteria proposed by Pinker for locative verb classes, and provides evidence from German that these criteria cannot be universal (with particular reference to object affectedness). However, the strength of these criticisms is considerably weakened by the fact that semantic verb class effects do seem evident in children and adult’s use of, and grammaticality judgments of, overgeneralisation errors (e.g. Ambridge et al., 2008). Thus, the semantic verb class hypothesis enjoys empirical support.

A more damaging criticism is the fact that the hypothesis cannot explain verb frequency effects,whichare also pervasive in the literature (as reviewed below). Indeed, some authors (e.g. Stefanowitsch, 2008) have argued that apparent semantic verb class effects are epiphenomenal, with learners acquiring verbs' argument structure restrictions solely on the basis of surface-based statistical learning mechanisms such as entrenchment and preemption. It is to these mechanisms that we now turn.

The frequency hypothesis

Various accounts have attempted to explain how children are able to learn which verbs can be used in which constructions based on statistical properties of the input (e.g. Clark, 1988; Naigles & Hoff-Ginsberg, 1998). For example, the entrenchment hypothesis (e.g. Braine & Brooks, 1995; Theakston, 2004; Ambridge, Pine, Rowland & Young, 2008) proposesthat, although children may be aware that it is possible to use certain verbs in two alternating constructions, such as the ground- and figure-locative constructions, they gradually learn that this is not the case for all verbs. While children hear figure-only verbs, such as pour, frequently in their input, they never hearthem in the ground-locative construction. Eventually, this leads children to infer that, if it were possible to use pour in this construction, they “would have heard it by now”, and hence that ground-locative uses of this verb are ungrammatical for adult speakers. A similar statistical-learning account, preemption (e.g. Goldberg, 1995,2006, 2011; Boyd & Goldberg, 2011), proposes that only uses of the verb in similar discourse contexts lead to the inference that the non-attested form is ungrammatical. For example, utterances such as She poured water into the cup would pre-empt *She poured the cup with water, but other semantically more distant uses (e.g., It's pouring with rain) would not (or, at least, would do so to a lesser degree).

Ambridge, Pine and Rowland (2012) attempted to distinguish between the effects of entrenchment and preemption on the retreat from overgeneralisation in the locative construction, suggesting that both may play a role.However, their entrenchment and preemption predictors were highly correlated, which made it difficult to distinguish effects of one from the other (see also Boyd, Ackerman & Kutas, 2012).For this reason, differentiating between entrenchment and preemption is beyond the scope of the present study (see also e.g. Wonnacott, 2011:2; Perfors, Tenenbaum & Wonnacott, 2011:612). For the remainder of this paper, we will therefore simply refer to the ‘frequency hypothesis’.Our findings and conclusions could apply equally to the entrenchment and preemption hypotheses.

Existing evidence for the two accounts

Previous studies have provided evidence in support of both the semantic verb class hypothesis and statistical learning accounts. However, these have primarily been restricted to overgeneralisation errors relating to the causative alternation (e.g. Ambridge Pine, Rowland & Young, 2008; Ambridge, Pine, Rowland, Jones & Clark, 2009; Ambridge, Pine & Rowland, 2011; Brooks & Tomasello, 1999; Brooks, Tomasello, Dodson & Lewis, 1999; Naigles, Fowler & Helm, 1992; Naigles & Lehrer, 2002).In the causative alternation, related ideas can be expressed by either the transitive causative construction (e.g. Homer broke the plate) or the intransitive inchoative construction (e.g. The plate broke).

Ambridge and colleagues (Ambridge, Pine, Rowland & Young, 2008; Ambridge, Pine & Rowland, 2011) investigated the mechanisms behind the retreat from causative overgeneralisation errors using a grammaticality judgment task. Participants (children aged 5-6 and 9-10, and adults) were asked to judge both transitive and intransitive sentences containing high- and low-frequency familiar verbs, some of which were ungrammatical in the transitive construction, for grammaticalacceptability. In both studies, participants were also taught novel verbs that were semantically consistent with existing verb classes (taken from Pinker, 1989), some of which were grammatical only in intransitive sentences and some of which could alternate between the two constructions. In both studies, grammaticality judgments from both children and adults for sentences containing novel verbs were consistent with the predictions of the semantic verb class hypothesis. Both studies also found support for the frequency hypothesis, as the rated acceptability of overgeneralisation errors was higher for low-frequency than high-frequency verbs.

Naigles, Fowler and Helm (1992) conducted an act-out study thatalso provided some support for a frequency-based account (see also Naigles & Lehrer, 2002, for a cross-linguistic perspective). Children (from age 5) and adults were asked to act out sentences using familiar verbs presented in various frames, some of which were ungrammatical (e.g. *The tiger goes the lion).The authors investigated whether participants would act out a scene that was Frame Compliant (in line with the syntax, so a causative action in this case, e.g. the tiger making the lion go) orVerb Compliant (in line with verb semantics, e.g. the tiger and the lion going together). If act-outs are Frame Compliant, this implies an acceptance of the overgeneralisation by the participant. The authors found that there was a general shift from Frame Compliance to Verb Compliance with increasing age but, more interestingly for our purposes, the rate at which this shift took place varied not only across frames but also across verbs. The shift to Verb Compliance happened earlier for the frequent verbs come and gothan for less-frequent verbs, such as stay and fall. Thus it could be that, the more frequent a particular verb is in the input, the less likely the semantics of an ungrammatical frame are to over-ride the semantics of that verb. Results from production studies have also shown support for the role of frequency in constraining transitive causative/intransitive verb argument structure generalisations (e.g. Brooks & Tomasello, 1999; Brooks & Zizak, 2002).