Simulating the temporal reference of Dutch and English Root Infinitives

Hoekstra & Hyams (1998) claim that the overwhelming majority of Dutch children’s Root Infinitives (RIs) are used to refer to modal (not realised) events, whereas in English speaking children, the temporal reference of RIs is free. Hoekstra & Hyams attribute this difference to qualitative differences in how temporal reference is carried by the Dutch infinitive and the English bare form. Ingram & Thompson (1996) advocate an input-driven account of this difference and suggest that the modal reading of German (and Dutch) RIs is caused by the fact that infinitive forms are predominantly used in modal contexts. This paper investigates whether an input-driven account can explain the differential reading of RIs in Dutch and English. To this end, corpora of English and Dutch Child Directed Speech were fed through MOSAIC, a computational model that has already been used to simulate the basic Optional Infinitive phenomenon. Infinitive forms in the input were tagged for modal or non-modal reference based on the sentential context in which they appeared. The output of the model was compared to the results of corpus studies and recent experimental data which call into question the strict distinction between Dutch and English advocated by Hoekstra & Hyams. Root Infinitives in Child Language A striking feature of the speech of children who are acquiring their native language, is that, in many languages, children go through a stage where they produce Root Infinitives (non-finite verb forms in contexts that require a finite verb form). Thus, English-speaking children may produce utterances such as (1), and Dutch children may produce utterances such as (2). (1) Daddy drink coffee (2) Papa koffie drinken (Daddy coffee drink-inf) This phenomenon has been subject to considerable linguistic theorizing, as the fact that it occurs in several languages (including English, Dutch, Swedish, German and French) suggests the operation of invariant principles. A particularly influential Nativist theory has been provided by Wexler (1994). According to Wexler’s Optional Infinitive (OI) Hypothesis, by the time children begin to produce multiword speech, they have already correctly set all the basic inflectional and clause structure parameters of their language. They thus have adult-like knowledge of the word order and inflectional properties of the language they are learning. However, there is a stage of development (the Optional Infinitive stage), during which the abstract features of Tense (TNS) and Agreement (AGR) can be absent from the underlying representation of the sentence. This results in children initially using both finite and non-finite verb forms in contexts in which a finite form would be obligatory in the adult language. The great strength of the Optional Infinitive Hypothesis is that it explains the data from a wide variety of languages, as well as the relative sparseness of other errors. However, the theory also has some important weaknesses. Firstly, the theory assumes a large amount of innate knowledge and ignores the possibility that the Optional Infinitive phenomenon may be understood as the result of an input-driven learning process without the need to assume large amounts of innate knowledge. Simulations with the MOSAIC model have already shown that a simple learning mechanism which is sensitive to the distributional characteristics of the input can give a close quantitative fit to the prevalence of the Root Infinitives in English and Dutch over a range of MLUs (Freudenthal, Pine & Gobet (2002a, 2003, in preparation). Secondly, while the Optional Infinitive phenomenon occurs in several languages, cross-linguistic differences exist in the finer detail of the phenomenon that are problematic for Wexler’s theory. One obvious way of explaining these differences is in terms of differences in the distributional characteristics of the language being learned. In this paper we assess the viability of such an explanation by simulating cross-linguistic differences in the fine detail of the OI-phenomenon in Dutch and English using MOSAIC. The central aim of this paper is therefore to investigate whether the same mechanism that captures one of the key similarities in the speech of children learning different languages can also capture differences in the way that this phenomenon patterns as a function of differences in the languages being learned, and hence can provide a unified account of patterns of cross-linguistic similarity and difference in children’s early multi-word speech. The Modal Reference of Root Infinitives The majority of Root Infinitives that Dutch children produce carry a modal meaning: they tend to express desires and wishes, or relate to unrealized events. Hoekstra & Hyams (1998) have dubbed this the Modal Reference Effect. The type of verbs that occur as Root Infinitives also differs from inflected verbs. Dutch speaking children appear to use Root Infinitives when referring to actions rather than static situations. This has been called the Eventivity Constraint. Wijnen (1996), analysed the speech of four Dutch children, and found that 95% of the children’s Root Infinitives contained eventive verbs, and 85% of the Root Infinitives had a modal reference, thus confirming the Modal Reference Effect and Eventivity Constraint. According to Hoekstra & Hyams, the Modal Reference Effect and Eventivity Constraint do not hold for English. They present data based on an (unpublished) paper by Ud Deen (1997), who found that only 13% of English Root Infinitives carry a modal meaning. Ud Deen also found that, while the majority of English Root Infinitives are eventive in nature, this effect is less pronounced than it is in Dutch, with 75% of English RIs containing eventive verbs. Hoekstra & Hyams explain this cross-linguistic difference by referring to differences between the English and Dutch infinitive form. The English infinitive, they claim, is not a true infinitive, but a ‘bare form’. Dutch has a true infinitive as it has an infinitival morpheme. This infinitival morpheme is thought to carry an irrealis feature which is responsible for the modal reference. This, they argue, is evident from the analysis of the following utterances: 3. I see John cross the street* 4. I saw John cross the street 5. I see John crossing the street Utterance (3) is ungrammatical in English, because the English bare form denotes ‘not only the processual part of the event, but includes the completion of that event’ (Hoekstra & Hyams 1998, p. 105). A correct description of an ongoing event in English would therefore require the use of the past tense as in (4), or the progressive as in (5). Sentence 6 makes it clear that this constraint does not operate in Dutch: an ongoing event may be described using a present tense construction. Apparently, the Dutch infinitive does not signal completion of the event. 6. Ik zie/zag Jan de straat oversteken I see/saw John the street cross-INF I see/saw John cross the street. This difference between the English and Dutch infinitival form also explains the difference with respect to the eventivity of Root Infinitives, as, according to Hoekstra & Hyams it is the modal reading of Dutch Root Infinitives that forces the selection of an eventive verb. Since English Root Infinitives are not exclusively modal, they can occur with stative as well as eventive verbs. Problems with Hoekstra & Hyams’ Account While the Hoekstra & Hyams’ account explains the differential reading of Dutch and English Root Infinitives, it predicts that the proportion of modal readings of RIs in Dutch and English is radically different. Theoretically, all RIs in languages with an infinitival morpheme should be modal, while the reference of English RIs is free. The proportion of modal RIs in Dutch and German appears to be considerably lower than 1.00 however. Wijnen (1996) reports a proportion of .85 averaged over 4 children, and Ingram & Thompson (1996) report a proportion of .55 using a strict criterion and .79 using a lenient criterion. Ingram & Thompson also suggest that the modal reading of RIs in German (and Dutch), is caused by the fact that infinitive forms in adult German and Dutch are typically used in conjunction with a modal, as in (7) and (8). Since Dutch-speaking children predominantly hear infinitive forms in modal contexts in the input, they come to associate these forms with the modal reading and use them predominantly to express desires. 7. Ik ga morgen werken (I go-FIN Tomorrow work-INF) 8. Wil je spelen? (Want-FIN you play-INF) The proportion of modal Root Infinitives in English may also be considerably higher than the .13 that was found in a corpus study by Ud Deen. An inherent weakness of corpus studies is that the modal/nonmodal reading of an utterance is assigned on the basis of the context in which it is produced. However, since the corpora are transcripts of spontaneous speech, the information required to discriminate between modal and non-modal readings is often lacking. For this reason, Blom, Krikhaar & Wijnen (2001) conducted an experiment in which children produced descriptions of modal and non-modal events. In the experiment, the majority of Dutch children’s Root Infinitives (68%) were used to describe modal events. For the English children this was 44%. While this difference was significant and in the expected direction, this finding is problematic for Hoekstra & Hyams, as it suggests that the difference between Dutch and English is not a qualitative difference, but a graded, quantitative one which may well be related to the distributional characteristics of the language rather than differences in infinitival morphology. In this paper, MOSAIC will be used to investigate the source of the differential reading of Dutch and English Root Infinitives. MOSAIC has a number of characteristics that make it a suitable candidate for such an investigation. Firstly, the model has already been shown to successfully simulate the developmental change in the prevalence of Root Infinitives in Dutch and English (Freudenthal, Pine & Gobet 2002a, 2003, in preparation), as well as phenomena related to Subject Omission in English (Freudenthal, Pine & Gobet 2002b). The model’s success in simulating the finer detail of the OI phenomenon therefore provides a strong test of an input-driven account of the OI phenomenon. Secondly, the model learns off Child Directed Speech. The use of Child Directed Speech ensures a realistic frequency distribution, so that differences in the surface characteristics of a language are reflected in the input in a quantitatively realistic way. This is of particular importance as the practice of using artificially created input sets (which is common in simulations of phenomena in child speech) may lead the researcher to misrepresent the distributional characteristics responsible for the phenomenon under investigation. Thirdly, MOSAIC uses no built-in linguistic knowledge. Whatever representations it builds up during learning are a result of the interaction between its learning mechanism and the distribution of the input it sees. This last characteristic is important because Hoekstra & Hyams’ explanation of the differential reading of RIs is dependent on the assumption that the child knows that the infinitival morpheme implies a modal interpretation (rather than learning the association through exposure to the input). MOSAIC will be described below, followed by the details of the simulation. Simulating Language Acquisition in MOSAIC Whilst the version used for the simulations discussed here has changed from the earlier simulations, the main theoretical underpinning of the model remains the same. The basic tenet of the model is that the learning of language is a performance-limited process which is heavily weighted towards the most recent elements in the speech stream (i.e., which has an utterance final bias). Several authors have argued that children are better at learning material that occurs towards the end of the utterance (Naigles & HoffGinsberg, 1998; Shady & Gerken, 1999; Wijnen et al. 2001). MOSAIC learns from orthographically coded input, with whole words being the unit of analysis. The model is a simple discrimination net (an n-ary tree) which is headed by a root node. At the start of learning the discrimination net consists of just the root node. More nodes (encoding words or phrases) are added as the model is shown more utterances. An important requirement for nodes to be added is that whatever follows the word to be encoded in the input, must already have been encoded in the model. That is, the model will only learn a new word, when it has already encoded the rest of the utterance. This results in the model building up its representation of the utterances it is shown by starting at the end of the utterance, and slowly working its way to the beginning. If the model were to see the utterance I go home three times, it would on its first pass encode the fact it has seen the word home at the end of an utterance. On the second pass, it would encode the sequence go home. After a third pass, it would have encoded the whole utterance. Figure one gives a graphical representation of the model at this stage. The fact that MOSAIC builds its representation of an utterance by starting at the end of the utterance is the major mechanism responsible for its simulation of the development of Root Infinitives in Dutch. Early in Dutch children’s development, 80-90% percent of their utterances containing verbs are Root Infinitives. This drops to 10-20% later in development (Wijnen et al, 2001). Early in training, the model encodes many utterance final phrases. Since the infinitive takes sentence-final position (as can be seen in examples 7 and 8), the model produces many utterances with only non-finite verb forms. As the model encodes longer and longer phrases, these Root Infinitives are slowly replaced by auxiliary/modal plus infinitive constructions. Figure 1: MOSAIC after it has seen the utterance I go home three times. In the example illustrated in Figure 1, a (sentence final) word is encoded after one exposure. In fact, MOSAIC actually learns much more slowly than this, and the input corpus is fed through the model several times, so output of increasing average length can be generated after consecutive exposures to the input corpus. The probability of creating a node in MOSAIC is given by the following formula: € NCP = 1 1+ e / c 