MODAL PARTICLES IN DUTCH AS A SECOND LANGUAGE: EVIDENCE FROM A PERCEPTION EXPERIMENT 1

This study examines the perception of modal particles in Dutch, contrasting native and non-native speakers. We will use the term ‘modal particle’ as a cover term for all those usually high-frequent, usually adverb-like words found in real-life discourse in languages such as German, Dutch and classical Greek, used to express meaning nuances in various areas such as modality, evidentiality and aspect (cf. e.g. Schiffrin, 1987; Foolen, 1996). Modal particles are probably difficult for second language learners to master, not only because of strong mismatches between languages as regards their usage of such lexical elements, but also because this class of words shows an intricate association between word meaning and prosodic parameters (Foolen, 1986, Cheon-Kostrzewa, 1996, Wenzel, 2002, Möllering, 2004). In the large majority of cases, modal particles in Dutch do not carry a pitch accent, i.e., they are not highlighted by a conspicuous melodic movement (König, 1991, Vismans, 1994. One could also say that they are deaccented, cf. Elffers, 1997). Very often, however, they have an accented counterpart which functions as a ‘normal’ adverb (cf. e.g. Abraham, 1984). For example:2


Introduction
This study examines the perception of modal particles in Dutch, contrasting native and non-native speakers. The term 'modal particle' will be used as a cover term for a group of high-frequent words found in real-life discourse in languages such as German, Dutch and classical Greek, used to express meaning nuances in various areas like modality, evidentiality and aspect (cf. e.g. Weydt (ed.), (1977) where comparable elements are called 'Modalpartikeln' or 'Abtönungspartikeln', Schiffrin (1987) where we find the term 'discourse markers ', and Foolen (1996) who speaks of 'pragmatic particles'). Whether or not modal particles constitute a proper subset of the class of adverbs (or of any other syntactic or semantic class of lexical items) is a question beyond the scope of this paper (cf. e.g Henschel & Weydt 1989).
Given that languages differ considerably both in the number of particles they have and in the ways they are deployed, it has been hypothesized that they constitute a problem for non-native speakers (Weydt (ed.), 1981;Foolen, 1986). Mastering Dutch particles is furthermore complicated by the fact that the grammatical and semantic properties of Dutch particles are treated incompletely (at best) in grammars and dictionaries (Van der Wouden and Caspers, 2010) and in learning materials (Foolen, 2010). Modal particles may be particularly difficult for second language learners to master, not only because of strong mismatches between languages as regards their usage of such lexical elements, but also because this class of words shows an intricate association between word meaning and prosodic parameters (Foolen, 1986;Cheon-Kostrzewa, 1996;Wenzel, 2002;Möllering, 2004). In the large majority of cases, modal particles in Dutch do not carry a pitch accent, i.e., they are not highlighted by a conspicuous melodic movement (König, 1991;Vismans, 1994).
Very often, Dutch modal particles have an accented counterpart in one or more other word classes (cf. e.g. Abraham, 1984). For our experiment, we have chosen particles with an adverbial counterpart. At times, we will be somewhat sloppy in our terminology and refer to the accented counterpart as '(normal) adverb' or 'adverbial interpretation'. For example: (1a) below exemplifies a use of the modal particle zeker,  (1a) there is a pitch accent on the noun pijn (indicated by the capitals), while the word zeker remains unaccented. As a result, zeker functions as a modal particle, giving the whole utterance an uncertain, testing flavor -it almost functions as a (yesno) question. In example (1b) pijn is unaccented and the sentence accent in this case is located on zeker, leading to an adverbial interpretation of the word (viz., 'certainly').
Dutch intonation has to do with information structure and focus (Keijsper, 1985;Baart, 1987). In light of the kind of things modal particles contribute to the semantics and pragmatics of utterances, it is hardly surprising that these lexical elements cannot be focused, which helps us understand why they cannot be accented. On the other hand, modal particles are sensitive to information structure: at least since Krivonosov (1963) we know that they tend to mark the borderline between theme and rheme, i.e., between old and new information. In Dutch, both phrases carrying old information and those that express new information can be, and often are, prosodically marked, e.g. by means of a pitch accent. The preferred position of modal particles is then by definition one between two accented phrases, and therefore unaccented (Vismans, 1994).
With respect to second language acquisition, prosody has been virtually ignored for a long time (Chun, 2002). Recently, however, it has been established that prosodic factors such as stress position and pitch accent placement are crucial for judgements of non-nativeness and comprehensibility of L2 speech (Anderson-Hsieh et al., 1992;Munro & Derwing, 1995;1998;Trofimovich & Baker, 2006).
The results and conclusions of the few studies on transfer of prosodic characteristics (from L1 to L2) that have been carried out do not all point into the same direction. Some experimental studies involving Dutch do report transfer (Mennen, 2004;Roosman, 2006;Rasier, 2006), others do not (Caspers & van Santen, 2006). Mennen (2004) found evidence for transfer of the phonetic implementation of pitch rises from L1 Dutch to L2 Greek. Roosman (2006) reports L1 influence on the realisation of word stress in L2 Dutch. Rasier (2006) concludes that the location of sentence accent is transferred from L1 French to L2 Dutch. On the other hand, Caspers & Van Santen (2006) did not find evidence for influence of L1 on the location of word stress in spoken L2 Dutch.
To our knowledge, no experimental work has been done on the perception by second language learners of different accent distributions in Dutch. On the basis of the available literature it is to be expected that -unaccented or deaccented -modal particles are difficult for L2 learners of Dutch to interpret, and the experiment we discuss in the current paper was explicitly set up to establish whether this problem actually exists.

Approach
We devised a perception experiment in which subjects were presented with a context sentence and two stimulus sentences, differing only in the presence versus absence of a pitch accent on the target word. The context sentences were constructed in such a way that they would fit either an adverbial (accented) or modal (unaccented) reading of the stimulus sentence. We selected the following modal particles for our investigation: wel ('rather', 'quite'), zeker ('probably') en toch (untranslatable, indicates that confirmation is expected). The words chosen share the property that their interpretation as adverb or modal particle depends on the presence versus absence of a pitch accent. Furthermore, they have been investigated earlier (cf. Schermer-Vermeer, 1984;Vismans, 1994;Abraham, 1984, van der Wouden, 1999Hogeweg, 2005Hogeweg, , 2009, which means that at least some of their properties are known.
We hypothesized that DSL-speakers would have problems with the distinction between modal and adverbial use of words like zeker. We expected the control group to select stimuli with accented target words in non-modal contexts in the majority of cases, and stimuli with unaccented target words in modal contexts. For the non-native subjects, the largest problems were expected for the modal contexts. We thought it possible that the non-modal contexts, i.e., the stimuli containing accented target words might be responded to correctly, if only because accented lexical items are more salient than unaccented ones.

Stimulus materials
For each of the three target words five different stimulus sentences were created. The sentences were ambiguous as to the interpretation of the target as a modal particle or a regular adverb. They were read aloud by a trained speaker, who realized pitch accents at indicated positions, resulting in either a modal or a non-modal version of the sentence. In addition, the stimulus sentences were preceded by context sentences that were intended to lead to either a modal or a non-modal interpretation of the target word in question. Below examples are given of context and stimulus sentences; for the stimulus sentences the intonation contour is indicated à la Bolinger (1989) The pitch accent on zeker in stimulus a) triggers an adverbial interpretation (i.e., 'certainly') and therefore the sentence is an appropriate continuation in the given context. In stimulus b), zeker remains unaccented and the pitch accent in this case falls on the verb komt. As a result, zeker can only function as a modal particle, indicating a form of uncertainty, which does not fit context A very well.
The alternative context for these stimulus sentences -fitting a modal use of zeker, i.e., stimulus b) -is: context B De stoel van Sander is nog leeg 'Sanders chair is still empty' For the complete set of stimulus materials see the appendix.
After verification and segmentation, using PRAAT (Boersma & Weenink, 2006), the digital recordings were placed in a single audio file. Every context sentence occurred once, followed by both versions of the corresponding stimulus sentence (the relative order of the accented and unaccented versions was varied quasi-randomly). Context and stimulus sentences were presented twice in succession, in a fixed random order, with fixed pauses, and separated by beeps.

Subjects
A control group of 25 speakers of Dutch as a first language (DFL-speakers) participated in the experiment. They were between 20 and 27 years old, and 11 of them were male. The 17 DSL-speakers (late learners of Dutch) were first year students at the department of Dutch Studies at Leiden University. They had all followed exactly the same educational programme from September 2006 to May 2007, containing a large amount of language instruction. Their level was between A2 and B1, according to the Common European Framework of Reference (Council of Europe 2001), 3 ages varied between 20 and 41, and 5 were male. Eight DSL-speakers had Chinese as their first language (7 Mandarin, 1 Cantonese), the remainder of the group spoke an Indo-European language. No hearing problems were reported and all subjects participated on a voluntary basis.

Procedure
The group of native subjects was tested in one session, the non-native subjects, for practical reasons, in two other ones. The sessions all started with the distribution of answer booklets. On the first page the subjects had to fill in a short questionaire (year and place of birth, mother tongue, L3, etc.). Then they had time to read the instructions and to ask questions. Subjects were instructed to listen to context and stimulus sentences and mark on the answer sheets which of the two versions of the stimulus sentence they thought fitted the presented context best. A practise item was then played, after which questions could again be asked. Next, the experiment was run, without any further pauses. Both context and stimulus sentences were printed in the answer booklet, without an indication of the pitch accent position, differing the two stimulus sentences with (a) and (b). Subjects had to circle either the (a) or the (b) version, depending on which version they thought best suited the presented context. Table 1 contains the number of 'correct' (i.e., unaccented targets in modal contexts and accented targets in non-modal contexts) and 'incorrect' responses for the native and non-native speakers. For the group of non-native subjects, one case was missing. The data show that the DFL-speakers have a mean score of 90% correct responses, while the DSL-speakers have a mean of 57% correct responses (Pearson χ 2 = 190.256, df= 1, p< .001).

Results
In figure 1 the data for both the native and non-native speakers are broken down further by context type and target word. An analysis of variance on percentage correct reveals a large effect of nativeness (F(1,1257)= 245.495, p< .001), and a small main effect of target word (F(2,1256)= 6.819, p< .005). There is no main effect of context type (F(1,1257)< 1, ins.), which means that overall there is no significant difference in the number of correct responses to the modal and non-modal contexts. However, there are significant two-way interactions between nativeness and target word (F(2,1251)= 8.342, p< .001), between nativeness and context type (F(1,1253)= 7.182, p< .001) and between target word and context type (F(2,1251)= 6.650, p< .005). Finally, there is a strong three-way interaction (F(2,1247)= 46.005, p< .001). It is clear that wel is responded to differently by the native speakers in the modal and non-modal conditions: in a context designed to trigger a preference for accented wel, the subjects indeed selected accented wel in 79% of the cases, but the responses rise to a full 100% preference for unaccented wel in the modal context condition. This means that native speakers feel very strongly about unaccented wel in these contexts: apparently, the stimulus with the accented version is completely unacceptable. The reverse is not true: in non-modal contexts the preference for the accented version of wel is not absolute, in a fifth of the cases subjects select the unaccented version. The difference between modal and non-modal contexts is significant for wel (Pearson χ 2 = 29.018, df= 1, p< .001). The same is true for toch¸ albeit that the difference is much smaller and reversed (Pearson χ 2 = 8.786, df= 1, p<.005). Closer inspection of the data reveals that this effect is caused by one ambiguous context sentence (context 1b for zeker allows a modal as well as a non-modal interpretation). For zeker the effect of context type is not significant (Pearson χ 2 = 3.363, df= 1, ins.).
The non-native speakers have a higher percentage correct responses for wel in nonmodal contexts than the native subjects (87% vs. 79%), but a much lower score in the modal condition (49% vs. 100%). The responses for the other two particles show mirrored effects for both groups of speakers as well, which explains the strong threeway interaction reported above.
Wel in a non-modal context is the only condition the DSL-group seems to be really comfortable with (87% correct); the difference between modal and non-modal contexts is significant for wel (Pearson χ 2 = 27.791, df= 1, p<.001). For zeker the responses do not differ in the two types of context (Pearson χ 2 = 48.839, df= 1, ins.); the scores appear to be random here (on average 51% correct), indicating that the stimuli in both context types are generally hard for the subjects to interpret. Finally, there is a significant difference in the number of correct responses to the target word toch per context type (Pearson χ 2 = 16.619, df= 1, p<.001); in de modal contexts 66% of the responses is correct, while in the non-modal contexts only 35% of the responses is correct. It seems that there is a general preference for unaccented toch, irrespective of context type. We do not have an full explanation for this finding, but it could be the case that unaccented toch is much more frequent in spoken Dutch than accented toch.

Discussion and conclusion
It became clear that non-native subjects have problems with selecting the best fitting stimulus in the contexts designed to evoke a modal use of the target word, as we expected. However, in the non-modal contexts they experienced difficulty as well, except with the stimuli containing the target word wel. It appears that the DSL-speakers find zeker and toch more difficult than wel. This could be explained by the fact that accented wel is in direct opposition with the negation niet (not), which makes it highly contrastive and rather hard to miss for DSL-learners. 4 Recent work on spontaneous speech elicited from DSL-speakers with Spanish as L1 revealed an unexpected but clear hierarchy in the acquisition of modal particles (van Balen et al. , 2010). It turned out that the higher the proficiency level of the subject, the more modal particles were used, and a ranking in difficulty was possible for the different particles. When a Spanish DSL-speaker uses modal particles, maar is the first one to appear, followed by wel; toch is tenth in line, and zeker is not used at all by the investigated group of speakers. Furthermore, the non-modal use of the target words shows a parallel pattern: wel is more frequent than toch, and zeker is completely absent from the DSL-data. It seems that wel is acquired much earlier than toch, and zeker clearly is the hardest.
We are inclined to interpret the results of our experiment as an indication that Dutch particles, as well as their interaction with prosody, merit more attention in didactic materials aimed at DSL-speakers. It seems reasonable to reserve some (more) time for the phenomenon of accentuation (its nature and function), and to try to make the students sensitive to the fact that the meaning of Dutch words can depend on whether or not they carry accent. Moreover, examples of Dutch modal particles and their accent-carrying adverbial counterparts should be offered (in context), in order to create a certain consciousness for modal particles and their contribution to the meaning and usage possibilities of utterances. The lack of descriptive work, however, will not make this a trivial task.