The Bird on the Pole by the Pond On Three-site NPs, Relative Clauses and Prosody Report of the course The Experimental Study of Language (Deficiencies) at the Utrecht University. Eric Auer Sarah Bernolet Sanne Bongers Mirte van Galen Abstract Frank Wijnen en Hugo Quene claim that the U-shaped attachment preference pattern is not caused by the recency principle and the predicate proximity principle (Gibson et al. 1996), but by implicit prosody. Research has shown that implicit prosody is used in ambiguity resolution. Wijnen and Quenes theory is, that the second NP in a row of three is not the preferred antecedent of a relative clause because there is no prosodic pattern that expresses this attachment. In their experiments their hypothesis is confirmed. In this paper we describe how we tried to make a useful contribution to their research. We did three experiments. In the first experiment we check whether the problem of pronouncing the middle NP occurs only in complex NPs including the relative clause. In the second experiment we copy Wijnen and Quenes production experiment in a slightly different way. Our third experiment tests wether the pronounced sentences from experiment 3 are understood. Experiment one gave few surprises: in a complex NP consisting of three nouns connected by prepositions it is not difficult to accentuate one of them. The results we got for experiment two and three were not as we predicted: disambiguating these sentences by means of prosody seems to be a very difficult task. So maybe it is not prosody that helps us to choose an antecedent for relative clauses. 1.Introduction Frank Wijnen & Hugo Quene based their research on earlier research on relative clause attachment. Several researchers have tested what would happen when people are confronted with ambiguous relative clauses, i.e. relative clauses with more than one possible head or antecedent (e.g. Someone shot the servant of the actress who was on the balcony). The researchers are all eager to know which one of the possible nouns people choose when confronted with a structural ambiguity like that. When comparing the attachment preferences across several languages, researchers discovered that attachment preferences differed somewhat across languages (Mitchell & Cuetos, 1988): in some languages (e.g. Spanish) people tend to attach the relative clause to the first possible antecedent (= high-attachment tendency), whereas in other languages (e.g. English) people opt for the most recent noun, the one that is closest to the relative clause (= low-attachment tendency). The strange thing is that there seems to be no obvious typological criterion that aligns with these attachment preferences. Moreover, the high-attachment tendency violates the recency principle. The recency principle is found in research on many kinds of structural ambiguities. The RC attachment ambiguity is the only principle that shows exceptions to this tendency: in several languages the first attachment site is the preferred one, not the most recent one. A possible explanation for this can be that, when it comes to RC attachment, the recency preference is modulated by another principle. According to Gibson et al. (1996), recency preference is modulated by a second structural parsing principle: predicate proximity. The principle of predicate proximity makes people choose for an attachment as close as possible to the head of the predicate phrase when they are confronted with more than one possible attachment site. In other words, when people have to choose from more than one attachment site, two principles are at stake: recency preference and predicate proximity. The first principle makes people choose the last noun as head of the relative clause, whereas the second principle makes them choose the first possible head. The strongest principle, the one that guarantees the least costly attachment, wins. Gibson et al. speculate that predicate proximity is a parameterized structural principle and that its strength is related to word order strictness: the tendency is stronger in free word order languages than in languages with a strict word order. In other words: when a language has a strict word order, it is less likely that the predicate proximity principle will win the race for the attachment. To test their hypothesis on the predicate proximity principle, Gibson et al. explore the relative clause attachment preferences in a three-site context. Example: The apple next to the sandwich with the cheese that tastes very good. If they are correct about the two conflicting principles, the preference function in a three-site context will be non-monotonous: irrespective of the relative preference for the marginal sites, the intermediate site will be less preferred than the marginal ones. The intermediate site (e.g. the second in a series of three sites) is neither the most recent one nor the one that is closest to the head of the predicate phrase; it has the two principles working against it and thats why the attachment preferences in a three-site context form a U-shaped pattern. The results of their experiments show that the attachment preferences are indeed not monotonous: the preferences for the first and the last noun vary across languages and, more important, attachment to the second noun is always least preferred. The problem, however, is that Gibson et al. cant generalize their hypothesis on the predicate proximity principle as a parameterized structural principle. It is true that in some languages the strength of the principle aligns with the word order strictness (e.g. in Spanish and in English), but there are several languages in which contrary results are found (e.g. Brazilian Portuguese). These results have led Frank Wijnen & Hugo Quene to the conclusion that In general the proposed principle of Predicate Proximity appears to be somewhat shaky or ad hoc. (2000: 11). Frank Wijnen & Hugo Quene want to give an alternative explanation for the U-shaped preference pattern that Gibson et al. found. They think that implicit prosody (or rather: the limitations of prosodic phrasing) has a great deal to do with this U-shaped pattern. Several studies have shown that prosody helps people to understand language. Prosodic structures like intonation, phrasing and rhythm influence the processing of spoken language: a pause signals where a sentence ends, intonation is used to mark what is important. A number of studies have shown that people tend to rely on prosody to understand sentences that contain local or global ambiguities: prosodic phrasing and focus accents, for example, can have strong effects on attachment decisions (Schafer, Carter, Clifton & Frazier 1996, [see Wijnen & Quene, 2000: 3]). When listeners are confronted with a relative clause with two possible antecedents, they tend to attach the relative clause to the noun that is marked by prosody. Prosody does not only influence the processing of spoken language: several studies have shown that prosody has its effects on visual language processing too. Some researchers assume that people compute a phonological representation of what they are reading and that this representation encodes prosodic characteristics of the read materials (e.g. Bader, 1999). Readers give an implicit prosodic structure to written language. The information provided by this prosodic structure can help readers whenever they are confronted with structural ambiguities. This brings us back to the experiments of Frank Wijnen & Hugo Quene. As stated before, they are convinced that the U-shaped preference pattern from the Gibson et al. experiment can be explained by implicit prosody. Wijnen & Quene noticed that the problem with the sentences in the Gibson et al. experiment ( NP1 prep NP2 prep NP3 RC ) is that it is highly difficult to set the middle NP apart prosodically. You can single out the first and the last noun by using prosodic boundaries: a boundary after the first noun sets the first noun apart from the other two nouns, a boundary before the last noun sets the first two nouns apart from the last one. In this way you can mark these nouns as head of the RC. If you want to single out the second noun, you need to use two boundaries and by doing that, you set the three nouns apart. Of course there are still other ways to mark the intended RC-attachment prosodically (pitch accent), but Wijnen & Quene suppose that the second noun cannot be felicitously and acceptably made prominent by pitch accent and that leaves them with the conclusion that by prosody alone, marking the middle NP in a three-site context as head of a subsequent relative clause is impossible, or at least highly unnatural (2000: 12). Because people have no prosodic pattern that favors or selects the second NP, RC-attachment to this noun is less preferred. According to Wijnen & Quene, the implicit prosodic structure makes readers choose the first or the last noun as head of the relative clause and they are convinced that this is why the attachment preferences in a three-site context form a U-shaped pattern. To prove their hypothesis, Wijnen & Quene conducted 4 experiments. In experiment 1 & 2, a questionnaire and a continuous grammaticality judgment procedure, they try to replicate the Gibson et al. experiments. Before they go deeper into the matter, they check whether they too find a non-monotonous attachment preference function in three-host constructions. Experiments 3 & 4 concentrate on the role of prosody in RC-attachment: experiment 3 is a speech production study and experiment 4 is an auditory questionnaire. By means of these experiments Wijnen & Quene want to investigate whether prosodic phrasing can be used to signal the intended RC-attachment in a three-site context. They want to see whether people are able to disambiguate these three-host constructions by using prosodic features. They expect that people are able to do so, but only when the relative clause has to be attached to the first or the last noun. On top of this, they want to see whether people succeed in communicating the intended RC-attachment to others. Experiment 4, the auditory questionnaire, investigates whether listeners can perceive the intended RC-attachment when the sentences are disambiguated by means of prosody. We based our research on those two experiments (experiments 3 and 4 from the Wijnen & Quene -study). Our research consists of three experiments: two speech production tests and one auditory questionnaire. In the first speech production test we confront people with series of three nouns, linked together by prepositions (NP1- prep NP2 prep NP3). The structure of these complex noun phrases is the same as the noun phrases in the Wijnen & Quene-experiment, without the relative clause. We ask people to read these complex noun phrases aloud and to make one of the three nouns more prominent than the other ones. We conduct this experiment to check whether people can set the middle noun apart by means of prosody. In this way we can see whether people have a prosodic pattern that selects the middle noun in a series of three as the head of a complex noun phrase. We do this in order to check whether there only arises a pronounciation problem when a complex noun phrase like that is combined with a relative clause. If problems arise when people have to pronounce these complex noun phrases, it is obvious that it isnt the relative clause that is the problematic factor. If they have no problems, the problematic factor must be the disambiguation of the complex NP with the relative clause. In the second experiment we used test sentences that are comparable to the test sentences that are used in the Wijnen & Quene-experiment ( NP1 prep NP2 prep NP3 RC ). The only difference is that we used a different method to convey the intended RC-attachment. Wijnen & Quene signaled the intended RC-attachment either through number agreement of the auxiliary verbs in the RC with one of the three antecedent noun phrases (experiment 2) or by underlining one of them (experiment 3). We think that these methods influence the behavior of the subjects. When you force agreement between one of the noun phrases and the relative clause, it is no longer necessary to disambiguate the sentences by means of prosody. According to us, people will try less hard to mark the intended RC-attachment prosodically if it is already marked otherwise, either by agreement or by underlining. And thats not a good thing when you are looking for prosodic patterns. We tried to keep the sentences globally ambiguous, so that the subjects have no other choice than to make use of prosody to mark one of the noun phrases. The test sentences are presented with an instruction that elicits one of the three noun phrases (e.g.: Maak duidelijk dat het het hondje is dat we zo leuk vonden. Make clear that it is the puppy that we liked very much. ] is the instruction for the sentence: Het lied over het kind met het hondje dat we zo leuk vonden. The song about the child with the puppy that we liked very much ] ). In this way we dont need to use visual or grammatical means to mark the intended RC-attachment. We think that because we use globally ambiguous sentences our results will differ from the results in the Wijnen & Quene-experiment. In the third experiment we listen to the recordings of the first two experiments. You can compare this experiment with experiment 4 in the Wijnen & Quene-study. There are two differences. One is, that the sentences listened to are not pronounced by a phonetically trained speaker, but by the subjects. Te second difference is, that in this experiment, we are our own subjects. Without looking at the instructions, we check whether we can perceive which one of the three noun phrases is marked by prosody (experiment 1) or which noun phrase is marked as the head of the relative clause (experiment 2). In fact, our research is more or less the same as what Wijnen & Quene did. We try to refine the research method, but we are no trained researchers yet. Nevertheless, we want to see whether we can replicate their findings, using globally ambiguous sentences. 2.The material In our experiment our goal is to find out if it is true that there is no prosodic structure for sentences in which the RC is attached to the second of three NPs. The basic material we made has the following form: [NP, Prep, NP, Prep, NP, RC]. Example: Het lied over het kind met het hondje dat we allemaal zo leuk vonden. The song about the child with the dog that we all liked so much. But to use any sentence of this form would not be sufficient. There are certain conditions that the sentences have to satisfy in order to be usable for our experiment. To get pure results we have to minimize influences like semantics or sentence length. The first condition is a semantic condition. For every NP in the sentence it must be equally semantically plausible to attach to the RC. So De lamp boven de tafel naast de vrouw die niet goed in het interieur past. The lamp above the table next to the woman that does not go with the interior. is not a good sentence to use, because a woman is not a thing that does or does not belong in an interior. To be sure we used the right words, we did a pretest for this condition. This pretest will be discussed in the next section. The second condition was the condition that every NP is equal in length. Previous research has shown that the length of words in sentences has influence on the prosody. We want only the structure of the sentence to have influence on prosody, not length. So we had to make sure every NP had the same length. The length we use for the NPs was one or two syllables. The length of the RC also matters. We divided the RCs in the material into two groups: one of five or six syllables and one of nine or ten syllables. By doing this we make a little sidestep in our experiment: we are also able to examine the difference between long and short RCs. A third condition on the material involves attachment between the NPs. If we want to examine attachment of the NPs to an RC, the NPs must not be attached to each other, because that would reduce the choice of three NPs to two. This is the case in De roman met de tekst van de schilder die ik erg goed vind. The novel with the text by the painter I like very much. In this sentence you can see de tekst van de schilder (the text by the painter) as one thing to which can be referred. But the reader has to see the three NPs as separate things. This is why we didnt use the word : it gives the wrong effect. Also it must not be possible for the third NP to control the first NP. When the first and third NP would be connected, this would mean that the second NP is less favourable in any case. An example De film met de acteur over de stuntman die populair was. The movie with the actor about the stuntman that was popular. In this sentence, it is the movie that is about the stuntman, so the actor will be less salient. In our material we have avoided this connection between the first and third NP as much as possible. When a sentence satisfies all the above conditions, and is of the right form, it can be used for our material. We have succeeded in making 48 sentences. They can be found in the appendix. The sentences described above are the sentences that are used for experiment 2 and 3. The sentences used for experiment 1 are the same sentences without RC. For an experiment to be valid fillers have to be used. If there are no fillers, the experiment will become too predictable for the subjects, which may influence the results. The fillers have to resemble the real material, so the same instruction can be used for them. For the first experiment we used two kinds of fillers. They have the following forms: [NP, Prep, NP, Prep, NP, Prep, NP] and [NP, Verb, NP, NP]. For the second and third experiment we used fillers of this form: [NP, Prep, Adj, NP, RC] and [Adj, NP, Prep, NP, RC]. We also used two kinds of ambiguities. 3.Pretest To verify the suitability of the experiment items, a pretest was done. As explained above, we first threw out items where, according to our intuition, it was plausible that the first and the third NP would belong together. The pretest was an offline plausibility rating experiment: 23 people filled in a questionnaire on the web, giving the plausibility for each of the three attachments for each potential experiment item. The goal was to remove items that showed a clear bias towards one or two of the attachment possibilities. Given the sentence De acteur in de film over de stuntman die populair was. The actor in the movie about the stuntman that was popular. the three cases to rank would be: De acteur die populair was. De film die populair was. De stuntman die populair was. The ranking was done on a scale from one (bad, not plausible) to five (good, plausible) for each of the cases. The instructions asked to tell how good each of the "words" (we avoided the technical term NP) would fit the relative clause, and the subjects were encouraged to give the same score for more than one case if appropriate. The complete questionnaire can be found on the web. When gaps were found in the questionnaire, all three cases of the concerning sentences were removed from the data for that questionnaire. After collecting enough results, a statistical measure was computed from the logged data of the questionnaires. For this, the data was first normalized: The mean for each subject was subtracted from all values, so that the new mean would be zero for each subject. Then, the values were scaled to give a standard deviation of one. In this way, individual differences in the general treating of the questionnaire were removed. If, for example, one subject considers ranking three as normal and one and five as bad and good, while another one thinks most cases are good for a ranking of four, and never worse than three, the further evaluation would have been biased without the normalization. After the normalization, the mean and standard deviation were calculated for each case of each sentence, using the data of all questionnaires together. This data was used for additional checking of the results by hand: One sentence was removed by hand: Though all three attachment possibilities were similar in plausibility there, the overall plausibility of the cases (De puree/De schaal/De taart waar over gemorst is) was quite bad (average of -1.15). The main check, however, was done automatically: For each sentence, an F-value telling how strongly the three cases differed was computed from the data. Then, the sentences were sorted using that value, giving a list with F-values from 0.02 to 6.17. All sentences with a value of more than 2.4 were removed, as well as the generally implausible sentence mentioned above. Our selection of p=0.1 was done to reject even items with a low possibility of bias towards one or two of the possible attachments. Finally we had 48 experiment items remaining for the use in the two production experiments. 4.Experiment 1 Methods Subjects: Six unpaid students took part in this experiment (4 women, 2 men). All of them were native Dutch speakers. None of them took part in the pretest. Materials: See the sections on the material and the pretest for the selection of the experimental sentences and the fillers. In the appendix is a list of the used material. In this experiment we used the selected experimental sentences, but without the relative clause. So the subjects were given sentences with the grammatical structure NP1-prep-NP2-prep-NP3. Procedure: Three different lists were constructed by a computer program that randomly assigned the three different cases of every single sentence to a list. This means that although all of the subjects were given the same sentences, subjects with a different list never had the same instruction for a particular sentence. There was no restriction for the number of instructions for a certain NP within a list, so that is why those numbers sometimes are a bit unbalanced. The experimental sentences were pseudo-randomly intermixed with the fillers, so that there never were more than three fillers or more than two experimental sentences in a row. Each list had a different order. Each list was given to two subjects. The experiment took place in a soundproof environment. Subjects were given a general instruction that explained that they had to make clear which word was important in a (fragment of a) sentence by means of their prosody (see appendix). Everything was presented on a computer screen and the subjects were able to decide themselves when they were ready for the next sentence (by way of clicking on a next-button on the screen). Each sentence had a specific instruction that showed the word that they had to make prominent. The utterances of the subjects were recorded on a DAT-tape. Data analysis: The recorded signal was analysed by us (all untrained listeners). In principle we did not use the computer for acoustic analysis, but for the utterances of three subjects the computer was used for extra information. We scored in the categories high pitch low pitch loud quiet stretched words breaks between words and the number of times the sentence was pronounced (we told the subjects that they could repeat the sentence as much as they wished). We did not use a statistical analysis for this experiment, because there was not enough time to do that. Results Because there were some individual differences between the subjects in the way they pronounced the sentences, we analysed the data first for each subject. In this section you will find a table with results for each subject. The first column shows the figures of the sentences that had an instruction to make the first NP prominent, the second column shows the figures of the sentences with a NP2-instruction and the third column shows the figures of the sentences with a NP3-instruction. The numbers in parentheses are the amount of experimental sentences with a certain NP-instruction. The words on the left denote the different categories that are listed under data analysis. There has been made a distinction in NP-prep (the preposition before the NP), NP-det (the determiner of the NP) and NP-noun (the noun of the NP) if necessary. Categories that did not show up in all of the utterances of a subject are not listed in the tables. The category remarks is used to give extra information. Words with two syllables that had a high pitch first syllable and a low pitch second syllable, were counted as high pitch (because it sounds high). Words with two syllables that had a low pitch first syllable and a high pitch second syllable, were counted as low pitch (because it sounds low). Subject 1 (list 1) A) Although in six cases the NP3-noun was not the only part of the sentence that had a high pitch, in four of these cases the NP3-noun was still higher than the rest of the sentence. Results This subject shows a consistent pattern. The NP that is important is made prominent by pronouncing its noun in a loud and high pitched way. There are some sentences with a NP3-instruction that have a high pitch NP1-noun as well as a high pitch NP3-noun. It could be that the noun of NP1 is easier made high pitched, as a kind of default. There are also a few sentences with a NP2-instruction that show a low pitch NP3-noun, this could be because the subject does not consider the experimental sentences as fragments of sentences. Whole sentences have a low pitch at the end of the sentence. The subject does not use breaks. Subject 2 (list 1) A) There were two cases where the NP1-noun was not the only part of the sentence that was loud, but in one of these cases the NP1-noun was still louder than the rest of the sentence. B) There were ten cases where the NP2-noun was not the only part of the sentence that was loud; in five of these cases the NP2-noun was louder than the rest of the sentence and in one case the NP1-noun was louder than the rest of the sentence. C) There were ten cases where the NP3-noun was not the only part of the sentence that was loud, but in six of these cases the NP3-noun was still louder than the rest of the sentence. Results This subject shows the same pattern as subject 1, although it is a little less clear. NPs are made prominent by pronouncing the noun louder and with a high pitch. Half of the sentences with a NP2-instruction have a loud NP1-noun as well as a loud NP2-noun. But in half of these latter cases, the NP2-noun is louder than the NP1-noun. Also a lot of the sentences with a NP3-instruction have a loud NP1-noun and a loud NP2-noun, as well as a loud NP3-noun. But again, in a lot of these cases the NP3-noun is the loudest one. The sentences with a NP2-instruction and the sentences with a NP3-instruction both have a few cases where the NP1-noun has a high pitch (as well as the intended NP-noun). There are some examples of the use of stretching to denote the intended NP, in sentences with instructions for NP1 or NP2. It seems sentences with a NP1-instruction have the most straightforward pattern. If NP1 really is the easiest sentence to produce, this is what would be expected. About breaks: If there is a break in sentences with a NP1-construction, this break is located between the first and the second NP. If there is a break in sentences with a NP2-construction, this break is located between the second and the third NP. There never are any breaks in sentences with NP3-instruction (there cannot be a break after NP3). Subject 3 (list 2) A) Although in four cases the NP1-noun was not the only part of the sentence that had a high pitch, in two of these cases the NP1-noun was still higher than the rest of the sentence. B) Although in two cases the NP2-noun was not the only part of the sentence that had a high pitch, in one of these cases the NP2-noun was still higher than the rest of the sentence. C) There was one sentence (with a high NP2-noun and a high NP3-noun) where the NP3-noun was the highest part of the sentence and there was another sentence (with a high NP1-noun and a high NP2-noun) where the NP2-noun was the highest part of the sentence. Results The pattern of the sentences with a NP1- or NP2-instruction is more or less the same as the pattern of the first two subjects: the noun is higher and louder. Specific for this subject is that there is a low pitch NP3-noun in most of the sentences, regardless of the instruction. So the subject drops his voice at the end of the sentence, sometimes even before the NP3-noun. There are a few cases in which the sentences with a NP1-instruction not only have a high pitch NP1-noun, but also a high pitch NP2-noun. The sentences with a NP3-instruction do not have a clear pattern. The noun that is high most of the times is the noun of NP2. The NP3-noun has a low pitch most of the time, but that is probably because the subject drops his voice at the end of a sentence. It is true that the NP3-noun is more often loud than the noun of the other NPs, but there are only 5 cases (of 16 sentences) where the NP3-noun is loud. Remark: If a preposition is pronounced high and loud, it is nearly always the preposition naast (next to). Subject 4 (list 2) A) Although in two cases the NP1-noun was not the only part of the sentence that was loud, in one of these cases the NP1-noun was still louder than the rest of the sentence. B) In nine cases both the NP1-noun and the NP2-noun were loud, in seven of these cases the NP2-noun was still louder than the NP1-noun. C) Although the NP3-noun was loud just as many times as the NP1-noun, in nine cases the NP3-noun was louder than the NP1-noun. Results The general pattern here is the same as the pattern of the other subjects, although this subject makes more use of louder NP-nouns than of higher NP-nouns. Sentences with a NP2- or a NP3-instruction have a loud NP1-noun (as well as a high noun of the intended NP) most of the times, but generally the intended NP-noun is louder. There are some cases of sentences with a NP2-instruction where a part of the NP3 is high pitched, which is remarkable. Sometimes the whole NP3 is quiet in sentences with a NP1- or NP2-instruction. The breaks show the same pattern as the breaks of subject 2: The breaks in sentences with a NP1-construction are located between the first and the second NP. The breaks in sentences with a NP2-construction are located between the second and the third NP. There never are any breaks in sentences with NP3-instruction. Subject 5 (list 3) A) Although in five cases the NP1-noun was not the only part of the sentence that was loud, in four of these cases the NP1-noun was still louder than the rest of the sentence. B) In all of the sentences both the NP1-noun and the NP2-noun were loud. In two of these cases the NP1-noun was higher than the NP2-noun and in three cases the NP2-noun was higher than the NP1-noun. C) In two cases the NP1-noun was the loudest part of the sentence, in one case the NP2-noun was the loudest part of the sentence and in six cases the NP3-noun was the loudest part of the sentence. Results Sentences with a NP1-instruction have a straightforward pattern: the NP1-noun is again louder and higher than the rest of the sentence. In sentences with a NP2-instruction there is far less use of pitch, compared to sentences with a NP1-instruction. In all of the NP2-sentences the NP1-noun and the NP2-noun have a high pitch. The preposition of NP2 is also loud most of the times. Sometimes the NP3-noun is stretched. So this pattern is different from the other subjects, although it has been noticed before that the NP1-noun gets a stress most easily (high pitch or louder). In sentences with a NP3-instruction there is no use of pitch at all. The volume is remarkable, because most of the times all three nouns are loud, although sometimes the NP3-noun is the loudest one. So this is also a different pattern. The subject does not use breaks. Subject 6 (list 3) A) There was one sentence where both the NP1-noun and the NP2-noun had a high pitch, but the NP2-noun was higher. B) There were nine cases where the NP3-noun was not the only part of the sentence that had a high pitch; in three of these cases the NP3-noun was still higher than the rest of the sentence. Results Generally the pattern is the same as the pattern of the other subjects: the noun of the intended NP is louder and higher than the rest of the sentence. This is most easily recognized in the sentences with a NP2-instruction. The NP1-sentences sometimes have a high pitch NP2-noun (3 times) as well as a high pitch NP1-noun. Sometimes the NP1-noun is stretched. The pattern of sentences with a NP3-instruction is a little bit less clear: the NP1-noun is also loud aud aud a *** sigh... now a part in unicode or something starts (mostly appendix) *** Sanne bluedread Microsoft Word 8.0 ...