Volgende: 7. Voorkeuren van fonologische Omhoog: Link Vorige: 5. Interpretatie als een
In this paper I will design an experiment to test processing of the Frisian reflexive him in comparison with its Dutch equivalent zich. A different feature composition of him causes a different behavior in the theory of `Primitives of Binding'. Thus, in the theory the interpretation of Frisian him is computed in a completely different way than Dutch zich. We expect effects of this difference to emerge in processing of Dutch and Frisian reflexives. The paper is organized as follows: I will first give a theoretical background of Binding and of the differences between him and zich that cause the differences in interpretation. I will then proceed to explaining my motivation in setting up the experiment, my hypothesis and possible expectations. Furthermore, I will give a detailed overview of the experiment and its setup. I will also include some background information on the method I plan to use in carrying out the experiment, the eye tracker. Since I have not (yet) tested subjects with this experiment, all the hypotheses here will remain rather speculative.
In Dutch, we find a three-way distinction in pronominal elements (pronouns and anaphora). It has clear pronouns (hem, haar, ik) on the one hand and anapora (zich, zichzelf) on the other hand. Within the class of anaphora we find simplex anaphora (zich) and complex anaphora (zichzelf). The Frisian paradigm looks a bit different: Frisian can express both Jan wast hem and Jan wast zich by using one form, him. Thus (1) can have the reflexive meaning, but himcan also refer to someone different than Jan.
(1) wasket
This sentence is problematic for principle B of the Binding Theory (Chomksy 1981, Haegeman 1994). How can one and the same form within one and the same Governing Category have both a bound and a free meaning? Within the GB framework this is impossible, unless we assume that him has two lexical entries, one as a pronoun and one as an anaphor.
A comparable problem with him for the standard Binding Theory arises in (2):
(2) Jan wasket /
Let's assume him is a pronoun, which is one of its meanings in Frisian. Then
him would need to be free in the exact same Governing Category where
himsels is bound. I don't see how we could find such a Governing
Category since Jan is either going to be included or excluded
depending on how one formulates the rules. And again, we are faced with the
question how one element can be both a pronoun and an anaphor in the same
domain. We can only solve this problem by assuming two lexical entries for
him. In section
![[*]](file:/usr/share/latex2html/icons/crossref.png){kind=icon}
I will show that we don't need to make this stipulation
if we follow a different theory on binding.
For the reasons mentioned above, among others, a new view of binding has been developed over the last years (Reinhart & Reuland 1993, Reuland 2001, Reuland & Everaert 2001). This new theory allows us to tackle the problems with Frisian him. It lets go of the sharp distinction between elements that always have to be free (pronouns) and elements that always have to be bound (anaphors). The principles of Chomsky (1981) are reformulated. One of the major changes lies in the way elements are classified. A certain element is no longer clearly a pronoun or an anaphor, but can have properties of both. As a result of this, the binding conditions are no longer stated specifically for pronouns and anaphors. To take the example of him, we can now no longer classify him as either a pronoun or an anaphor, from which its behavior with respect to the Binding Theory would follow. The distribution of him now follows from its feature composition and the interaction of these features with certain economy rules. With these two important components of the Minimalist Program we can deal with a number of facts that were highly problematic for the classic view on binding, including the facts on the distribution of Frisian him. I will now turn to an explanation of how these problems are solved.
In the Minimalist Program a lot of the work that was done before by the
principles and parameters of Chomsky (1981) is now taken over by rules of
Economy such as the Minimal Link Condition and Procrastinate (Chomsky 1995).
These rules mainly describe the economy reasons for not pursuing a certain
derivation within the computational system. There are also economy rules
regulating the division of labor between the different modules language
consists of. These modules are displayed in figure .
Two of these economy rules are relevant to our experiment: Rule I and Rule BV.
(3) Rule I : Intrasentential Coreference: NP A cannot corefer with NP B
if replacing A with C, C a variable A-bound by B, yields an
indistinguishable interpretation.
(4) Rule BV: Bound variable representation: T may not translate an
expression E' in SEM' with syntactically independent NPs A' and B' into
an expression E in SEM in which A is A-bound by B, if there is an
expression E'' resulting from replacing A' in E' with C', C' an NP
such that B' heads a A-CHAIN tailed by C' and T also translates E''
into E.
These rules concern the various ways a dependency between two NPs can be computed. As we have seen in the model above, language consists of three modules: PF, LF and the computational system (CHL). All three of these modules work together to process sentences. Every module is responsible for part of the interpretation of language. The question now is, how is this labor divided? As is argued in Reuland (2001), the CHL component will take all the work if possible. Using syntax in the interpretation of certain dependencies is cheapest. If syntax is impossible the system can in principle revert to the Interface between CHL and LF. Interpreting a dependency via this Interface will be costlier than using syntax, but it will still be cheaper than using the third way of getting an interpretation: through discourse (LF). So only if the Interface interpretation is impossible will the system use discourse. This is what the two rules are saying: Rule BV states that if you have a syntactic way of interpreting a dependency, you have to use this instead of forming the same interpretation by using a costlier way, namely the Interface. Rule I states that if you can interpret a dependency by using the Interface, you can not form the same interpretation by using discourse. Economy always picks the cheapest way.
In the Minimalist Program, important operations such as Move are driven by the feature composition of words. Uninterpretable features have to be deleted and erased, and this can only be done in a checking relation with another element. The only way of interpreting a dependency in syntax is by (5) :
(5) i. The only way to establish a dependency: Move / Attract and
checking
ii. The only way to force a dependency: checking grammatical
features in a checking configuration.(Chomsky 1995)
To form a syntactic dependency all features of a moved element have to be
checked by the features of its antecedent. Thus, whether the syntactic
encoding of a dependency is possible depends completely on the feature
composition of the elements involved in that dependency. If and only if the
features of the moved element are a subset of or equal to the feature set of
the antecedent, a CHAIN can be formed. In other words, only if the features
of the moved element are not a subset of or equal to the feature set of the
antecedent, can the system use a costlier way of establishing a dependency.
All of this will become relevant in section
,
where I will discuss the
feature composition of Frisian him and the way it interacts with the
Economy rules.
The above rules force the system to use the cheapest computation for a dependency. But what does it mean if the cheapest option of all three is not available? Then the system has to opt for the costlier one. And if even this one is not available it has to pick the costliest one. Why is syntax the cheapest option? Because in the case of a CHAIN, only one cross-modular operation has to take place. A syntactic CHAIN has to be translated into a variable. The costlier options involve either one or two extra cross modular operations. This is schematized in (6) (taken from Reuland 2001), where (6)a represents interpretation via discourse storage, (6)b the interface and (6c) interpretation via CHAINs.
| (6) | a. | discourse storage (values) | a | a |
| C-I objects (variables) | x1 | x2 | ||
| syntactic objects (CHAINs) | C1 | C2 | ||
| basic expressions | ||||
| b. | discourse storage (values) | a | ||
| C-I objects (variables) | x1 | x1 | ||
| syntactic objects (CHAINs) | C1 | C2 | ||
| basic expressions | ||||
| c. | discourse storage (values) | a | ||
| C-I objects (variables) | x1 | |||
| syntactic objects (CHAINs) | C1 | C1 | ||
| basic expressions |
If the system is forced to use more cross modular operations, then it will need more processing resources. Thus, it seems interesting to gain evidence for the theory set up above by using processing. After all, if we are right about the way Frisian is derived, and if we are right about the economy rules we expect processing differences to show up when testing subjects.
In Dutch a reflexive relation can be expressed by the SE-anaphor zich 6.1. This anaphor has features for Case (weak feature), category (+D) and 3rd person only. Thus, it is not fully specified for j-features since it lacks a Gender 6.2and Number feature6.3. The feature specification of zich allows it to enter into a CHAIN relation with its antecedent in (7).
| (7) | wast | ||
| Jan | washes | zich | |
| `Jan | washes | himself' |
In this configuration the features of zich can move all the way up to
a checking relation with the antecedent Jan since they first move to
V and then the configuration V-FFzich moves further up to I. This is
shown in figure .
The position where FFzich ends up is in an automatic checking relation with Jan. Since the features of zich are all inherent and interpretable, they can be deleted provided they do not violate the PRD. This is indeed not the case, since the meaning of the features can be recovered from the exact same features in Jan. Deletion and erasure of a feature always take place if possible (Chomsky 1995).
Through Move/Attract there is a Chain dependency between zich and FFzich. And there is a Chain containing FFzich and Jan. Then Chomsky and Lasnik (1993) state that these two can be linked up in a CHAIN:
(8) If (1, 2) is a Chain and (1, 2) is a
Chain and 2 = 1 then (1, 2/1, 2)
is a CHAIN.
In this situation 2/1 = FFzich. Ultimately there is a CHAIN between Jan and zich. To conclude, the dependency between Jan and zich can, and therefore must, be interpreted within CHL. As we have seen this is the cheapest way of interpreting a dependency. If syntax is allowed, interpretation must take place using syntax and Rule BV prevents the system from using different methods such as the Interface or discourse strategies 6.4.
Now let's look at the Frisian (9):
| (9) | wasket | ||
| Jan | washes | him | |
| `Jan | washes | himself' |
Let us first consider only the reflexive interpretation of this sentence. In the interpretation of the dependency between Jan and him we first try the cheapest option, namely CHAIN formation. The syntactic situation is in principle the same as in Figure 2, only no Chain is formed between FFhim and Jan. And since there is no Chain, CHAIN formation between Jan and him will not be possible either. The absence of a Chain is due to the feature composition of him. Just like Dutch hem it has a Number feature. No formal dependency is formed since the Number feature prevents deletion of him's features and recoverability through the features of Jan 6.5. Rule BV now becomes irrelevant, since CHAIN formation does not prevent us from trying to interpret the dependency using the Interface. Binding between Jan and him is possible according to Reinhart (2000).
(10) Logical Syntax definition of A-binding
a A-binds b iff a is a sister of a
?-predicate whose operator binds b
Binding gets us the following representation:
(11) Jan ?x (x wasket x)
Now Rule I states we cannot get coreference in (12) yielding the same interpretation. To conclude, we have to use the C-I interface to interpret (9).
(12) Jan ?x (x wasket a) where a = Jan
The double meaning of (9) now follows. We can use a representation like (12) only if a refers to a different entity. Only then do we get a distinguishable interpretation. This gets us wasket . One single lexical entry him can now get different interpretations thanks to a change in formulating the binding theory.
Having discussed the theoretical background for Dutch as well as Frisian, we see that (7) and (9) are interpreted in completely different ways.
Now that we know how Dutch and Frisian reflexives are interpreted in these kind of examples, we can set up an experiment to test their actual processing. A first approach would be to simply try to test (7) and (9) and look for differences in processing times between the two reflexives. We cannot, however, test these sentences within the same population since this would automatically mean that we are dealing with bilingual speakers. This in turn could imply different levels of mastering of one language versus the other. Imagine we would get a processing difference for one speaker where Frisian him takes longer to interpret. How would we know whether these results are due to our hypothesis or to the fact that this speaker is a fluent speaker of Dutch, but rather inexperienced in Frisian? It seems much safer, therefore, to test each population separately. To solve this problem, we could consider testing Frisian native speakers for (9) and Dutch native speakers for (7). Then we could statistically generalize and see whether there are any significant differences. An even better experiment, though, seems to compare within one language differences in processing. I have designed a set of sentences where Dutch would show differences in processing and Frisian doesn't. If we get these results within one language as well as the expected differences between (7) and (9), we can double-check whether the results are due to the hypothesis we are proposing.
As we shall see shortly, the only sentences that qualify for our experiment
will be unequal in length. An ideal method for comparing such sentences is
the eye tracker. This allows detailed information on the online processing
of all the various subparts of a sentence. In paragraph
I will give some background information on the eye tracker.
15 native speakers of Dutch and 15 native speakers of Frisian will be tested. We have to pay attention to test only subjects that have experience in reading in their native language, otherwise slow processing could be ascribed to lack of reading skills. This is important especially for Frisian since, for most people, this language is only used in speech. Subjects are preferably inexperienced with linguistics.
On the basis of the theoretical background discussed in section
I have developed a set of basic sentences. For Dutch, I plan to test
sentences such as (7) compared to sentences where no syntactic
interpretation of zich is possible. This is the case in logophoric
constructions such as (14).
A possible factor that could influence processing is the mostly reflexive
meaning of the verb in (7). Wassen is mostly used in its reflexive
meaning although it can also be used in a purely transitive sense. This
could well have a positive influence on processing time of reflexive
zich / him. To exclude this possibility, I will use verbs such
as snijden (to cut) and verwonden (to hurt) which do not bias
interpretation.
In psycholinguistic literature, it is a given fact that the effects of
longer processing often show up only some time after the relevant word has
occurred. In our case, we need to make sure we can measure the effects some
time after zich / him. Therefore we need to include a
subordinate clause at the end of each sentence. This subordinate clause will
be the same in each pair of sentences and similar for all sentences except
for the fillers (see also note 6 on introductory stories).
One basic set of conditions for Dutch will thus be:
(13) Jan sneed zich nadat hij Piet had geslagen.
(14) Jan legde de bal naast zich nadat hij Piet had geslagen.
The same type of sentences can be used for Frisian native speakers. Thus we get:
(15) Jan snie him nei't er Pier slein hie
(16) Jan lei de bal neist him nei't er Pier slein hie
All the subjects will be tested for 5 different verbs per type of sentence, to exclude the possibility that the results are influenced by a different knowledge of one of the verbs. As such, in (13) and (15) I will use the verbs snijden, verwonden, beschilderen, inschrijven, verstoppen. In (14) and (16) I will use zetten, opzetten, planten, gooien, leggen. Each subject will be tested for 5 sets of two sentences, in random combinations. The sets of sentences will be randomly mixed with 50 percent filler sentences. In total, each subject will get 20 sentences to read. One possible set of combinations is displayed in the Appendix.
Subjects will not be informed in any way about the type of experiment or the goal of the experiment they are going to participate in. They will be told that they will be reading a set of Dutch / Frisian sentences and they randomly have to answer simple questions to test their understanding, not for competition in any way. It will be emphasized they should read the sentences at a natural pace. Furthermore they will get instructions to push a button after each story and after they have finished reading a sentence. All subjects will be seated in front of a computer screen. In the center of the screen the sentences will appear one by one, and the complete sentence will be displayed at once. Each sentence will be introduced by a short story which is unique for this particular sentence. This is needed to make sure the Frisian subjects get the reflexive reading, since the non-reflexive meaning is irrelevant for our purposes. More importantly, it could be that Frisian him takes more time if subjects have to choose in some way or another between the reflexive and non-reflexive reading. This could influence the results. To make the conditions completely parallel the same story will be displayed before the Dutch sentences, as well as before the Dutch and Frisian fillers. An example of a story can be found in the Appendix. After reading the story, subjects push a button and the sentence will appear. After completely reading the sentence, they again push a button, and at random a question will appear to make sure the sentences are read and understood.
The processes that occur during reading have been widely studied and reading methods have been used to study various syntactic phenomena and the parsing procedures that are involved in syntax. There are various methods available in principle to study the interpretation of reflexives, among them are word-by-word reading (subjects control the rate of presentation by pressing a button), rapid serial visual presentation (RSVP) of sentences (subjects are presented words at a set rate in the same spatial location) and completion responses (subjects silently read a passage and then make a standard word recognition response, such as naming or lexical decision, to a subsequent target word) (source: Gernsbacher 1994). These methods all have the disadvantage of using a more or less unnatural way of reading and thus run the risk of giving unnatural results. Eye tracking is a method that largely circumvents this unnaturalness. Subjects simply read sentences and have no additional artificially induced tasks to do. Monitoring subject's eye movements while reading does not perturb the speed or way they are reading. Another advantage is that this method provides a wealth of data on the online process of reading. Eye trackers can measure exactly where the eyes are looking during the reading process, thus allowing the researcher to find out exactly how long and how often a word is being fixated. Thus it is not necessary to draw inferences about comprehension processes on the basis of global data such as total reading time, for example. The eye tracker can determine exactly what it is in a sentence that causes a shift in the reading time.
Our hypothesis is that in Frisian, both sentences are interpreted with the same processing times since they are interpreted in the same way. In Dutch the logophoric sentences will take longer than the simple reflexive ones, since they are interpreted in the Interface which is costlier. Furthermore, we expect Frisian simple reflexive sentences such as (9) to take longer than their Dutch equivalents such as (7) since in Frisian they are interpreted in a costlier way. If the theory on Binding and Coreference is correct, and if the Economy rules are indeed at work here, we expect to find the following processing ``hierarchy'' for him and zich (where < means ``takes less time than'')
(17) Jan waste zich < Jan waskete him + Jan legt de bal naast zich +
Jan lei de bal neist him
In terms of the eye tracker, we expect to see results in the complete reading times to result from differences in fixation times on the reflexive. My hypothesis is that we find longer fixation on Dutch zich in the simple reflexive sentence, and a longer reading time for the remainder of the sentence, starting from the target area, the reflexive (see section 3.2: the effects of longer processing can show up in the region after the relevant word).
In this paper I have proposed an eye movement recording experiment to test processing of Dutch and Frisian reflexives. Since I have not yet carried out this experiment, I have no results to comment on. But we can reasonably expect, I believe, that we will actually find the results we are looking for. If this is indeed the case, it is good evidence from linguistic practice that the underlying theory seems to be on the right track, and what more could we ask for? I acknowledge that this study only looks at two languages, but if results match the theory, the road is open for further research. We expect that in any human language the facts can be described with a combination of limited minimalist Economy principles and a clear feature specification of the relevant elements. The question then is, can we find any language that does not match this kind of description? Another interesting question to ask is how children would score on the proposed sentences. Since it has been proposed in the literature that children at some point in their linguistic development have limited processing resources, it would be interesting to connect this proposal to our findings and see how children react on reflexives that are interpreted in discourse, C-I interface and syntax. If we strictly follow the hierarchy we would expect them to score best on syntax interpretation and worst on discourse interpretation, but then again there could be numerous other factors involved that have to be investigated. A further and similar question would regard aphasics: can we link aphasia deficits to strategies for interpreting anaphora in some way? To conclude, it is clear that the new view on the interpretation of binding and coreference offers promising new insights in the way the human language system works.
The fillers are in italics. The sentences will be displayed in random order and for each subject a different order will be used.
Dutch:
Frisian: vertalingen.
An introductory story with sentence (1) for Frisian / Dutch:
``Jan skylt in apel and krijt spul mei Pier. Jan wurdt tige lilk. Hy
slaat Pier. Pier rint lilk de keamer út. Dan krijt Jan it mês wêr en giet
fierder te skilen. Jan snijt him nei't er Pier slein hat.''
Sentence: Jan snijt him nei't er Pyt slein at.
``Jan is een appel aan het schillen en krijgt ruzie met Piet.
Jan wordt heel boos, hij slaat Piet! Piet loopt kwaad de kamer uit.
Dan pakt Jan het mes weer en gaat verder met schillen. Per ongeluk
gaat het mis!''
Sentence: Jan snijdt zich nadat hij Piet heeft geslagen.
Eric Auer 2003-06-18
![\includegraphics[width=10cm,keepaspectratio]{grammar-plaatje1.eps}](img8.png)
![\includegraphics[width=12cm,keepaspectratio]{grammar-plaatje2.eps}](img9.png)