Simone Teufel, Elina Feleki
Centre for Cognitive Science
University of Edinburgh
2 Buccleuch Place
Edinburgh, EH8 9LW
April 19, 1996
Recent theories of sentence production distinguish the level of grammatical function assignment from the positional level, where the concepts expressed by certain grammatical functions are linearly ordered. [Bock and Warren(1985)] show the influence of conceptual accessibility, a semantic property, on the assignment of grammatical function for English. We report a similar experiment for German, which is a free word order language. Our results confirm [Bock and Warren(1985)]'s basic hypothesis. We speculate about the potential of German syntax to settle a more controversial question: the potential influence of conceptual accessibility on positional ordering. We suggest further research that could settle the question more convincingly once free of the restrictions of English syntax.
Sentence production is the conversion of thoughts into sentences, the mapping of cognitive concepts into a linear order of speech sounds. It involves a number of different levels that interact to produce the linguistic output: syntactic, semantic, lexical and phonological. Understanding the way these levels interact and the implications of this interaction for the architecture of the mind has been the target of much research in linguistics, psycholinguistics, and cognitive psychology. We will concentrate on the interaction between the syntactic and semantic levels in the production of German sentences.
Recent models of sentence production ([Garrett(1975)], [Garrett(1980)], [Bock and Levelt(1994)]) assume assignment of conceptual elements to grammatical roles at a different point in processing from the determination of word order. These models include the following levels:
This section provides a summary of Bock's and Warren's theoretical assumptions and their experiment.
Various experiments have shown the special prominence of surface subjects: they tend to be more animate, concrete and imageable. But there is a complete hierarchy of grammatical relations, descending from subject to (direct) object, indirect object, oblique object, object of comparison. [Keenan and Comrie(1977)] introduce the notion of a hierarchy of NPs based on relativizability: If a language permits a relative clause to be formed on a noun phrase associated with a grammatical function low in the hierarchy, it will permit relativization on NPs representing all grammatical functions above it. They argue with crosslinguistic evidence as well as with a within-language ordering of acceptability. For example, the man who Sandy is taller than is less acceptable than a relative clause where the relativized part is the direct object. [Pullum(1977)] argues that the hierarchy is reflected in the dominant NP constituent orders of the world's languages, where only the position of the verb varies.
[Bock and Warren(1985)] define conceptual accessibility as the ease with which the mental representation of some potential referent can be activated in, or retrieved from, memory. Arguments from language acquisition ([Keil(1979)]) predict a link between the development of knowledge about entities and conceptual roles, and higher-level grammatical relations. Bock's and Warren's hypothesis is that concepts at lower levels of ontological hierarchy (i.e. the hierarchy of conceptual accessibility) should be those that are most commonly used in higher-level grammatical roles:
Findings from child development and the acquisition of ontological knowledge support this hypothesis: Prototypical sentence subjects are actors, agents, moving objects. Visual attention is centred on the actor ([Robertson and Suci(1980)]); it facilitates children's learning of word-referent relationships for words denoting agents and actors. Children also have more knowledge about animate than about inanimate objects. [Keil(1979)] stresses that knowledge of the physical object category precedes knowledge of the event category, from which abstract objects emerge. As a result, the adult's knowledge of lower (more concrete) levels in the hierarchy of conceptual accessibility are richer than higher levels (events and abstractions).
[Bock and Warren(1985)] used imageability as an index for conceptual accessibility, partly because of the ready existence of imageability norms ([Paivio et al.(1968)]). According to Bock and Warren, objects which belong to lower levels in the predicability/imageability hierarchy are more accessible in the sense that it is easier to remember them, and talk about them as well as to think of associates for them. Moreover, differences in imageability have been shown to influence selection of surface subjects.
Bock and Warren tried to find variations in syntactic structure of sentences which would provide evidence for the conceptual pressure of assigning a higher grammatical function to more accessible nouns. Their experiment made use of 3 types of sentences:
Simple transitive declaratives & passivization:
Double dative constructions:
Each sentence contained one highly imageable and one less imageable noun. The first two sentence pairs were used to examine the effects of imageability on the placement of major constituents representing three levels in the relational hierarchy (subject, direct and indirect object). Subjects were presented with target sentences and were asked to recall them later. Inversions of the noun phrases in the subjects' responses to the target sentence were interpreted as an effect of conceptual accessibiliy on grammatical function assignment. Phrasal conjuncts (third sentence pair) were used as control for word order effects: simple leftward movement is possible without any change in surface grammatical roles. If variations in the syntactic structure of the other two sentence types resulted indirectly from word order effects and not from effects of conceptual features on grammatical roles, then phrasal conjuncts (which both shared the same grammatical role) should show effects of conceptual accessibility that are comparable to those for simple declaratives and datives.
The results confirmed the hypothesis that subjects produce significantly more inversions in simple declaratives and datives, placing the more imageable noun before the less imageable one rather than vice versa. They interpreted this as evidence that conceptual accessibility is closely related to the hierarchy of grammatical relations. Having found no such effect for phrasal conjuncts, they concluded that word order is only indirectly affected by conceptual accessibility. In other words, Bock and Warren maintain that conceptual effects on syntax take place at functional level only. We will return to this highly controversial point in our discussion.
The theory about sentence production recapitulated in a previous section ([Garrett(1980)]) claims universality. Bock's and Warren's experiment only speaks for English, a fixed word order language, where it is difficult to separate effects on grammatical functional level from effects on positional level. We want to address two questions:
Our experiment, described in the next section, addresses the first question. We have not conducted experiments towards the second question but we will present a proposal for such research later.
In our experiment we used the concept of animacy as an index for conceptual accessibility. Animacy, and more specifically ``humanness'', is highly related to this notion: at a very early stage in conceptual learning, children combine animate entities (especially humans) with the idea of agency (especially human agents, whose range of action is even higher than that of animals); human entities, as explained before, therefore appear particularly low in the imageability hierarchy.
Our experiment was not an exact duplication of Bock's and Warren's experiment. First, we did work with spoken but rather with written language (due to practical pressures). Second, we did not use sentences with double object constructions or phrasal conjuncts, because there are better ways in German to experiment with word order. Rather, we used the active/passive variation in order to compare the effect of animacy on grammatical function of animate nouns vs. inanimate nouns. Our hypothesis is that subjects will prefer to assign the subject to animate nouns, even if this implies that they will have to convert active sentences to passives.
Two pairs of active/passive sentences were constructed out of each pair of animate/inanimate target nouns. Each item appeared in four conditions. The verb was held constant across all four sentences, and each sentence contained one animate and one inanimate entity. The first pair of sentences described an action where the animate entity was the agent and the inanimate entity the patient whereas in the second pair the reverse relationship held. Each pair of sentences expressed the same proposition, but with different syntactic constructions (one passive, the other active). The following tables give an overview of the four conditions:
------------------------------------------------------------ | | Animacy ----------+----------+-------------------------------------- | | Animate Agent Inanimate Agent ----------+----------+-------------------------------------- Sentence | Active | A B | | Type | Passive | D C ------------------------------------------------------------
----------------------------------------------------------- Condition | A | B | C | D ----------+-----------+-----------+------------+----------- Agent | Anim | Inanim | Inanim | Anim Patient | Inanim | Anim | Anim | Inanim ----------+-----------+-----------+------------+----------- Subjekt | Anim | Inanim | Anim | Inanim ----------+-----------+-----------+------------+----------- Type | Active | Active | Passive | Passive ----------+-----------+-----------+------------+----------- Example | Man | Barrel | Man | Barrel | squashed | squashed | squashed | squashed | barrel | man | by barrel | by man -----------------------------------------------------------
We should remark here that the only parameters we varied were animacy of agent (animacy of patient followed automatically as opposite) and sentence type. Whether or not the subject of the sentence was animate or not followed as a function of these 2 parameters. Crucially, conditions A and D are paraphrases of each other, and so are conditions B and C. We were interested in whether our test persons would recall sentences unchanged or whether they would prefer the paraphrase of the sentence (i.e. D for A, A for D, B for C or C for B), thus assigning target nouns the opposite grammatical functions. Our hypothesis was that the test persons will prefer to assign the grammatical role of subject to animate nouns, even if they are not the agents of the sentence, and even if this implies that they will have to passivize the sentence. We therefore predict that the likelihood of inverting conditions B and D (where the animate target is not realized as subject) is higher than the likelihood of inverting conditions A and C (where the animate target is realized as subject).
The two variables we manipulated were sentence construction type (active vs passive) and animacy (animate agent vs inanimate agent), resulting in a 2x2 experiment.
32 items were construed in these 4 conditions (that is a total of 128 sentences). As explained above, in two of these conditions, the animate target was the subject, whereas in two of them it was the direct object. The sentences differed only in the order of noun phrases.
The sentences consisted of a matrix sentence expressing an action of saying/thinking/believing, and a subordinate transitive clause (always introduced by ``daß''). The sentences we were interested in were the subordinate clauses; the matrix sentences were used in our recall/production experiment as prompts. Targets were the subject and the object of the subordinate clauses, where it was always the case that the one was animate and the other inanimate, as the following example item demonstrates.
Four different lists of sentences were created, where each list contained only one of the four conditions of each of the 32 items. The order of the sentences was randomised. Each list contained an equal number of sentences in each condition. The resulting 48 sentences per list were divided into 8 blocks of 6 sentences. Our fillers were intransitive clauses, e.g.
Der Doktor sagte, daß die Krankenschwester um 5
nach Hause gegangen war.
The doctor said that the nurse had gone home at 5.
Fillers were identical for all lists. Each block contained 2 fillers. The order of fillers and items within blocks was random, but it was kept constant across all four lists. Not more than 2 of the same conditions occurred per block, but two sentences of the same conditions were allowed to occur consecutively. A prompt list containing the matrix sentence up to and including the word ``daß'' from the paired sentences was constructed for use with all four sentence lists. It was also divided into 8 blocks, but within blocks, the prompts occurred in an order different from that of the corresponding sentences in the sentence list.
Each sentence list was presented to 5 subjects. There were 2 within-subjects and within-items factors (construction type and animacy), with each subject receiving 8 sentences in each of the 4 conditions. Each item was presented to all 20 subjects in one of the four conditions.
A pilot study was conducted in order to find the correct timing and to detect ambiguities, pragmatic anomalities or other factors which could interfere with the results. 8 subjects were examined. They were given the blocks of sentences (either blocks of 6 or of 8 sentences and were allowed different time periods, ranging from 45 seconds to 2 minutes. The correct timing is crucial in a recall/sentence production experiment of this kind; too much time would have allowed the subjects to recall the sentences by rote and too little time would have placed too much demand on subjects' memory, thereby leading to overall failure of recall. Thus, we had to make sure the task was adequately demanding and that a sufficient amount of time elapsed between the reading and the reproduction of sentences in order to make sure that subjects reconstructed the sentences from semantic representation rather than from surface rote memorization (memorization of the exact syntactic structure of the sentences). Furthermore, the pilot study helped to identify pragmatically marked and ambiguous items which were removed. Moreover, some sentences had to be shortened. In general, it was made sure that there were no points that could lead to confusion or uncertainty about the intended meaning of the sentences.
The participants were 20 German native speakers between 20 and 30 years old, recruited from the Edinburgh University community. They were given a chocolate bar, either Mars, Bounty, Twix or StarBar. :-)
Subjects were given one minute to read one block of sentences. After each block, the experimenter asked them to do a simple 3 digit by one digit multiplication. Then, subjects were given the prompt list for that block and were asked to reproduce the six sentences from memory. To write down the sentences, they were given as much time as they wished (usually between 2 and 4 minutes). Subjects were run in groups of one to four. They were told that the experiment was a memory task on sentences, and the instructions emphasised remembering the ideas expressed in the sentences rather than the exact wording.
Like in Bock's and Warren's experiment, four categories were employed in scoring the recalled sentences, including corrects, inversions, errors and omissions. To be considered correct, a recalled sentence had to preserve the basic syntax and word order of the presented sentence. Thus, the same or similar phrases had to appear as subjects or objects. Changes in tense, number, definiteness were permitted as well as synonym substitutions and minor deletions or additions that did not alter the major grammatical relations of the sentence. We did not score sentences as correct when the noun phrases were in the correct order, but there was just an ``X'' or ``...'' representing one of the noun phrases (which occurred several times).
An inversion was scored when the sentence met all the criteria for a correct sentence, but was presented in the alternative form that represented the same meaning (condition D instead of A, or C instead of B, or vice versa). Note that preservation of the same semantics was a vital condition here. All other responses were scored as errors. If there was no answer, or at least one of the 3 vital parts of the sentence (verb, object, subject) was missing, an omission was scored.
Only the correct and the inverted sentences were taken into account in the evaluation of the results - i.e. we performed the analyses on proportions of inversions with respect to the total number of corrects and inversions (discarding errors and omissions). This proportion expresses the tendency to invert a sentence, given that the basic idea of the sentence was remembered.
These data were evaluated in analyses of variance that treated both subjects and items as random factors. Overall, there were significantly more inversions when the effect of the inversion was to make the animate target the subject (12.50% for condition B, 28.45% for condition D) than when the effect was to make it the object (4.35% for condition A, 6.45% for condition C).
------------------------------------------------------------ Condition | A | B | C | D -----------+-----------+------------+------------+----------- presented | Man | Barrel | Man | Barrel as | squashed | squashed | squashed | squashed | barrel | man | by barrel | by man -----------+-----------+------------+------------+----------- recalled | Barrel | Man | Barrel | Man as | squashed | squashed | squashed | squashed | by man | by barrel | man | barrel -----------+-----------+------------+------------+----------- Avg. % | 4.35 | 12.50 | 6.45 | 28.45 inversions | | | | -------------------------------------------------------------As predicted, we found a main effect of animacy (F1 (1,19) = 17.20, p < .001 for subjects, F2 (1,31) = 23.67, p< .0001 for items), which we interpret as a preference to express the animate target as a subject.
We also obtained a main effect of sentence type (F1 (1,19) = 9.02, p < .008 for subjects, F2 (1,31) = 6.00 , p < .02 for items), even though this result was not a primary goal of our study. What this amounts to is that subjects are more likely to invert a sentence to active if it was in passive beforehand, i.e. a strong overall tendency to produce active sentences.
We found an interaction between sentence type and animacy (F1 (1,19)= 11.99 , p < .01 for subjects and F2 (1,31)= 7.19 , p < .01 for items), indicating that animacy has a stronger effects on passives than on actives, as expected.
We will start this section by giving some interpretation of the results and commenting on our methodology and alternative experimental setups. Then we will elaborate on our main point here, possible experiments to address the question of conceptual influence on word order.
As the results show very strong tendencies, we can but repeat that they provide additional evidence that the conceptual accessibility is closely related to the hierarchy of grammatical functions for German also, in the line of [Bock and Warren(1985].
A difference in design in our experiment to theirs concerns the active/passive alternation. As opposed to their experiment, where the active/passive difference was just a between-item factor, we have designed it as an within-item factor, allowing us to compare all four conditions (Bock and Warren had only counterbalanced active and passive sentences). As a consequence, they had to assume that actives are the preferred structure. On the other hand, the fact that we had 4 conditions in our experiment helped us prove this assumption (though this is more a side effect of the experiment). The strong inversion pattern from passive to active that was extracted from our experiment was anticipated. The use of passive involves a pragmatic factor that cannot easily be reproduced under the conditions of the experimental environment. The passive involves an intention on the part of the speaker to place more emphasis on the patient and much less on the agent, there being cases where the agent is completely omitted. Taking into account that these factors which could trigger the special usage of passive are absent in the experimental environment of sentence production, it is not surprising that more inversions occurred from the passive condition to the active: it simply confirms that active is the default. Therefore, the interaction we found (the fact that there were significantly more inanimate than animate passives converted into active) is not surprising: the inclination of subjects to make the animate target the subject became even stronger when the result was an active sentence.
The data does not allow conclusions about effects of animacy on word order. This is because in our sentences, subjects are always the first NP, and objects are always the second NP. Claims about word order effects can only be made if there are dislocated NPs.
One of the greatest difficulty in the design of the materials of our experiment was isolating verbs that would allow an action to involve the possibility of both pairs of animate-agent/inanimate-patient and inanimate-agent/animate-patient alternatively, holding the target NPs constant. Most verbs in the vocabulary of a language describe actions which are performed by animate agents.
Of the actions that can be performed by inanimate agents, there is only a restricted number of transitive verbs where the object can be animate. We have classified the actions we found to obey these criteria:
These rigid semantic restrictions may in some cases have led to awkward sentences in the absence of context. However, such cases were limited in number and we don't think this factor influenced the experiment significantly. Our sentences were written rather than spoken, thus they included no prosodic cues such as stress and intonation that can make sentences sound more natural. In any case, we are confident that in spoken language there would not have been any grounds for questioning semantic soundness.
The passivization of certain verbs involved a small number of problematic cases as well (e.g. German dative objects can't be passivized). First of all, in the cases where the logical subject was the inanimate entity, it could not always be expressed through a von-phrase; some of the passivized verbs would only subcategorize for a durch-phrase:
Der Feldwebel betonte, daß der Soldat (*von dem/ durch den)
Bunker gesch\"utzt wurde.
The sergeant emphasized that the soldier was protected by the bunker.
This is an indication that the true thematic role of the inanimate entity was not agent but rather instrument. This weakens our point a bit, as we are depending on the notion of agency; however, this was only the case for a small number of sentences
Another point where we were not sure about the validity of the concept of animacy as a measure of conceptual accessibility occurred during the design of the experiment. We originally wanted to include action taking place between individuals and institutions, as in
Die Handelskammer gab an, daß der Architekt die Firma
The chamber of commerce explained that the architect had hired the company.
We did not use them in our final materials, as we suspected that even though the institution nouns are abstract, the fact that they represent groups of humans makes them similar in animacy/imageability to concrete human agents. As we had no possibility to settle this question, we decided conservatively to exclude these sentences.
In order to make claims about whether or not semantic factors influence the determination of word order (positional level), it is necessary to isolate the effects of word order from the effects of grammatical function assignment. For fixed order languages, this is impossible: in English, the subject is always the first NP in a sentence (with the exception of constructions such as topicalization). The only way to make any claims about word order effects in English is by looking at conjunct order like Bock and Warren did. We believe that their argument is not very strong; apart from the fact that all they can show is a null effect, the variables grammatical function and word order are not independent in their setup. We believe that the syntax of certain other languages allows for more insightful experiments.
Languages with richer morphology usually also employ more freedom with respect to word order; as the morphological marking on NPs flags the thematic role of that NP it can be dislocated (e.g. scrambled, extraposed or topicalized) without making the sentence ambiguous. In German, for example, sentence 3 is perfectly grammatical; that is, it is possible to place a given target (here: den Mann) in front, even though it is not in the most salient grammatical function, but rather an accusative object.
Er sagte, daß den Mann der Schlag getroffen hat.
he said that the man(acc) the blow(nom) hit had.
There is indeed considerable evidence from other free word order languages such as Catalan, Spanish, cf. [Prat-Sala et al.(1996)]) that the positional level is also influenced by conceptual accessibility.
We therefore suggest a 2x2-experiment manipulating the two factors grammatical function and word order. We want to clarify if there are situations where the test persons violate the condition that animate targets are expressed as subjects, in favour of placing them prior to other NPs in the sentence. We want to find out what is more attractive to the test persons, to make the animate target the subject or to position it as the first noun phrase (in whatever grammatical role).
The proposed sentences all have the following properties:
----------------------------------------------------------- Condition | E | F | G | H ----------+-----------+-----------+------------+----------- Agent | Anim | Inanim | Inanim | Anim Patient | Inanim | Anim | Anim | Inanim ----------+-----------+-----------+------------+----------- Subject | Anim | Inanim | Inanim | Anim ----------+-----------+-----------+------------+----------- First | Anim | Inanim | Anim | Inanim Position | | | | ----------+-----------+-----------+------------+----------- Pref. | | | | grammat. | + | - | - | + function | | | | ----------+-----------+-----------+------------+----------- Pref. | | | | word | + | - | + | - order | | | | -----------------------------------------------------------
Conditions E and H are variants of each other, only distinguished by dislocation, and so are F and G.
If there is no influence on word order, conditions E and H should be the preferred conditions, as they express the animate target as the subject (even though dislocated in condition H); we would expect to get more inversions for conditions F and G. In addition, E and H should have about the same number of inversions, and similarly F and G, as they have the same subject/object structure.
If there is influence on word order, condition G should be attractive as well, since it places the animate target in front (even though not in subject role). This should result in more inversions for F than G. Clearly, in both cases, F will result in more inversions than E, as E is a more favourable situation with respect to both word order and grammatical function, and F is less preferred under both conditions.
The two factors compete when G and H are compared. The question then is: Which of the two factors is stronger? Can preference for word order override the preference for high grammatical roles?
Similar experiments for Spanish and Catalan have been conducted recently ([Prat-Sala et al.(1996)]), with one difference being that they did not involve recall tasks. Prat-Sala et al. utilised pictures instead of sentences and asked subjects to describe the actions depicted in the words, thus avoiding a bias towards a particular type of target sentence. Thus there cannot be doubt that syntactic structure was not computed afresh but only memorized, which is always a possibility with experiments conducted along our lines. Prat-Sala et al.'s results illustrate that conceptual accessibility does indeed influence word order for these languages.
Even though we first designed our experiment along these lines, we later felt that the sentences with dislocation were too marked to present them to subjects without context (which we could not provide due to time pressure), and therefore we decided on the active/passive variation.
Context for these sentences can be provided by making the second NP the topic, either via stress or by adding a contrastive phrase:
daß der Golfer von dem Ball getroffen wurde
that the golfer was hit by the ball
The stressed item should be the last one, which is consistent with the accepted notion that old information usually precedes new information: in sentence 7 the NP the ball represents old information and the golfer is the new information, and hence it's the golfer which should be stressed. Maybe it would also be necessary to set the context by introducing the agents before the relevant item : On the golf course, there was a golf player, practising with a golf ball, and a basketball player, playing with a basketball. Overall, the sentences seem very much improved by manipulation of the context.
What would it mean for theories of sentence production if the influence of semantic factors on word order could be proved, especially if it were additionally proved that word order can override grammatical function? This is a tricky question. Do we have to assume different sentence production mechanisms for fixed and free order languages, or does the model need to be changed? These are open questions.
Our experiment has brought additional support for the hypothesis expressed in [Bock and Warren(1985)], namely that there is a strong effect of conceptual accessibility on grammatical function assignment. However, we suspect that there is semantic influence on word order. We believe that further experimental research conducted in free word order languages will have more to say on this matter.
In this situation, prosodic features are the sole determinants of the meaning of a sentence, a further argument for preferring spoken language for further experiments.