In this paper, we investigate German constructions composed of an adjective and a two-part nominal compound, such as katholisches Kirchenoberhaupt ('catholic church.leader'), focusing on two issues: (i) what are the prerequisites for semantically possible adjective-nominal compound constructions; (ii) which semantic factors determine the availability of bracketing paradox readings (i. e., the adjective modifies the first noun) in such constructions. We test theory-driven hypotheses using a corpus-based frequency model and evaluate the performance of the model with respect to human annotations.
Nominal compounds which are modified by an adjective, as in Ex. (1), can have
an iconic or an anti-iconic reading. Here, the phrase may refer to a church leader,
who is Catholic (iconic reading), or to the leader of the Catholic Church
(antiiconic reading). The latter reading is discussed in the literature as a bracketing
paradox (BP) because the semantic bracketing [[adjective noun] noun] does
not iconically match the structural bracketing [adjective [noun noun]], the
latter of which is invariant in German. Formally explicit approaches to this
phenomenon focus on systematically deriving their ambiguity ([2, 4]), but fail to
satisfactorily account for the puzzling ungrammaticality of Ex. (
Catholic church.leader
(
We investigate such German constructions composed of an adjective (A) and a two-part nominal compound (N1N2), focusing on two issues: (i) what are the prerequisites for semantically possible A-N1N2 constructions; (ii) which semantic factors determine the availability of the iconic or anti-iconic reading. We analyze these constructions from a theoretical perspective (Section 2) and test the theoretical assumptions using a corpus-based frequency model (Sections 3 and 4). In a second step, the performance of the model is evaluated with respect to human annotations (Sections 5 and 6). 2
The perceived ungrammaticality of Ex. (
This restriction does not make any assumptions regarding the distinction between the iconic or anti-iconic interpretation of A-N1N2 constructions. We hypothesize that this distinction is based on the relative semantic plausibility ([5],[8]) of the A-N1 and A-N2 constructions. We formulate this intuition as Hypothesis 2 (H2): H2: for examples where H1 holds, the higher the semantic plausibility of A-N1 relative to A-N2 is, the more likely it is that A-N1N2 is a bracketing paradox.
The effects of H2 are exemplified by Ex. (1), (
(
Catholic table.prayer
(
Catholic company.leader 3
We verify H1 and H2 using a frequency-based model derived from the 11.6 billion tokens decow14ax corpus [6]. The model considers the lemmatised form of the words for computing the frequencies. For an A-N1N2 construction we compute the following: f reqA N1 , the frequency (number of corpus occurrences) of A-N1 f reqA N2 , the frequency of A-N2 f reqA N1N2 , the frequency of A-N1N2
f reqA N1 , the relative frequency of A-N1 and A-N2.
r fA N1N2 = f reqA N2
We use the frequency of a construction in the corpus to model the notions of semantic possibility and semantic plausibility and to make judgments regarding the two hypotheses formulated in Section 2. If the frequency of a construction is higher than 0, the construction is considered semantically possible. We consider constructions that do not occur in the corpus to be semantically impossible.
The semantic plausibility score for an adjective-noun pair is given by the frequency count. The relative semantic plausibility score is the relative frequency of the two adjective-noun pairs in an A-N1N2 construction ( ff rreeqqAANN12 ). 4
We test our two hypotheses using a dataset of 198 A-N1N2 constructions compiled based on the theoretical literature.
To test H1, an A-N1N2 construction is considered semantically possible if its
corpus frequency is greater than 0. The dataset contained 77 semantically
possible A-N1N2 constructions (i. e., constructions that actually occurred in the
corpus1). For 70 of these examples, our model predicted A-N2 being also semantically
possible, resulting in a 90.9% prediction accuracy for H1. An interesting case is
the phrase in Ex. (
For H2, the model computes a relative semantic plausibility score for each semantically possible A-N1N2 construction in our dataset. We identify this score with the BP-ness of an A-N1N2 construction. An initial inspection of the constructions with a high relative semantic plausibility score shows that these are indeed 1The discrepancy between the initial and the attested number of constructions can be explained as follows: on the one hand, many of the examples were ungrammatical contrasts to grammatical examples; other examples were constructed by analogy with existing examples, which of course does not imply their actual occurrence in the corpus. constructions with an anti-iconic reading: katholisches Kirchenoberhaupt has a score of 1328, europäischer Auslandsaufenthalt (‘European foreign-country.stay’) has a score of 7327. In contrast, constructions with a very low score, like verrückter Chemieprofessor (‘crazy chemistry.professor’, score 0:003) or ambulante Unfallbehandlung (‘ambulant accident.treatment’, score 0:0003), clearly have no anti-iconic reading. Borderline cases include bedrohliches Krankheitssymptom (‘menacing desease.symptom’, score 4:86) and politische Satiresendung (‘political satire.broadcast’, score 1:81).
The model confirmed H1, and thereby provided good evidence for the assumption that for all A-N1N2, A-N2 must be semantically possible, irrespective of whether A can modify N1 or not. The initial observations also suggest that the frequency model can be successfully used to test H2. To test H2, we annotated the set of 77 A-N1N2 constructions with regard to their perceived iconicity. The annotation is described in the next section. 5 5.1
The dataset contained those 77 A-N1N2 constructions that are semantically possible according to the frequency model. These items were annotated by 5 PhD students in linguistics (3 women, 2 men; native speakers of German); two of them are co-authors of this paper. The items were presented to the annotators in an Excelspreadsheet in one of 5 randomized orders. The annotators worked independently from each other.
The annotators annotated each item according to the following two questions2: Q1 Is the A-N1N2 construction as a whole grammatical? yes/no
Q2 Which reading is preferred? anti-iconic, iconic, equal preference 5.2
Grammaticality (Q1) We categorized the 77 items according to the annotation question Q1. As a prerequisite, at least 4 (out of 5) annotators had to agree upon an answer. If there was no corresponding agreement, the item was not categorized. 73 out of 77 items are judged as grammatical. 2 items are judged as ungrammatical, and 2 items could not be categorized; these 4 examples were excluded from the evaluation of Q2. The inter-rater agreement had a Fleiss’ k value [3] of 0.45 (moderate agreement). Notably, the 2 items that were judged as ungrammatical are vierstöckiger Hausbesitzer and verregnete Feriengefahr. These are exactly those cases that are falsely classified as semantically possible by the frequency-based 2We also asked the annotators whether A-N1 and A-N2 are semantically possible. As the answers do not directly bear on H2, we will not report them here. model as they occur in metadiscussions on their semantic impossibility (see Section 4).
Preference (Q2) We categorized the remaining 73 grammatical items according to the annotation question Q2. As a prerequisite, at least 4 (out of 5) annotators had to agree upon an answer (if all 5 annotators judged the item as grammatical), or, at least 3 (out of 4) annotators had to agree upon an answer (if only 4 annotators judged the item as grammatical). 11 items could not be categorized, as the required majority was not obtained. From the remaining 62 items, 16 were perceived to have an anti-iconic reading, while the other 46 were perceived to have an iconic reading. The inter-rater agreement had a Fleiss’ k value of 0.58 (moderate agreement). The results yield a two-way distinction between anti-iconic and iconic readings. Notably, no items were annotated as being truly ambiguous. However, for 11 items, the annotators did not agree upon a preferred reading. Among these, several examples are prototypes for bracketing paradoxes according to the theoretical literature; in fact, 4 have a tendency for being ambiguous (e. g., politische Satiresendung) and 2 have a tendency for being anti-iconic (e. g., katholisches Kirchenoberhaupt). As this result is in need of further clarification, we excluded the problematic data points from the evaluation of the frequency-based model. The remaining 62 items annotated for iconicity will be further used to train and test the frequency-based model. 6
This section presents the results of using the frequency-based semantic plausibility model introduced in Section 3 to predict the iconicity of A-N1N2 constructions. The 62 annotations which were considered to be grammatically correct (Q1) and were assigned the same preferred reading by the majority of the human annotators (Q2) are used as a dataset.
The task at hand is to predict if a particular A-N1N2 construction is considered a bracketing paradox or not using only the frequency information, in particular the relative semantic plausibility score which we normalize across the examples in our dataset.
We use logistic regression, a widely-used linear machine learning method, to train a prediction model. Table 1 reports the average F1 score [7] and accuracy figures obtained for 10-fold cross-validation. The model hyper-parameter (the regularization coefficient) is chosen individually for each fold, using a grid search over 10 equally spaced values in the interval [1e 4, 1e + 4]. The results show that despite the imbalanced number of instances for each class, the relative semantic plausibility score is a very good predictor for the preferred interpretation of a particular construction.
16 BP, 46 non-BP (62 total)
0.90
95.71 The corpus-based frequency model confirmed both H1 and H2. First, for all AN1N2 constructions, A-N2 must be semantically possible, irrespective of whether A can modify N1 or not (H1). Second, the higher the relative semantic plausibility score of A-N1N2 constructions, the more likely it is that the construction is a bracketing paradox (H2). This result is based on our evaluation of the frequency-based model using human annotation as a gold standard.
Our study also pointed to issues that need to be further investigated in future work. From a theoretical perspective, it is surprising that katholisches Kirchenoberhaupt, the prototypical textbook example for bracketing paradoxes, received mixed ratings from the annotators. This suggests that the distinction might not necessarily be a binary one. We plan to complement our results via a rating study that elicits graded judgments for the perceived iconicity of A-N1N2 constructions. From the perspective of the computational modeling, we discovered some limitations of the frequency-based model. For example, in the construction intelligenter Tierarzt ‘intelligent animal.doctor’, the pair intelligenter Arzt is very infrequent, as the adjective ‘intelligent’ spells out an implied attribute of the ‘doctor’, whereas the locution intelligentes Tier is much more frequent. Thus, the model predicts that this construction should have an anti-iconic interpretation (‘doctor for intelligent animals’), which is clearly wrong. Another shortcoming relates to the inability of the model to make predictions in absence of the frequency information, which resulted in analyzing only a part of the initial dataset. In order to circumvent these shortcomings, we plan to use distributional semantics models, which have the ability to integrate information about the semantics of the construction’s constituents.
Financial support for the research reported in this paper was provided by the German Research Foundation (DFG) as part of the Collaborative Research Center “The Construction of Meaning” (SFB 833), projects A1 and A3. We thank the anonymous reviewers for their comments and our annotators for their support. [1] Rolf Bergmann. Verregnete Feriengefahr und Deutsche Sprachwissenschaft.
Zum Verhältnis von Substantivkompositum und Adjektivattribut. Sprachwis