We aim to model the results from a selfpaced reading experiment, which tested the effect of semantic type clash and typicality on the processing of German complement coercion. We present two distributional semantic models to test if they can model the effect of both type and typicality in the psycholinguistic study. We show that one of the models, without explicitly representing type information, can account both for the effect of type and typicality in complement coercion.
Complement coercion (The author began the book
→ reading the book) has been shown to cause an
increase in processing cost
A second influence on processing cost is the
thematic fit or typicality of the fillers of the verb’s
argument slots
Distributional semantic models (DSMs) can
successfully model a range of psycholinguistic
phenomena, including the effect of typicality
on complement coercion
In this paper, we consider experimental results from a study on complement coercion in German that manipulates both type and typicality. We discuss the performance of existing DSMs and a novel DSM combination. We also discuss how type information can be emerge from distributional information. 2
In a self-paced reading study on German complement coercion (Zarcone et al., in preparation), we have manipulated both type and typicality. The dataset consists of 20 pairs of subjets (S) and aspectual verbs (V). Each pair is combined with four nominal objects (O) in SOV order: [S Das Geburtstagskind] hat [O mit den Geschenken [S The birthday boy] has [O with the presents / der Feier / der Suppe / der Schicht] [V angefangen]. / party / soup / work shift] [V begun].
The objects are: a high-typicality entity (presents); a high-typicality event (party); a lowtypicality entity (soup); and a low-typicality event (work shift). The low-typicality objects are drawn from the high-typicality objects of other S-V pairs.
The self-paced reading study yielded the following significant effects: (1) an effect of typicality on reading times (t = 2.28, p = .02) at the object region (indicating subject-object integration), (2) an effect of object type on reading times (t = −2.5, p = .01) at the verb region (the region of the type clash), (3) an interaction of type and thematic fit at the verb region (t = 2.04, p = .04). Mean reading times per condition are reported in Table 1. In sum, the study shows that complement coercion involves both type and typicality. Thus, computational models of complement coercion need to account for both. 3
Distributional semantic models (DSMs) represent word meaning as high-dimensional vectors recording co-occurrences with elements of their high-fit entity high-fit event low-fit entity low-fit event
Object region mit den Geschenken with the presents 642 655 667 710 usage contexts. Semantic similarity is defined in terms of a vector similarity metric such as cosine.
Distributional Memory (DM, Baroni and Lenci
(2010)) is a DSM that includes syntactic
knowledge into the word representations. More
concretely, the TypeDM version of DM records
wordrelation-word tuples hw1 r w2i. The tuples are
weighted by Local Mutual Information
Complement Coercion and DSMs.
DM has been extended into the Expectation Composition and Update model (ECU, Lenci (2011)), a family of procedures that can be used to predict the typicality of one sentence part given other sentence parts. E.g., to model the typicality at the verb region in a German sentence with SOV word order (e.g. Das Geburtstagskind hat mit dem Geschenk angefangen / The birthday boy has with the present begun), ECU determines the thematic fit for the verb given subject and object: • compute an expectation for the verb given the subject s, as the distribution over verbs v defined by the weights of the tupleshs subj vi • compute an expectation for the verb given the object o, as the distribution over verbs v defined by the weights of the tuplesho obj vi. To combine the subject and object expectations, we combine the two distributions component by component, typically either by sum or products. This distribution is then represented in a vector space by computing the centroid or prototype of the vectors of the 20 most expected verbs. Finally, the thematic fit for a verbv given the subject s and the object o is its cosine similarity to the centroid. Subject
Object
Verb
JE
ECU. We call the models following the ECU
procedure SOV+ and SOV*, depending on their
combination operation (sum and product,
respectively). Simpler models only consider the
influence of subject or object on the verb (SV and OV
respectively), just by leaving out the combination
step. These models can successfully account for
reading time results on a dataset of complement
coercion in German that manipulates typicality but
not type
In order to test ECU on a dataset which manipulates both type and typicality, we evaluate the following ECU models on the complement coercion data in (Zarcone et al., in preparation): SO to model effects at the object given the subject; SOV+, SOV* and OV to model effects at the verb. We expect these models to account for the typicality effect at the object (1), but not for the type effects at the verb (2,3).
The results are summarized in Table 2 (left and middle). In accordance with our prediction, SO correctly yields the typicality effect at the object (F = 7.38, p < 0.01). Neither SOV+, SOV*, nor OV can model the type-typicality interaction at the verb (3). Surprisingly, though, SOV* and OV yield (2), an effect of type at the verb (F = 5.3228, p < 0.05 and F = 20.388, p < 0.001, respectively). Joint Expectations. The reading time study found that the subject-object typicality effects linger at the verb, interacting with type. The main shortcoming of ECU is its inability to model the typicality effects at the verb. This is due to the architecture of the SOV models (cf. Fig. 1, top): they compute the expectations for the verb first from the subject (SV) and update them with the object’s expectations (OV). They ignore the interaction between subject and object (SO) – the source of typicality effects (1,3) – corresponding to the assumption that this interaction should only matter at the object. In order to account for this, we draw an analogy to the concept of joint probability: P (S, O, V ) (1) effect of typicality at the object region (SO interaction) (2) effect of type at the verb region (type clash) (3) type x thematic fit interaction at the verb region non-compos.
SO OV X × × X × × SOV+ × × × which is equivalent (by the chain rule), to which we can interpret distributionally as motivation to reweight the typicality of the verb given the object with the typicality of the object given the subject, thus re-introducing the subject-object interaction into the verb prediction (cf. Figure 1, bottom).
In the Joint Expectation (JE) model, the thematic fit score assigned to the target verb is influenced both by the verb’s thematic fit with the object (the verb’s initial thematic fit score, equivalent to the ECU weight for the hobject obj verbi tuple) and by the object’s thematic fit with the subject (equivalent to the ECU weight for the hsubject verb objecti tuple), which in turn is used to reweight the verb’s score.
Similar to ECU, there is a choice of combination operations in JE (sum or product). Since JE can be formulated as a simple wrapper around ECU, ECU can be used to compute the individual components (e.g. SO, OV, or more complex ones) and these then just need to be combined additively (SO+OV) or multiplicatively (SO*OV).
The right-hand side of Table 2 shows the results for JE. SO+OV yields an effect of typicality (F = 6.777, p < 0.05) but no effect of type (2) or interaction (3). SO*OV yields two main effects of (2) type (F = 7.2359, p < 0.05) and typicality (F = 7.2359, p < 0.01), although no interaction (3).
Comparing the two models, we see that ECU SO accounts for the results obtained at the object (1), but the SOV models cannot explain the interaction with typicality on the verb (2,3). JE (SO * OV) models the effects of both type (2) and typicality at the verb, but does not (yet) account for their interaction (3).
We found that the SO model successfully accounts
for the effect of typicality at the object. This is
not surprising: one of the most typical tasks
successfully performed by distributional models such
as ECU is predicting verb-argument plausibility,
and ECU had already been successful in modeling
effects of typicality on reading times in German
complement coercion
On the other hand, the ECU SOV models were not able to account for the type–typicality interaction at the verb. The JE model (SO * OV), which we presented as an alternative to the ECU model to better account for the typicality effects at the verb, yielded effects of both type and typicality at the verb, but did not account for their interaction.
Our most surprising result is that the OV, SOV*, and SO*OV models explain the effect of type. As DSMs do not represent this concept explicitly, a possible interpretation suggested by our results is that type and typicality are not distinct categories, but capture properties of predicate-argument combinations at different granularity levels.
Distributional models can account for types
because they emerge from the observed corpus
distributions. Specifically, for the aspectual verbs used
in the present data set, the distribution over their
objects – namely that event nouns occur much
more frequently that object nouns
In conclusion, the picture that emerges from our experiments is one where (1) expectations for predicate-argument combinations have a hierarchical structure, with types as a high-level distinction and typicality as a low-level distinction, (2) both levels are different, but interact early during processing, influencing reading times, and (3) both type and typicality can emerge from the “same same” distributional model.
This research was funded by the German Research Foundation (DFG) as part of SFB 732 ”Incremental Specification in Context” and SFB 1102 ”Information Density and Linguistic Encoding”.