Studies across multiple languages show that overt morphological priming leads to a speed-up only for transparent derivations but not for opaque derivations. However, in a recent experiment for German, Smolka et al. (2014) show comparable speed-ups for transparent and opaque derivations, and conclude that German behaves unlike other Indo-European languages and organizes its mental lexicon by morphemes rather than lemmas. In this paper we present a computational analysis of the German results. A distributional similarity model, extended with knowledge about morphological families and without any notion of morphemes, is able to account for all main findings of Smolka et al. We believe that this puts into question the call for German-specific mechanisms. Instead, our model suggests that cross-lingual differences between morphological systems underlie the experimentally observed differences.
Priming is a general property of human language
processing: it refers to the speed-up effect that
a stimulus can have on subsequent processing
A recent study by Smolka et al. (2014) investigated overt morphological priming on prefix verbs in German, where the base verb and derived verb can be semantically related (transparent derivation: schließen – abschließen (close – lock) ) or not (opaque derivation: fu¨hren – verfu¨hren (lead – seduce)). Experiment 1, an overt visual priming experiment (300 ms SOA) involved 40 six-tuples that paired up a base verb with five prefix verbs of five prime types (see Figure 1). The verbs were normed carefully, e.g., for association, to exclude confounding factors. The authors reported three main findings: (a), no priming for Form and Unrelated; (b), no priming for Synonymy; (c), significant priming of the same strength for both Transparent and Opaque Derivation.
These findings suggest that morphological
priming on German prefix verbs use a mechanism that
is different from lexical priming, which assumes
that the strength of the semantic relatedness is the
main determinant of priming – i.e., lexical
priming would predict finding (a), but neither (b) nor
(c). The findings by Smolka et al. are also at odds
with overt priming patterns found in similar
experimental setups for other languages such as French
Our paper investigates this claim on the computational level. We present a simple model of corpusbased word similarity, extended with a database of morphological families, that is able to predict the three main findings by Smolka et al. outlined above. The ability of the model to do so, even though it operates completely at the word level without any notion of morphemes, may put into question Smolka Target binden (bind) 1 Transparent Derivation zubinden (tie) 2 Opaque Derivation entbinden (give birth) 3 Synonym zuschnu¨ren (tie) 4 Form abbilden (depict) 5 Unrelated abholzen (log) et al.’s call for novel morpheme-level mechanisms for German. 2
We model the priming effects shown in Smolka et al. by combining two computational information sources: A distributional semantic model, and a derivational lexicon.
Distributional Semantics and Priming.
Distributional semantics builds on the distributional
hypothesis
A classical test case of distributional models is
exactly lexical priming, which has been modeled
successfully in a number of studies
priming DISTSIM(p, t) ∝ cos p~, ~t
Regarding morphological priming, this model
predicts the result patterns for French and Dutch but
should not be able to explain the German results.
Derivational Morphology in a Distributional
Model. In Pado´ et al. (2013), we proposed to
extend distributional models with morphological
knowledge in the form of derivational families D,
that is, sets of lemmas that are derivationally
(either transparently or opaquely) related
The MORGEN model should have a better chance of modeling Smolka et al.’s results than the DISTSIM model. Note, however, that it remains completely at the word level, with derivational families as its only source of morphological knowledge. 3
Setup. We compute a DISTSIM model by
running word2vec
Following Smolka et al., we analyze the predictions with a series of one-way ANOVAs (factor Prime Type with reference level Unrelated). As appropriate for multiple comparisons, we adopt a more conservative significance level (p=0.01). Results. Table 1 reports the experimental results and model predictions (average experimental reaction times, cosine model predictions, and significance of differences). Model contrasts that match experiment contrasts are marked in bold.
As expected, DISTSIM predicts the patterns of classical lexical priming: we observe significant priming effects for Transparent Derivation and Synonymy, and no priming for Opaque Derivation. This is contrary to Smolka et al.’s experimental results.
Our instance of the MORGEN model does a much better job: It predicts highly significant priming effects for both Transparent and Opaque derivations (p<0.001) while priming is not significant at p<0.01 for Synonyms (p=0.04). These predictions correspond very well to Smolka et al.’s findings (cf. Table 1). We tested for two additional contrasts analyzed by Smolka et al.: the difference in priming strength between Transparent and Opaque Derivation (not significant in either experiment or model) and the difference between Transparent Derivation and Synonym (highly significant in both experiment and model). 4
In sum, we find a very good match betweenMORGEN and the experimental results, while the DISTSIM model cannot account for the experimental evidence. Recall that the main difference between the two models is that MORGEN’s includes all members of the prime’s derivational family into the prediction of the priming strength. This leads to the following changes compared to DISTDIM: 1. For Opaque Derivation, MORGEN typically predicts stronger priming than DISTSIM, since prime and target are typically members of the same derivational family (assuming that there are no coverage gaps in DERIVBASE), and the average similarity between the target and the words in the family is higher than the similarity to the prime itself. Taking Figure 1 as an example, the Opaque Derivation pair entbinden (give birth) – binden (bind)is relatively dissimilar, and the similarity increases when other pairs like binden (bind) – zubinden (tie) are taken into consideration. 2. For Synonymy, MORGEN typically predicts weaker priming than DISTSIM, since the average similarity between target and all members of the prime’s family tends to be lower than the similarity between target and original prime. Again considering Figure 1, the Synonym pair binden (bind) – zuschnu¨ren (tie) is relatively similar, while including terms derivationally related to the prime zuschnu¨ren (tie) like schnurlos (cordless) introduces lowsimilarity pairs like schnurlos (cordless) – binden (bind).
MORGEN is not the only model that takes a distributional stance towards morphological derivation. Marelli and Baroni (2014) propose a compositional model that computes separate distributional representations for the meanings of stems and affixes and is able to compute representations for novel, unseen derived terms. The morpheme-level approach of Marelli and Baroni’s model corresponds more directly to Smolka et al.’s claims and might also be able to account for the experimental patterns.
However, our considerably simpler model, which only has knowledge about distributional families, is also able to do so. This at the very least means that morpheme-level processing is not an indispensable property of any model that explains Smolka et al.’s experimental results and that the evidence for a special organization of the German mental lexicon, in contrast to other languages, must be examined more carefully.
In fact, our model provides a possible alternative source of explanations for the cross-lingual differences: Since the MORGEN predictions are directly influenced by the size and members of the derivational families, German opaque morphological priming may simply result from the high frequency of opaque derivations. In the future, we plan to apply the model to Dutch and French to check this alternative explanation.
We gratefully acknowledge funding by Deutsche Forschungsgemeinschaft through Sonderforschungsbereich 732, project B9.
Lexicon, pages 7–8, Niagara on the Lake, Canada.
Abstract.