Improving Zero-Shot-Learning for German Particle Verbs by using Training-Space Restrictions and Local Scaling

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 91–96, Berlin, Germany, August 11-12, 2016. Improving Zero-Shot-Learning for German Particle Verbs by using Training-Space Restrictions and Local Scaling Maximilian Köper and Sabine Schulte im Walde and Max Kisselew and Sebastian Padó Institut für Maschinelle Sprachverarbeitung, Universität Stuttgart Pfaffenwaldring 5B, 70569 Stuttgart, Germany {koepermn,schulte,kisselmx,pado}@ims.uni-stuttgart.de Abstract Recent models in distributional seman- tics consider derivational patterns (e.g., use → use + f ul) as the result of a compositional process, where base term and affix are combined. We exploit such models for German particle verbs (PVs), and focus on the task of learning a mapping function between base verbs and particle verbs. Our models apply particle-verb motivated training-space restrictions relying on nearest neighbors, as well as recent advances from zero- shot-learning. The models improve the mapping between base terms and de- rived terms for a new PV derivation dataset, and also across existing deriva- tion datasets for German and English. 1 Introduction Lazaridou et al. (2013) were the first to apply distributional semantic models (DSMs) to the task of deriving the meaning of morphologically complex words from their parts. They relied on high-dimensional vector representations to model the derived term (e.g., useful) as a result of a compositional process that combines the meanings of the base term (e.g., to use) and the affix (e.g., ful). For evaluation, they compared the predicted vector of the complex word with the original, corpus-based vector. More recently, Kisselew et al. (2015) put the task of modeling derivation into the perspective of zero-shot-learning: instead of using cosine similarities they predicted the derived term by learning a mapping function between the base term and the derived term. Once the predicted vector was computed, a nearest neighbor search was applied to validate if the prediction cor- responded to the derived term. In zero-shot- learning the task is to predict novel values, i.e., values that were never seen in training. More formally, zero-shot-learning trains a classifier f : X → Y that predicts novel values for Y (Palatucci et al., 2009). It is often applied across vector spaces, such as different domains (Mikolov et al., 2013; Lazaridou et al., 2015). The experiments by Kisselew et al. (2015) were performed over six derivational patterns for German (cf. Table 1), including particle verbs (PVs) with two different particle prefixes (an and durch), which were particularly difficult to predict. PVs such as anfangen (to start) are compositions of a base verb (BV) such as fan- gen (to catch) and a verb particle such as an. Predicting PV meaning is challenging because German PVs are highly productive (Springorum et al., 2013b; Springorum et al., 2013a), and the particles are notoriously ambiguous (Lech- ler and Roßdeutscher, 2009; Haselbach, 2011; Kliche, 2011; Springorum, 2011). Furthermore, the particles often trigger meaning shifts when they combine with base verbs (Springorum et al., 2013b), so the resulting PVs represent frequent cases of non-literal meaning. In this paper, we focus on predicting the meanings of German …

Content truncated to 500 words