Abstract
This thesis is about computational morphology for German derivation. Derivation is a word formation process that creates new words from existing ones, where the base and the derived word share the same stem. Mostly, derivation is conducted by means of relatively regular affixation rules, as in to bake - bakery. In German, derivation is highly productive, thus leading to a high language variability which can be employed to express similar facts in different ways, as derivationally related words are often also semantically related (or transparent). However, linguistic variance is a challenge for computational applications, particularly in semantic processing: It makes it more difficult to automatically grasp the meaning of texts and to match similar information onto each other. Thus, computational systems require linguistic knowledge. We develop methods to induce and represent derivational knowledge, and to apply it in language processing. The main outcome of our study is DErivBase, a German derivational lexicon. It groups derivationally related words (words that are derived from the same stem) into derivational families. To achieve high quality and high coverage, we induce DErivBase by combining rule-based and data-driven methods: We implement linguistic derivation rules to define derivational processes, and feed lemmas extracted from a German corpus into the rules to derive new lemmas. All words that are connected - directly or indirectly - by such rules are considered a derivational family. As mentioned above, a derivational relationship often implies semantic relationship, but this is not always the case. Semantic drifts can cause semantically unrelated (opaque) derivational relations, such as to depart - department. Capturing the difference between transparent and opaque relations is important from a linguistic as well as a practical point of view. Thus, we conduct a semantic refinement of DErivBase, i.e., we determine which lemma pairs are derivationally and semantically related, and which are not. We establish a second, semantically validated version of our lexicon, where families are sub-clustered according to semantic coherence, using supervised machine learning methods: We learn a binary classifier based on features that arise from structural information about the derivation rules, and from distributional information about the semantic relatedness of lemmas. Accordingly, the derivational families are subdivided into semantically coherent clusters. To demonstrate the utility of the two lexicon versions, we evaluate them on three extrinsic - and in the broadest sense, semantic - tasks. The underlying assumption for applying DErivBase to semantic tasks is that derivational relatedness is a reasonable approximation to semantic relatedness, since derivation is often semantically transparent. Our three experiments are the following: 1., we incorporate DErivBase into distributional semantic models to overcome sparsity problems and to improve the prediction quality of the underlying model. We test this method, which we call derivational smoothing, for semantic similarity prediction, and for synonym choice. 2., we employ DErivBase to model a psycholinguistic experiment that examines priming effects of transparent and opaque derivations to draw conclusions about the mental lexical representation in German. Derivational information is again incorporated into a distributional model, but this time, it introduces a kind of morphological generalisation. 3., in order to solve the task of Recognising Textual Entailment, we integrate DErivBase into a matching-based entailment system by means of a query expansion. Assuming that derivational relationships between two texts suggest them to be entailing rather than non-entailing, this expansion increases the chance of a lexical overlap, which should improve the system's entailment predictions. The incorporation of DErivBase indeed improves the performance of the underlying systems in each task, however, it is differently suitable in different settings. In experiment 1., the semantically validated lexicon yields improvements over the purely morphological lexicon, and the more coarse-grained similarity prediction profits more from DErivBase than the synonym choice. In experiment 2., purely morphological information clearly outperforms the other lexicon version, as the latter cannot model opaque derivations. On the entailment task in experiment 3., DErivBase has only minor impact, because textual entailment is hard to solve by addressing only one linguistic phenomenon. In sum, our findings show that the induction of a high-quality, high-coverage derivational lexicon is beneficial for very different applications in computational linguistics. It might be worthwhile to further investigate the semantic aspects of derivation to better understand its impact on language and thus, on language processing.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.