GeCoTagger: Annotation of German Verb Complements with Conditional Random Fields

GeCoTagger: Annotation of German Verb Complements with Conditional Random Fields Monica F¨urbacher, Roman Schneider Institute for the German Language (IDS) R5 6-16, 68161 Mannheim, Germany {fuerbacher, schneider}@ids-mannheim.de Abstract Complement phrases are essential for constructing well-formed sentences in German. Identifying verb complements and categorizing complement classes is challenging even for linguists who are specialized in the field of verb valency. Against this background, we introduce an ML-based algorithm which is able to identify and classify complement phrases of any German verb in any written sentence context. We use a large training set consisting of example sentences from a valency dictionary, enriched with POS tagging, and the ML-based technique of Conditional Random Fields (CRF) to generate the classification models. Keywords: Grammar and Syntax, Verb Valency, Machine Learning Methods 1. Introduction Verb complements are indispensable for constructing a cor- rect grammatical sentence in German. The appropriate us- age of complements is one fundamental skill for language learners, so the concept of verb complements - or verb va- lency - is not only an established field of linguistic research, but also often used for didactical purposes. The popular valency dictionary VALBU (=Valenzw¨orterbuch deutscher Verben) (Schumacher et al., 2004) and its expanded online counterpart E-VALBU (=Elektronisches Valenzw¨orterbuch deutscher Verben) (Kubczak, 2009) support both linguists and language learners by providing detailed descriptions of nearly 700 German verbs with more than 3,000 read- ing variants. Besides other linguistically motivated infor- mation, the dictionaries contain authentical example sen- tences, extracted from DeReKo ( (Kupietz et al., 2010)), with a manually added fine-grained markup of verb com- plement classes. Unfortunately, compiled dictionaries are naturally limited and cannot cover all possible sentences of a living language and even not the range of all existing verbs. Filling this gap manually seems to be an unpromising task, because it consumes much time and is error-prone. An automatic classification of complements for each verb in any sentence would solve this sophisticated problem. Though we see a remarkable increase of machine learning (ML) tools for natural language processing, we do not know of any empir- ical approach for the automatic classification of verb com- plements. For the development of our ML-based classification algo- rithm, we compile a corpus of 28,649 example sentences provided by the XML representation of E-VALBU (IDS- Mannheim, 2010) (M¨uller-Spitzer and Schneider, 2009). The corpus will then be POS-tagged and lemmatized. With this data set, we will train ML-models with different pa- rameters, based on Conditional Random Fields (CRF). The result will not only indicate whether complements can be identified at all by our algorithm, but also prove whether the complements will be correctly classified. 2. Complement classification The following section briefly addresses theoretical knowl- edge for the task of classifying verb complements. It cov- ers some linguistic basics regarding complements and gives necessary background information for developing a classi- fication algorithm based on machine learning. 2.1. Linguistic background A grammatically correct German sentence consists of three main components: the verbal complex, at least one verb complement and the facultative supplements. …

Content truncated to 500 words