Combining Deep and Shallow Approaches in Parsing German

Combining Deep and Shallow Approaches in Parsing German Michael Schiehlen Institute for Computational Linguistics, University of Stuttgart, Azenbergstr. 12, D-70174 Stuttgart mike@adler.ims.uni-stuttgart.de Abstract The paper describes two parsing schemes: a shallow approach based on machine learning and a cascaded finite-state parser with a hand-crafted grammar. It dis- cusses several ways to combine them and presents evaluation results for the two in- dividual approaches and their combina- tion. An underspecification scheme for the output of the finite-state parser is intro- duced and shown to improve performance. 1 Introduction In several areas of Natural Language Processing, a combination of different approaches has been found to give the best results. It is especially rewarding to combine deep and shallow systems, where the for- mer guarantees interpretability and high precision and the latter provides robustness and high recall. This paper investigates such a combination consist- ing of an n-gram based shallow parser and a cas- caded finite-state parser1 with hand-crafted gram- mar and morphological checking. The respective strengths and weaknesses of these approaches are brought to light in an in-depth evaluation on a tree- bank of German newspaper texts (Skut et al., 1997) containing ca. 340,000 tokens in 19,546 sentences. The evaluation format chosen (dependency tuples) is used as the common denominator of the systems 1Although not everyone would agree that finite-state parsers constitute a ‘deep’ approach to parsing, they still are knowledge-based, require efforts of grammar-writing, a com- plex linguistic lexicon, manage without training data, etc. in building a hybrid parser with improved perfor- mance. An underspecification scheme allows the finite-state parser partially ambiguous output. It is shown that the other parser can in most cases suc- cessfully disambiguate such information. Section 2 discusses the evaluation format adopted (dependency structures), its advantages, but also some of its controversial points. Section 3 formu- lates a classification problem on the basis of the evaluation format and applies a machine learner to it. Section 4 describes the architecture of the cas- caded finite-state parser and its output in a novel underspecification format. Section 5 explores sev- eral combination strategies and tests them on several variants of the two base components. Section 6 pro- vides an in-depth evaluation of the component sys- tems and the hybrid parser. Section 7 concludes. 2 Parser Evaluation The simplest method to evaluate a parser is to count the parse trees it gets correct. This measure is, how- ever, not very informative since most applications do not require one hundred percent correct parse trees. Thus, an important question in parser evaluation is how to break down parsing results. In the PARSEVAL evaluation scheme (Black et al., 1991), partially correct parses are gauged by the number of nodes they produce and have in com- mon with the gold standard (measured in precision and recall). Another figure (crossing brackets) only counts those incorrect nodes that change the partial order induced by the tree. A problematic aspect of the PARSEVAL approach is that the weight given to particular constructions is again grammar-specific, since some grammars …

Content truncated to 500 words