Inside-Outside Estimation of a Lexicalized PCFG for German

Inside-Outside Estimation of a Lexicalized PCFG for German Franz Beil, Glenn Carroll, Detlef Prescher, Stefan Riezler and Mats Rooth Institut ffir Maschinelle Sprachverarbeitung, University of Stuttgart Abstract The paper describes an extensive experiment in inside-outside estimation of a lexicalized proba- bilistic context free grammar for German verb- final clauses. Grammar and formalism features which make the experiment feasible are de- scribed. Successive models are evaluated on pre- cision and recall of phrase markup. 1 Introduction Charniak (1995) and Carroll and Rooth (1998) present head-lexicalized probabilistic context free grammar formalisms, and show that they can effectively be applied in inside-outside es- timation of syntactic language models for En- glish, the parameterization of which encodes lexicalized rule probabilities and syntactically conditioned word-word bigram collocates. The present paper describes an experiment where a slightly modified version of Carroll and Rooth's model was applied in a systematic experiment on German, which is a language with rich in- flectional morphology and free word order (or rather, compared to English, free-er phrase or- der). We emphasize techniques which made it practical to apply inside-outside estimation of a lexicalized context free grammar to such a language. These techniques relate to the treat- ment of argument cancellation and scrambled phrase order; to the treatment of case features in category labels; to the category vocabulary for nouns, articles, adjectives and their projections; to lexicalization based on uninflected lemmata rather than word forms; and to exploitation of a parameter-tying feature. 2 Corpus and morphology The data for the experiment is a corpus of Ger- man subordinate clauses extracted by regular expression matching from a 200 million token newspaper corpus. The clause length ranges be- tween four and 12 words. Apart from infiniti- val VPs as verbal arguments, there are no fur- ther clausal embeddings, and the clauses do not contain any punctuation except for a ter- minal period. The corpus contains 4128873 to- kens and 450526 clauses which yields an average of 9.16456 tokens per clause. Tokens are auto- matically annotated with a list of part-of-speech (PoS) tags using a computational morpholog- ical analyser based on finite-state technology (Karttunen et al. (1994), Schiller and StSckert (1995)). A problem for practical inside-outside esti- mation of an inflectional language like German arises with the large number of terminal and low-level non-terminal categories in the gram- mar resulting from the morpho-syntactic fea- tures of words. Apart from major class (noun, adjective, and so forth) the analyser provides an ambiguous word with a list of possible combina- tions of inflectional features like gender, person, number (cf. the top part of Fig. 1 for an exam- ple ambiguous between nominal and adjectival PoS; the PoS is indicated following the '+' sign). In order to reduce the number of parameters to be estimated, and to reduce the size of the parse forest used in inside-outside estimation, we collapsed the inflectional readings of adjec- tives, adjective derived nouns, article words, and pronouns to a single morphological feature (see of Fig. 1 for an example). This reduced the …

Content truncated to 500 words