Parsing Coordinations

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 406–414, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Parsing Coordinations Sandra K¨ubler Indiana University skuebler@indiana.edu Erhard Hinrichs Universit¨at T¨ubingen eh@sfs.uni-tuebingen.de Wolfgang Maier Unversit¨at T¨ubingen wo.maier@uni-tuebingen.de Eva Klett Universit¨at T¨ubingen eklett@sfs.uni-tuebingen.de Abstract The present paper is concerned with sta- tistical parsing of constituent structures in German. The paper presents four ex- periments that aim at improving parsing performance of coordinate structure: 1) reranking the n-best parses of a PCFG parser, 2) enriching the input to a PCFG parser by gold scopes for any conjunct, 3) reranking the parser output for all possi- ble scopes for conjuncts that are permissi- ble with regard to clause structure. Exper- iment 4 reranks a combination of parses from experiments 1 and 3. The experiments presented show that n- best parsing combined with reranking im- proves results by a large margin. Provid- ing the parser with different scope possi- bilities and reranking the resulting parses results in an increase in F-score from 69.76 for the baseline to 74.69. While the F-score is similar to the one of the first ex- periment (n-best parsing and reranking), the first experiment results in higher re- call (75.48% vs. 73.69%) and the third one in higher precision (75.43% vs. 73.26%). Combining the two methods results in the best result with an F-score of 76.69. 1 Introduction The present paper is concerned with statistical parsing of constituent structures in German. Ger- man is a language with relatively flexible phrasal ordering, especially of verbal complements and adjuncts. This makes processing complex cases of coordination particularly challenging and error- prone. The paper presents four experiments that aim at improving parsing performance of coor- dinate structures: the first experiment involves reranking of n-best parses produced by a PCFG parser, the second experiment enriches the input to a PCFG parser by offering gold pre-bracketings for any coordinate structures that occur in the sen- tence. In the third experiment, the reranker is given all possible pre-bracketed candidate struc- tures for coordinated constituents that are permis- sible with regard to clause macro- and microstruc- ture. The parsed candidates are then reranked. The final experiment combines the parses from the first and the third experiment and reranks them. Improvements in this final experiment corroborate our hypothesis that forcing the parser to work with pre-bracketed conjuncts provides parsing alterna- tives that are not present in the n-best parses. Coordinate structures have been a central is- sue in both computational and theoretical linguis- tics for quite some time. Coordination is one of those phenomena where the simple cases can be accounted for by straightforward empirical gen- eralizations and computational techniques. More specifically, it is the observation that coordination involves two or more constituents of the same cat- egories. However, there are a significant number of more complex cases of coordination that defy this generalization and that make the parsing task of detecting the right scope of individual conjuncts and correctly delineating …

Content truncated to 500 words