Bitext-Based Resolution of German Subject-Object Ambiguities

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 737–740, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Bitext-Based Resolution of German Subject-Object Ambiguities Florian Schwarck Alexander Fraser Institute for Natural Language Processing University of Stuttgart {koehlefn,fraser}@ims.uni-stuttgart.de Hinrich Sch¨utze Abstract We present a method for disambiguating syn- tactic subjects from syntactic objects (a fre- quent ambiguity) in German sentences taken from an English-German bitext. We exploit the fact that subject and object are usually eas- ily determined in English. We show that a simple method disambiguates some subject- object ambiguities in German, while making few errors. We view this procedure as the first step in automatically acquiring (mostly) cor- rect labeled data. We also evaluate using it to improve a state of the art statistical parser. 1 Introduction Ambiguity of grammatical role is a problem when parsing a number of natural languages. In German, subject-object ambiguities are frequent. The sen- tence “Die Maus jagt die Katze” “the – mouse – chases – the – cat” exhibits such an ambiguity. Be- cause word order is freer in German than in English, the sentence has two possible meanings: (i) The cat is chasing the mouse and (ii) the mouse is chasing the cat. We exploit the fact that such ambiguities are much less frequent in languages that possess a less flexible syntax than German. In English, the trans- lation of the sentence “Die Maus jagt die Katze” is not ambiguous. If we have access to this translation, we can use this information to disambiguate the Ger- man sentence. The English translation is viewed as a surrogate for both contextual knowledge from the text and for world knowledge. We present a method for disambiguating the sub- ject and object roles in German sentences. We use an English-German bitext and exploit the fact that subject and object roles are rarely ambiguous in En- glish. Using a new gold standard we created we show that our method disambiguates a significant proportion of subject-object ambiguities in German with high precision. We view this procedure as the first step in automatically acquiring (mostly) correct labeled data for training a statistical disambiguator that can be used on German text (even when no translation is available). In addition to measuring algorithm performance directly, we present experi- ments on improving the disambiguation of BitPar, a state of the art statistical parser. 2 Algorithm Data and Word Alignment. We use the aligned English and German sentences in Europarl (Koehn, 2005) for our experiments. The corpus contains long and complex sentences. To establish translational correspondence between parallel sentences we use GIZA++ (Och and Ney, 2003). Its input is a tok- enized parallel corpus. We lemmatized the text prior to aligning it. Procedure. Figure 1 shows the architecture of our system. The boxes signify data sets, while the lines are processes applied to the data sets. The pa- per presents two applications. The first is the cre- ation of a set of disambiguated German …

Content truncated to 500 words