A Logistic Regression Model of Determiner Omission in PPs

The realization of singular count nouns without an accompanying determiner inside a PP (determinerless PP, bare PP, Preposition-Noun Combination) has recently attracted some interest in computational linguistics. Yet, the relevant factors for determiner omission remain unclear, and conditions for determiner omission vary from language to language. We present a logistic regression model of determiner omission in German based on data obtained by applying annotation mining to a large, automatically and manually annotated corpus. 1 The problem and how to deal with it Preposition-Noun Combinations (PNCs, sometimes called determinerless PPs or bare PPs) minimally consist of a preposition and a count noun in the singular that – despite requirements formulated elsewhere in the grammar of the respective language – appears without a determiner. The noun in a PNC can be extended through prenominal modification (1) and postnominal complementation (2). Still, a determiner is missing. The following examples are given from German. (1) auf parlamentarische Anfrage (‘after being asked in parliament’), mit beladenem Rucksack (‘with loaded backpack’), unter sanfter Androhung (‘under gentle threat’) (2) Er wehrt sich gegen die Forderung nach he defies REFL against the demand for Stilllegung einer Verbrennungsanlage. closedown an incineration plant ‘He defies the demand for closing an incineration plant.’ PNCs occur in a wide range of languages (Himmelmann, 1998); the conditions for determiner omission, however, have not been detected yet, and conditions applying to one language do not carry over to other languages. In addition, speakers only reluctantly judge the acceptability of newly coined PNCs, so that reliance to introspective judgments cannot be assumed. For English, Stvan (1998) and Baldwin et al. (2006) have claimed that either the semantics of the preposition or of the noun play a major role in determining whether a singular count noun may appear without a determiner in a PNC. Stvan (1998) assumes that nouns determine the well-formedness of PNCs (3) if the denotation of the noun occurs in a particular semantic field, while Baldwin et al. (2006) assume that certain prepositions impose selection restrictions on their nominal complements that allow for determiner omission (4). (3) from school, at school, in jail, from jail, ... (4) by train, by plane, by bus, by pogo stick, by hydro-foil ... Interestingly, Le Bruyn et al. (2009) have observed that basic assumptions of Stvan’s analysis do not apply to Dutch, French, or Norwegian. With regard to German, we observe that neither the pattern in (3) nor in (4) is productive. Constructions like (4) cannot be realized as PNCs in German, but require full PPs. In the following, we propose an analysis of PNCs that combines corpus annotation, annotation mining (Chiarcos et al., 2008), and logistic regression modeling (Harrell, 2001). Annotation mining assumes that linguistically relevant generalizations can be derived in a bottom-up fashion from a suitably annotated corpus. Relevant hits in the corpus are mapped into a feature vector that serves as input for logistic regression classification. In the present case, the input consists of sentences containing either PNCs or PPs. Binary logistic regression suggests itself as a classification method since the problem of PNCs can be rephrased as the following question: Under which conditions can an otherwise obligatory determiner be omitted? The majority of required annotations can be derived automatically, but there are no available systems for the automatic determination of preposition senses in German, so preposition sense annotation has to be carried out manually and requires a language-specific tagset for preposition senses. While our initial analysis is based on German data, the general methodology can be applied to other languages, provided that corpora receive proper annotation.