A Logistic Regression Model of Determiner Omission in PPs

Tibor Kiss, Katja Keßelmeier, Antje Müller, Claudia Roch, Tobias Stadtfeld, Jan StrunkCOLING 2010

TL;DR

This paper tackles the linguistic puzzle of why English determiners are sometimes omitted in prepositional phrases (e.g., "from school" instead of "from the school"). Using a logistic regression model trained on a large German corpus, the authors identify a complex interplay of factors, including preposition meaning, noun interpretation, and syntactic structure, that govern this phenomenon. Their work moves beyond simple explanations to offer a nuanced, data-driven understanding of determiner omission.

What Problem Does This Paper Address?

In many languages, including German, prepositional phrases (PPs) typically require a determiner before a singular count noun (e.g., "in the park"). However, there are instances where this determiner is omitted, forming what the authors call "determinerless PPs" or "Preposition-Noun Combinations" (PNCs), such as "from school" or "at home." While this phenomenon has been observed across languages, the underlying linguistic rules and the specific factors that license determiner omission remain unclear. Previous attempts to explain this have often focused on either the preposition's semantics or the noun's semantics in isolation, but these approaches have proven insufficient and language-specific. This paper aims to build a more comprehensive, data-driven model to explain determiner omission in German PPs.

Methods

The authors employ a multi-stage approach combining corpus annotation, annotation mining, and logistic regression modeling.

  1. Corpus Annotation: They utilize a large corpus of Swiss German newspaper text. Manual annotation is performed to identify PNCs and corresponding full PPs. Crucially, they annotate various linguistic features at different levels:
  2. Lexical: Part-of-speech, morphology, count/mass distinction, noun interpretation, preposition interpretation.
  3. Syntactic: Phrase embedding, noun dependents, noun modification.
  4. Global: Whether the phrase appears in headlines, titles, or quotations (which can be prone to truncation and different rules).
  5. Automatic Annotation: Several tools are used for automated feature extraction, including POS tagging, morphological analysis, and chunk parsing. For noun meaning, they leverage GermaNet (a German WordNet) and the HaGenLex lexicon. A classifier is also used for the count/mass distinction of nouns.
  6. Annotation Mining & Feature Engineering: Relevant linguistic generalizations are extracted from the annotated corpus. These extracted features form a "feature vector" for each instance.
  7. Logistic Regression Modeling: A binary logistic regression model is trained to predict the probability of determiner omission (DET = no) versus determiner realization (DET = yes). This method is chosen because it allows for the identification of multiple contributing factors and their relative influence. They start with a full model and then use feature elimination to arrive at a parsimonious yet explanatory model.

Datasets

The primary dataset used is a large corpus of the Swiss-German newspaper Neue Zürcher Zeitung from 1993 to 1999, comprising approximately 230 million words. This corpus was manually annotated for prepositional phrases and their properties. No new benchmark datasets were introduced; rather, the authors leverage existing linguistic resources like GermaNet and HaGenLex for feature extraction.

Key Insights and Ideas

  • Multicausality is Key: Determiner omission is not governed by a single factor but by a complex interplay of prepositional meaning, noun interpretation, and syntactic structure.
  • Syntactic Complexity Hinders Omission: The presence of syntactic complements, especially sentential complements (DEP-S), significantly increases the probability of determiner realization, thus hindering omission. This suggests that structural complexity acts as a barrier to bare PPs.
  • Morphological Derivation Favors Omission: Nouns derived via suffixation (like "-ung" in German, captured by the NOMINALIZATION feature) tend to favor determiner omission. This might reflect underlying semantic distinctions between events and objects.
  • Prepositional Senses Matter, But Not Exclusively: While prepositional senses do play a role (e.g., PRESENCE favoring omission for "ohne"), their influence is not always straightforward and can be overridden by other factors. For "unter," specific senses were found to block omission rather than favor it.
  • Language-Specific Patterns: The authors highlight that patterns observed in English (like those proposed by Stvan or Baldwin et al.) do not directly apply to German, underscoring the need for language-specific modeling.

Key Results

  • The logistic regression models achieved high predictive power, with Somers' Dxy values of 0.753 for "ohne" and 0.874 for "unter," indicating good performance in distinguishing between determiner omission and realization.
  • For the preposition "ohne" (without), features like NOMINALIZATION, ADJA (presence of modifying adjectives), CAUSAL, COMITATIVE, and PARTICIPATION were significant predictors. PRESENCE and GN-RELATION, GN-ATTRIBUTE, GN-EVENT, GN-ARTEFACT favored omission.
  • For the preposition "unter" (under/below), a more complex model with more predictors was required due to its higher polysemy. Features like CAUSAL, SUBORD., ORDER, and TRANSGRESSION were found to block omission, while DEP-S, DEP-NP, and DEP-PP favored realization.
  • The study found that while some prepositional senses favor omission, others block it, and the syntactic structure of the noun phrase (especially sentential complements) strongly favors determiner realization.

Limitations and Future Work

  • Limited Scope: The study focuses on two prepositions ("ohne" and "unter") and German. The authors acknowledge that the methodology can be applied to other languages and prepositions, but this requires proper annotation.
  • Ambiguity Handling: The approach does not disambiguate polysemous nouns, allowing them to contribute to multiple features. While this is justified by the model's evaluation of individual feature effects, it's a point of note.
  • Deeper Referential Requirements: The authors speculate that the dependency of determiner realization on sentential complementation might reflect deeper referential requirements, which warrants future investigation.
  • General Mechanism for Complementation: They suspect a general mechanism relating sentential complementation to determiner realization across prepositions, a topic for future research.
  • Role of "ADJA" and "NOMINALIZATION": The specific roles of the ADJA and NOMINALIZATION features are noted as requiring further clarification in future work, particularly regarding whether they reflect formal properties or underlying denotational distinctions.

Summary updated

BibTeX
@inproceedings{kiss2010logistic,
  author = {Tibor Kiss and Katja Keßelmeier and Antje Müller and Claudia Roch and Tobias Stadtfeld and Jan Strunk},
  title = {{A Logistic Regression Model of Determiner Omission in PPs}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2010},
}
Figure