Abstract
This study investigates the identification of "topic expressions"—the specific words representing what a sentence is about—in Danish doctor-patient dialogues. By analyzing surface-level linguistic features like word order and definiteness, researchers trained decision tree models that achieve near-human performance. The results highlight the "preverbal" position in Danish as a primary signal for topicality, demonstrating that complex pragmatic categories can be mapped onto concrete syntactic and morphological features.
Introduction
In linguistics, the "topic" is intuitively understood as what an utterance is "about." However, formalizing this concept is difficult because "aboutness" is a pragmatic category dependent on context and discourse flow. Danish, a "verb-second" (V2) language, provides a unique case study: while its word order is relatively fixed, the first position in a main clause is flexible. This paper seeks to bridge the gap between theory and data by using machine learning to determine if surface-level features can reliably predict the pragmatic topic in spoken Danish.
Methodology
The authors employed a three-stage supervised learning approach:
- Identification: Two coders manually identified "topic expressions" in a corpus using a modified "about test."
- Annotation: Every noun phrase (NP) was tagged with 16 surface features, including grammatical roles, morphological traits (e.g., definiteness, pronominalization), and positional data.
- Machine Learning: A decision tree generator was used to find correlations between these features and the manually labeled topics, tested via 10-fold cross-validation.
Datasets and Reliability
The study utilized dialogues from a corpus of doctor-patient conversations regarding weight problems. For the evaluation of the model, a dataset of 449 NPs was used. Given the subjectivity of "topic," the authors measured inter-coder agreement to establish a baseline for reliability.
As shown in the contingency table above, the coders achieved a Kappa ($\kappa$) of 0.70, indicating "reasonable reliability" for such a complex pragmatic task. This agreement formed the basis for the subsequent machine learning training and testing.
Key Insights
- The Power of the Preverbal Position: The strongest predictor of a topic in Danish is the "preverbal" position. An NP appearing before the finite verb in a main clause has a high statistical probability of being the topic.
- The "Interpersonal" Filter: In "epistemic matrix clauses" (e.g., "I think..."), the subjects are rarely the topic. Instead, the topic typically resides in the subsequent subordinate clause.
- Feature Bundling: The "ideal" topic expression in Danish is a definite pronoun located in the preverbal position of a non-subordinate clause.
Results
The decision tree models performed remarkably well, approaching the limits of human agreement. The most critical features for the classifier were preverbal position (+pre), definiteness (+def), pronominalization (+pro), and non-subordination (-sub).
As illustrated in the table above, the system achieved 84–89% accuracy (success rate) in identifying topic expressions, with F1-scores ranging from 0.63 to 0.72. For comparison, the human agreement "gold standard" reached 89% accuracy with an F1 of 0.78, suggesting the model is capturing the majority of the linguistic signals used by experts.
Conclusion and Future Work
This study successfully demonstrates that high-level pragmatic topicality can be mapped onto searchable linguistic features with high accuracy. The authors suggest that future work should link these findings to Centering Theory, specifically the "Backward-looking Center" ($C_b$). Such integration could significantly improve automated natural language processing tasks, such as anaphora resolution, in Danish systems.
Summary updated