Extraposition: A Case Study in German Sentence Realization

Michael Gamon, Eric Ringger, Zhu Zhang, Robert Moore, Simon Corston-OliverCOLING 2002

TL;DR

This paper tackles the challenge of generating more natural-sounding German sentences by modeling "extraposition," a grammatical phenomenon where clauses are moved to the end of a sentence. They demonstrate that machine learning, specifically decision trees, can effectively learn to predict when and how to extrapose clauses, significantly improving fluency compared to simpler approaches.

What Problem Does This Paper Address?

Natural language generation (NLG) systems aim to produce human-like text from abstract representations. A crucial, yet complex, part of this process is sentence realization, where syntactic structures are transformed into surface strings. For German, a phenomenon called "clausal extraposition" is essential for fluent output. This involves moving subordinate clauses (like relative, infinitival, or complement clauses) from their typical position to the end of the main clause. While grammatically correct, leaving these clauses in their "base" position can sound awkward or less fluent, especially when the clauses are long or complex. Existing NLG systems often struggle to handle this phenomenon effectively, leading to unnatural-sounding German. This paper investigates how to model extraposition using machine learning to improve German sentence realization.

Methods

The authors explore two machine learning strategies for modeling extraposition, both using decision tree learning:

  • Multi-step Movement: This approach models extraposition as a series of decisions. For each potential attachment site, a decision tree asks whether the clause should move up one level (to the grandparent node) or stay put. This process repeats until the decision is to stop moving.
  • One-step Movement: This strategy simplifies the process by making a single classification decision for each node in the parent chain of the extraposable clause. The question is whether the clause should move directly to that target node from its base position. If multiple nodes qualify, the one with the highest probability is chosen.

Both approaches leverage features extracted from syntactic and semantic analyses of parsed sentences.

Datasets

The study uses two large German corpora for training and evaluation:

  • Technical Manual Corpus: 100,000 aligned English-German sentence pairs from Microsoft technical manuals.
  • Encarta Corpus: 100,000 randomly selected English and German sentences from the Encarta encyclopedia.

These datasets were split into training and parameter tuning sets. A separate blind test set of 10,000 sentences from each domain was used for evaluation.

Key Insights and Ideas

  • Extraposition is Pervasive in German: Unlike English, where extraposition is relatively rare, it's a common and important phenomenon in German, especially in technical and encyclopedic texts. Ignoring it significantly impacts fluency.
  • Machine Learning Outperforms Rule-Based Approaches: Decision tree classifiers, trained on automatically extracted features, can effectively learn the complex patterns governing extraposition, surpassing simple baseline models that never extrapose.
  • Multi-step Movement is Slightly Better: The multi-step movement model consistently achieved slightly higher accuracy than the one-step model across both domains, suggesting that modeling the gradual movement of clauses is more effective.
  • Domain Specificity Matters: Models trained on one domain (e.g., technical manuals) perform significantly worse when applied to another domain (e.g., Encarta), indicating that the linguistic generalizations for extraposition are domain-dependent.
  • Feature Importance: Beyond just the "heaviness" of a clause, structural context, verbal properties (like verb position and separable prefixes), and semantic roles play crucial roles in determining extraposition likelihood.

Key Results

  • Improved Accuracy: Both the one-step and multi-step models significantly outperformed the baseline (no extraposition) in both the technical and Encarta domains.
  • In the technical domain, the multi-step model achieved an overall accuracy of 86.12% compared to a baseline of 67.58%.
  • In the Encarta domain, the multi-step model achieved an overall accuracy of 90.61% compared to a baseline of 84.15%.
  • Multi-step Advantage: The multi-step model consistently showed a slight accuracy advantage over the one-step model (e.g., 1.7% higher overall in the technical domain).
  • Domain-Specific Models: Models trained on the Encarta domain showed higher accuracy when tested on Encarta data (around 90%) compared to when tested on technical domain data (around 83%). Similarly, technical domain models performed better on technical data (around 86%) than on Encarta data (around 83%).

Limitations and Future Work

  • Inconclusive Results for INFCL and COMPCL: The number of infinitival (INFCL) and complement clauses (COMPCL) that were extraposed in the test sets was very small, making it difficult to draw firm conclusions about the models' performance on these clause types. More data, especially for these types, would be beneficial.
  • Model Complexity: The decision trees for the technical domain were more complex (more branching nodes) than those for the Encarta domain, potentially due to NLPWin assigning a higher percentage of spanning parses to Encarta data, leading to more reliable input.
  • Future Work: The authors suggest that the current setup, which uses the syntactic label of the extraposable clause as a feature, will automatically adapt to new clause types or domains as more training data becomes available. They also highlight that the qualitative inspection of selected features provides valuable linguistic insights into the phenomenon of extraposition. The multi-step model was chosen for inclusion in their end-to-end German sentence realization system, "Amalgam."

Summary updated

BibTeX
@inproceedings{gamon2002extraposition,
  author = {Michael Gamon and Eric Ringger and Zhu Zhang and Robert Moore and Simon Corston-Oliver},
  title = {{Extraposition: A Case Study in German Sentence Realization}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure