Tree Kernel-based Negation and Speculation Scope Detection with Structured Syntactic Parse Features

Bowei Zou, Guodong Zhou, Qiaoming ZhuEMNLP 2013 · OpenReview

Abstract

Scope detection is a key task in information extraction. This paper proposes a new approach for tree kernel-based scope detection by using the structured syntactic parse information. In addition, we have explored the way of selecting compatible features for different part-of-speech cues. Experiments on the BioScope corpus show that both constituent and dependency structured syntactic parse features have the advantage in capturing the potential relationships between cues and their scopes. Compared with the state of the art scope detection systems, our system achieves substantial improvement.

Introduction: The Challenge of Scope Detection

In Information Extraction, identifying the presence of a concept is only half the battle. One must also determine if that concept is being negated ("not expensive") or is merely speculative ("possible future scenarios"). This task, known as Scope Detection, is critical for accurately processing clinical and scientific texts.

Traditional systems often relied on rigid rules or "flat" feature lists. However, language is inherently hierarchical. Consider the sentence: "The chair is not expensive but comfortable." Here, the scope of the negation cue "not" is limited to "expensive."

Dependency parse tree for the sentence "The chair is not expensive but comfortable."

As shown in the dependency structure above, the relationship between the cue ("not") and its scope ("expensive") is defined by grammatical dependencies rather than simple linear distance. Capturing these structural nuances is the primary challenge addressed by Zou, Zhou, and Zhu.

Methodology: Leveraging Tree Kernels

The authors frame scope detection as a binary classification task: for every token in a sentence, the system determines if it falls "inside" or "outside" the scope of a given cue. The core innovation lies in two areas:

1. Structured Syntactic Features

Instead of treating sentences as sequences of tokens, the model uses Tree Kernels to compare sub-structures of constituent and dependency parse trees. Since standard tree kernels are designed for constituent trees, the authors developed transformational rules to convert dependency relations into nodes that a kernel can process effectively.

2. Part-of-Speech (POS) Optimization

The "shape" of a scope varies significantly based on the cue's grammatical role. For instance, an adjective cue like putative behaves differently than a verb cue like suggests. The authors built discriminative classifiers tailored to specific POS categories, allowing the system to select the most compatible features for each cue type.

Key Insights: The Critical Path and Sub-Trees

A major finding of the study is that specific structural segments of the parse tree are more informative than others. Two structures stood out:

  • The Completed Sub-Tree: The smallest sub-tree that contains both the cue and the token being classified.
  • The Critical Path: The direct route between the cue and the token within the tree.
Visual comparison of the Completed Sub-Tree and the Critical Path within a parse structure.

By focusing on these structures, the model can ignore irrelevant branches and focus on the "indirect clues" hidden in the grammar. For example, the presence of a coordinating conjunction like "but" often signals the end of a negation scope—a relationship clearly visible in a tree but difficult to generalize in a flat list.

Results and Performance

The system was evaluated on the BioScope corpus using the Percentage of Correct Scopes (PCS) metric, which requires a perfect match for the entire scope.

  • Negation: On the Abstracts sub-corpus, the system achieved 76.90% PCS, a significant improvement over the previous state-of-the-art (73.36%).
  • Speculation: Accuracy reached 84.21% PCS, outperforming the best previous system (79.89%).
  • Cross-Domain Robustness: The system performed exceptionally well on clinical radiology reports (85.31% for negation), suggesting that syntactic structures are more universal across domains than specific vocabulary.

Conclusion and Future Work

While the tree-kernel approach is highly effective, performance dips on the Full Papers sub-corpus due to the extreme complexity and length of sentences, which can lead to "noisy" parse trees. Future research is focused on developing pruning algorithms to "trim" these trees, helping the model maintain focus on the most important structural signals in complex scientific prose.

Summary updated

Abstract

Scope detection is a key task in information extraction. This paper proposes a new approach for tree kernel-based scope detection by using the structured syntactic parse information. In addition, we have explored the way of selecting compatible features for different part-of-speech cues. Experiments on the BioScope corpus show that both constituent and dependency structured syntactic parse features have the advantage in capturing the potential relationships between cues and their scopes. Compared with the state of the art scope detection systems, our system achieves substantial improvement.
BibTeX
@inproceedings{zou2013tree,
  author = {Bowei Zou and Guodong Zhou and Qiaoming Zhu},
  title = {{Tree Kernel-based Negation and Speculation Scope Detection with Structured Syntactic Parse Features}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
  url = {https://openreview.net/forum?id=SyEV0-z_-B},
}
Figure