A Lexicalized Tree Kernel for Open Information Extraction

Ying Xu, Christoph Ringlstetter, Mi-Young Kim, Grzegorz Kondrak, Randy Goebel, Yusuke MiyaoACL 2015 · OpenReview

Abstract

Traditional Open Information Extraction (Open IE) systems often rely on syntactic structures while ignoring lexical information due to data sparsity. This paper introduces a lexicalized tree kernel that incorporates word embeddings to bridge this gap. By smoothing the relationship between words using neural representations, the proposed model overcomes the limitations of purely syntactic parsing and achieves significant performance gains, with F-measure improvements ranging from 4% to 16% across various datasets.

Introduction

Relation Extraction (RE) typically operates on a fixed schema of predefined relationships (e.g., born_in). In contrast, Open IE aims to extract all possible relations from text without a restricted domain. While syntactic features like part-of-speech (POS) tags and dependency trees are robust, they are often insufficient for distinguishing semantically distinct relations that share identical syntax. For instance, "Federer hired Annacone" and "Federer considered Annacone" share the same dependency structure, but only the former implies a definitive professional relationship. This paper addresses the sparsity of lexical data by using word embeddings within a tree kernel framework to capture these semantic nuances.

System Architecture

The proposed system follows a supervised pipeline consisting of three primary stages: entity extraction and sentence parsing, relation candidate extraction, and final filtering via a Lexicalized Tree Kernel.

The three-module pipeline for relation extraction.

The process begins with identifying entities and generating dependency parses. The shortest dependency path between two entities is then extracted to form the basis of the relation candidate, which is then classified by the SVM Tree Kernel.

The Lexicalized Tree Kernel

The core innovation lies in the representation of the dependency path. While standard models use un-lexicalized trees—containing only POS tags and dependency labels—this model attaches actual words as terminal nodes to provide semantic context.

Comparison between un-lexicalized (a) and lexicalized (b) dependency trees.

To handle the resulting sparsity, the authors employ a modified SubSet Tree (SST) kernel. Instead of requiring exact string matches, the kernel uses word embeddings and a Gaussian function to calculate semantic similarity. This "smoothing" allows the model to recognize that words like "hired" and "employed" are functionally similar in a relational context, even if they appear as different strings in the training data.

Experimental Results

The model was evaluated on the Penn Treebank, New York Times (NYT), and ClueWeb datasets. Key findings include:

  • Performance Gains: The system consistently outperformed previous benchmarks, showing absolute F1 improvements of 4% to 16%.
  • Precision: On in-domain data (Penn Treebank), the lexicalized model reached 95.9% precision, a significant jump from the 85.7% un-lexicalized baseline.
  • The Named Entity Paradox: Interestingly, excluding specific entity names (e.g., "Federer") while keeping relational words (e.g., "hired") improved performance. Specific names often introduced noise, whereas the relationship is primarily defined by the verbs and prepositions connecting them.
  • Robustness: The embedding-based approach proved more effective than traditional Brown clustering, especially on noisy web data (ClueWeb), suggesting that continuous vector representations capture more nuance for semantic tasks.

Conclusion

By integrating lexical information with syntactic trees through word embeddings, this research demonstrates that semantic "smoothing" is a powerful tool for Open IE. While the system is sensitive to the quality of the initial dependency parse, it provides a robust framework for extracting relations from diverse text sources. Future work will explore the inclusion of context words outside the shortest dependency path to further refine the model's accuracy.

Summary updated

BibTeX
@inproceedings{xu2015lexicalized,
  author = {Ying Xu and Christoph Ringlstetter and Mi-Young Kim and Grzegorz Kondrak and Randy Goebel and Yusuke Miyao},
  title = {{A Lexicalized Tree Kernel for Open Information Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2015},
  url = {https://openreview.net/forum?id=iVYIEMFHqTb},
}
Figure