Learning Feature Representations for Keyphrase Extraction

Corina Florescu, Wei JinAAAI 2018 · DOI

Abstract

Keyphrase extraction (KE) is a critical task for identifying the most significant terms within a document. Traditional methods rely heavily on manual feature engineering or unsupervised graph centrality measures, which often struggle with domain generalization. This paper introduces SurfKE, a framework that treats documents as graphs and employs biased random walks and representation learning to automatically discover structural features. By learning the topological "role" of words within a document's specific context, SurfKE outperforms both supervised and unsupervised baselines across diverse datasets.

Introduction

Historically, Keyphrase Extraction has been approached through two primary lenses: unsupervised methods using graph centrality (e.g., PageRank) or supervised methods utilizing hand-crafted features like part-of-speech tags and word frequency. However, hand-crafted features are often brittle and require significant domain expertise. While semantic embeddings like Word2Vec capture meaning, they do not inherently reflect the structural importance of a word within a specific document. SurfKE bridges this gap by learning features directly from the document's internal structure, moving away from manual engineering toward automated feature discovery.

Methodology

The SurfKE framework follows a three-step pipeline designed to capture the structural characteristics of a text:

  1. Graph Construction: The document is converted into an undirected word graph where unique words (filtered by part-of-speech) serve as vertices. Edges are established between words appearing within a sliding window, with weights determined by co-occurrence frequency.
  2. Learning Feature Representations: Instead of direct ranking, the model performs biased random walks across the graph to generate node sequences. These sequences are processed by a representation learning model to produce low-dimensional embeddings that capture the topological role of each word.
  3. Candidate Phrase Scoring: For multi-word phrases, the model calculates a phrase-level feature vector by averaging the constituent word vectors. A Gaussian Naive Bayes classifier is then trained to distinguish keyphrases from non-keyphrases based on these learned representations.

Experimental Setup

SurfKE was evaluated on a synthesized dataset of 1,308 documents compiled from three diverse benchmarks:

  • DUC: News articles.
  • Inspec: Scientific abstracts.
  • MEDLINE/PubMed: Medical and biological research papers.

This variety ensures that the model's performance is tested across different writing styles and technical domains.

Results and Discussion

SurfKE was compared against standard supervised baselines (KEA, Maui) and unsupervised baselines (KpMiner, PositionRank). The experimental results, summarized in the table below, demonstrate that SurfKE consistently outperforms these methods.

Performance comparison of SurfKE against baseline models across Precision, Recall, and F1-score.

As shown in the results, SurfKE achieved an F1-score of 0.260, surpassing the next best baseline, PositionRank (0.240). Notably, the model showed a significant improvement in Precision (0.220), representing a roughly 50% relative improvement over KEA and an 8.9% improvement over PositionRank.

Key Insights

  • Structure over Semantics: Graph-based embeddings effectively capture how a word functions within the specific architecture of a text, providing a stronger signal for keyphrase importance than general semantic meaning.
  • Automated Feature Discovery: Biased random walks allow the model to discover complex patterns (such as positional importance) without explicit manual instruction.
  • Generalization: By learning from graph structures rather than domain-specific statistics, SurfKE maintains robust performance across news, scientific, and medical texts.

Conclusion and Future Work

SurfKE demonstrates the power of structural representation learning for keyphrase extraction. Future work may explore different random walk strategies, such as focusing on document-initial neighborhoods, or incorporating explicit topical guidance to further refine the embeddings. Additionally, exploring more complex downstream classifiers like Random Forests or Neural Networks could potentially yield even higher performance gains.

Summary updated

BibTeX
@inproceedings{florescu2018learning,
  author = {Corina Florescu and Wei Jin},
  title = {{Learning Feature Representations for Keyphrase Extraction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.12144},
}
Figure