Abstract
Existing methods for single document keyphrase extraction usually make use of only the information contained in the specified document. This paper proposes to use a small number of nearest neighbor documents to provide more knowledge to improve single document keyphrase extraction. A specified document is expanded to a small document set by adding a few neighbor documents close to the document, and the graph-based ranking algorithm is then applied on the expanded document set to make use of both the local information in the specified document and the global information in the neighbor documents. Experimental results demonstrate the good effectiveness and robustness of our proposed approach.
Introduction
Keyphrase extraction is the task of automatically identifying the most important phrases in a document to serve as a summary or index. While authors of academic papers usually provide these, the vast majority of web content—specifically news articles—does not have them.
The challenge is that a single news article is often too short to provide enough statistical evidence (like word frequency or co-occurrence patterns) to reliably identify the best keyphrases. Existing unsupervised methods (like TFIDF or TextRank) treat documents as independent islands. This paper argues that this is a missed opportunity: if you are reading an article about an earthquake, other articles about the same earthquake contain valuable clues (like the words "victim" or "aftershock") that can help validate which phrases are truly salient.
The ExpandRank Framework
The proposed framework, ExpandRank, operates in two main stages:
- Neighborhood Construction: Given a target document $d_0$, the system uses cosine similarity (based on TFIDF) to find the $k$ most similar documents in a larger corpus. This creates an "expanded document set."
- Keyphrase Extraction:
- Neighborhood-level Word Evaluation: The system builds a Global Affinity Graph. Nodes are candidate words (filtered to include only nouns and adjectives). Edges are drawn between words that co-occur within a sliding window. Crucially, the edge weights are adjusted by the similarity of the documents in which the words appear—meaning words from highly similar "neighbors" carry more weight.
- Graph Ranking: A PageRank-like random walk algorithm is run on this graph to calculate a saliency score for every word.
- Document-level Extraction: Finally, the system looks back at the original document $d_0$. It identifies candidate phrases (sequences of nouns/adjectives) and scores them by summing the saliency scores of their constituent words.
Experimental Setup
The researchers used the DUC2001 dataset, which contains 308 news articles across 30 topics. Since no "gold standard" for keyphrases existed for this set, the authors had two graduate students manually annotate the documents, achieving a Kappa inter-agreement score of 0.70. This annotated version of DUC2001 serves as the benchmark for their evaluation.
Key Insights and Parameter Sensitivity
- The Power of Context: Global information from neighbors can compensate for local sparsity. A word might appear only once in the target document, but if it appears frequently in the neighborhood, its importance is boosted.
- The "Goldilocks" Zone for Neighbors ($k$): The authors found that $k=5$ is the sweet spot. Using too few neighbors doesn't provide enough context, while using too many (e.g., $k=15$) introduces "noise" from documents that aren't truly related, which actually hurts performance.
- Similarity as Confidence: The algorithm doesn't treat all neighbors equally. It uses the cosine similarity between the target and the neighbor as a "confidence value," ensuring that the most relevant neighbors have the strongest influence on the final scores.
- Window Size Robustness: Interestingly, once the co-occurrence window size ($w$) is greater than 2, the performance remains very stable. This suggests the method is robust and doesn't require hyper-fine tuning of the window parameter.
Results
ExpandRank consistently outperformed the baselines across Precision, Recall, and F-measure. As shown in the table below, ExpandRank with $k=5$ achieves the best F-measure of 0.317, a significant improvement over TFIDF (0.254) and SingleRank (0.272).
Conclusion
ExpandRank demonstrates that "neighborhood knowledge" is a potent feature for NLP tasks, particularly when individual documents are short. While building a graph for an expanded set is more computationally expensive, this can be mitigated by batch processing. Future work includes testing on larger, more diverse datasets and integrating other extraction algorithms into this neighborhood framework.
Summary updated