[Paper Review] Unsupervised Keyphrase Extraction with Multipartite Graphs

Florian Boudin|arXiv (Cornell University)|Mar 23, 2018

Advanced Text Analysis Techniques4 citations

TL;DR

This paper proposes an unsupervised keyphrase extraction model that uses a multipartite graph to jointly represent keyphrase candidates and topics, leveraging their mutual reinforcement to improve ranking. It introduces a novel edge-weight adjustment mechanism to incorporate selection preferences (e.g., early occurrence), achieving state-of-the-art performance on three benchmark datasets by enhancing topical diversity and ranking accuracy.

ABSTRACT

We propose an unsupervised keyphrase extraction model that encodes topical information within a multipartite graph structure. Our model represents keyphrase candidates and topics in a single graph and exploits their mutually reinforcing relationship to improve candidate ranking. We further introduce a novel mechanism to incorporate keyphrase selection preferences into the model. Experiments conducted on three widely used datasets show significant improvements over state-of-the-art graph-based models.

Motivation & Objective

• To address the lack of topical diversity and coverage in existing graph-based keyphrase extraction models.
• To improve keyphrase ranking by modeling the mutual reinforcement between topics and keyphrase candidates.
• To integrate selection preferences (e.g., position, lexical features) into the graph ranking process without supervision.
• To reduce reliance on topic clustering quality by implicitly enforcing diversity through graph structure.
• To achieve superior performance on standard benchmarks without requiring labeled data or complex parameter tuning.

Proposed method

• Constructs a directed multipartite graph where nodes are keyphrase candidates and topics, with edges only between nodes of different topics.
• Assigns edge weights based on co-occurrence frequency between candidates and topics.
• Applies a graph-theoretic ranking algorithm (TextRank) to score keyphrase candidates.
• Introduces a weight adjustment mechanism that boosts candidates with desirable features (e.g., early occurrence) via edge weight modification.
• Tunes the strength of the weight adjustment using a hyperparameter α, optimized on a validation set.
• Uses hierarchical agglomerative clustering on word stems to group candidates into topics, avoiding reliance on external knowledge bases.

Experimental results

Research questions

• RQ1 Can a multipartite graph structure effectively model the mutual reinforcement between keyphrase candidates and topics to improve ranking?
• RQ2 Does incorporating selection preferences (e.g., position, lexical features) through edge-weight adjustment enhance keyphrase extraction performance?
• RQ3 Can topical diversity be implicitly enforced without hard constraints or topic clustering, reducing sensitivity to clustering errors?
• RQ4 How does the proposed model compare to state-of-the-art unsupervised baselines in terms of F1 and MAP scores?
• RQ5 To what extent does the model reduce over-generation errors caused by word-weight summing in baseline models?

Key findings

• The proposed model achieves the highest F1 and MAP scores across all three datasets (SemEval-2010, Hulth-2003, Marujo-2012), significantly outperforming TopicRank, PositionRank, and Single Topical PageRank.
• On the SemEval-2010 dataset, the model achieves an F1@10 score of 0.568, surpassing the best baseline (PositionRank) by 4.2% in relative improvement.
• The model achieves over 92% topic coverage in the top-10 keyphrases, indicating effective implicit diversity enforcement without hard constraints.
• Removing the weight adjustment mechanism results in a significant drop in performance, especially on F1@5 for SemEval-2010, confirming the importance of preference integration.
• The model reduces over-generation errors by avoiding simple word-weight summing, as evidenced by improved precision and MAP scores.
• Manual inspection reveals that topic-redundant keyphrases are often due to clustering errors or semantic relations (e.g., hypernym-hyponym), suggesting limitations in gold-standard annotations.