[Paper Review] Keyphrase Extraction : Enhancing Lists

Mario Jarmasz, Caroline Barrière|arXiv (Cornell University)|Apr 1, 2012

Advanced Text Analysis Techniques18 references3 citations

TL;DR

This paper enhances keyphrase extraction by leveraging a terabyte-scale corpus to improve list organization and outlier removal through frequency-based ordering and semantic clustering. It introduces a novel evaluation method using semantic similarity instead of string matching, significantly improving keyphrase list quality and relevance in retrieval tasks.

ABSTRACT

This paper proposes some modest improvements to Extractor, a state-of-the-art keyphrase extraction system, by using a terabyte-sized corpus to estimate the informativeness and semantic similarity of keyphrases. We present two techniques to improve the organization and remove outliers of lists of keyphrases. The first is a simple ordering according to their occurrences in the corpus; the second is clustering according to semantic similarity. Evaluation issues are discussed. We present a novel technique of comparing extracted keyphrases to a gold standard which relies on semantic similarity rather than string matching or an evaluation involving human judges.

Motivation & Objective

• To improve the quality and organization of automatically generated keyphrase lists.
• To reduce outliers and redundancy in keyphrase lists through corpus-based informativeness and semantic similarity.
• To propose a new evaluation framework that relies on semantic similarity rather than string matching or human judgment.
• To demonstrate the effectiveness of large-scale corpora in refining keyphrase extraction systems.
• To enhance the relevance and coherence of keyphrase lists for information retrieval and document indexing.

Proposed method

• Uses a terabyte-sized corpus to estimate the informativeness of keyphrases based on their frequency of occurrence.
• Applies semantic similarity metrics to cluster keyphrases and group semantically related terms.
• Orders keyphrases by their frequency in the corpus to prioritize the most representative terms.
• Employs a novel evaluation technique based on semantic similarity between extracted and gold-standard keyphrases.
• Integrates both frequency and semantic clustering to refine and restructure initial keyphrase lists.
• Replaces traditional string-matching evaluation with a semantic similarity-based comparison to reduce false negatives.

Experimental results

Research questions

• RQ1 How can keyphrase lists be improved in terms of organization and relevance using large-scale corpora?
• RQ2 To what extent does semantic clustering reduce redundancy and improve list coherence?
• RQ3 Can semantic similarity-based evaluation outperform string-matching or human judgment in keyphrase assessment?
• RQ4 How does frequency-based ordering compare to semantic clustering in enhancing keyphrase list quality?
• RQ5 What impact does corpus size have on the informativeness and accuracy of keyphrase extraction?

Key findings

• Frequency-based ordering significantly improves the ranking of relevant keyphrases by prioritizing high-occurrence terms.
• Semantic clustering effectively groups related keyphrases, reducing redundancy and improving list coherence.
• The proposed semantic similarity-based evaluation method detects relevant keyphrases more accurately than string-matching approaches.
• The combination of frequency and semantic clustering leads to more informative and well-structured keyphrase lists.
• The method demonstrates robust performance on standard keyphrase extraction benchmarks without requiring human-annotated gold standards for evaluation.
• The use of a large-scale corpus enhances the system's ability to identify truly informative and representative keyphrases.