A Graph-Based Approach to Skill Extraction from Text

Ilkka Kivimäki, Alexander Panchenko, Adrien Dessy, Dries Verdegem, Pascal Francq, Hugues Bersini, Marco SaerensEMNLP 2013

TL;DR

This paper introduces "Elisit," a system that extracts professional skills from text documents by leveraging Wikipedia's hyperlink graph and a curated list of LinkedIn skills. It addresses the challenge of automatically identifying expertise in unstructured text, offering practical applications in HR and knowledge management.

What Problem Does This Paper Address?

In today's data-rich environment, organizations struggle with efficiently identifying and managing the expertise within their workforce. This is crucial for tasks like recruiting the right talent, assigning employees to projects based on their skills, and finding suitable reviewers for academic submissions. While "expertise retrieval" is an active research area, there's a gap in methods that can automatically extract specific professional skills from arbitrary text documents, such as an individual's publications or professional profiles. Existing approaches often focus on matching queries to people or broad expertise areas, rather than pinpointing granular skills from unstructured text.

Methods

The Elisit system employs a two-phase approach:

  1. Text-to-Wikipedia (text2wiki) Module: This module first associates an input text document with relevant Wikipedia pages. It achieves this by representing both the input document and Wikipedia pages using various vector space models (VSM), including TF-IDF, LogEntropy, and combinations with LSA and LDA. Cosine similarity is used to measure the semantic similarity between the input text and Wikipedia articles. To handle the vastness of Wikipedia, an inverted index is used for efficient retrieval of the top 200 most similar Wikipedia pages.
  2. Wikipedia-to-Skill (wiki2skill) Module: This module then maps the Wikipedia pages identified in the first phase to a predefined list of professional skills. It utilizes a biased, hub-avoiding version of the Spreading Activation algorithm on Wikipedia's hyperlink graph. Spreading Activation simulates a cognitive associative memory, where activation "spreads" from initial nodes (representing the input text's associated Wikipedia pages) to neighboring nodes (other Wikipedia pages). The "bias" is crucial to prevent popular "hub" pages (like "ISBN") from dominating the results, ensuring more relevant skills are identified. The algorithm's parameters, such as decay factor, friction factor, and the popularity index used for biasing, are carefully tuned through experiments.

Datasets

  • Wikipedia: The system uses a Wikipedia dump from May 2nd, 2012, focusing on encyclopedia pages and their hyperlinks.
  • LinkedIn Skills: A list of 27,153 professional skills was extracted from LinkedIn in September 2012. This list is crucial as it includes links to corresponding Wikipedia articles, forming the basis for skill identification.

Key Insights and Ideas

  • Wikipedia as a Semantic Network for Skills: The paper demonstrates that Wikipedia's hyperlink structure, when combined with a curated skill list, can effectively serve as a rich semantic network for skill extraction.
  • The Importance of Biasing Spreading Activation: A key finding is that a biased Spreading Activation algorithm, specifically designed to avoid popular "hub" nodes, significantly improves retrieval results. This addresses a common pitfall in graph-based methods where overly connected nodes can skew outcomes.
  • Simpler VSMs Outperform Complex Ones for Text-to-Wikipedia: For the initial text-to-Wikipedia mapping, simpler vector space models like TF-IDF and LogEntropy performed better than more complex LSA and LDA models on the "See also" dataset. This suggests that for this specific task, a direct term-based representation might be more effective.
  • Bidirectional Links Matter: The experiments suggest that bidirectional links in the Wikipedia graph carry significant semantic weight, and increasing their influence improves skill retrieval.

Key Results

  • The Elisit system successfully extracts relevant skills from text documents, as demonstrated by qualitative examples and quantitative evaluations.
  • The biased Spreading Activation approach, particularly with a negative biasing parameter (e.g., α = -0.4), consistently improved retrieval metrics like Precision and Recall.
  • Using PageRank as the popularity index for biasing yielded better results than input degree or HITS.
  • The TF-IDF and LogEntropy models for the text2wiki module outperformed LSA and LDA in the context of the whole Elisit system.
  • The system can provide results within a few seconds, making it practical for web service applications.

Limitations and Future Work

  • Ambiguity of LinkedIn Skills: Some LinkedIn skills are not well-represented in Wikipedia due to ambiguity, which can lead to misguiding results. The paper addresses this by selecting a fixed set of representative skills for evaluation.
  • Parameter Optimization: The paper acknowledges that a fully automatic optimization scheme for all parameters would be challenging due to the number of parameters and the size of the Wikipedia graph.
  • Future Work:
    • Computing link weights based on the content of linked pages using their vector space representations.
  • Extending the methodology to other languages.
  • Applying the core methodology to different domains by substituting the LinkedIn skills with Wikipedia pages from other areas.

Summary updated

BibTeX
@inproceedings{kivimaki2013graph,
  author = {Ilkka Kivimäki and Alexander Panchenko and Adrien Dessy and Dries Verdegem and Pascal Francq and Hugues Bersini and Marco Saerens},
  title = {{A Graph-Based Approach to Skill Extraction from Text}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
}
Figure