A Data Driven Approach for Person Name Disambiguation in Web Search Results

Agustín D. Delgado, Raquel Martínez, Víctor Fresno, Soto MontalvoCOLING 2014

TL;DR

This paper introduces an unsupervised approach for clustering web search results to disambiguate individuals with the same name. Their method leverages capitalized n-grams and a novel heuristic for social media profiles, achieving competitive results without requiring labeled training data.

What Problem Does This Paper Address?

When searching for a person's name online, especially common ones, search engines often return results for multiple individuals sharing that name. This ambiguity makes it difficult for users to find information about the specific person they are looking for. The challenge lies in automatically grouping web pages that refer to the same individual and determining how many distinct individuals are represented in the search results, all without human intervention or pre-existing knowledge bases. Existing state-of-the-art methods often rely on supervised learning, requiring labeled data and manual parameter tuning, which is costly and time-consuming. Furthermore, many current approaches struggle with the increasing prevalence of social media profiles in search results, which can introduce noise and lead to inaccurate clustering.

Methods

The proposed approach, named UPND (Unsupervised Person Name Disambiguator), consists of two main phases: web page representation and web page grouping.

  • Feature Selection: The core idea is that capitalized n-grams (sequences of capitalized words) are strong indicators of an individual's identity, often representing named entities (people, organizations, locations) or other relevant information not always captured by standard Named Entity Recognition (NER) tools. The authors hypothesize that longer capitalized n-grams are more indicative of the same individual.
  • Weighting Functions: They experiment with TF, TF-IDF, and a z-score weighting function. The z-score is particularly interesting as it measures how frequently an n-gram appears in a specific document relative to its general distribution across all documents for that person's name.
  • Similarity Functions: Cosine distance and a weighted Jaccard coefficient are used to measure the similarity between web pages.
  • Clustering Algorithm: The UPND algorithm iteratively groups web pages based on the similarity of their capitalized n-gram representations. A key innovation is a threshold mechanism that dynamically adjusts based on the number of shared n-grams between two pages, avoiding penalization due to page size differences.
  • Social Media Heuristic (SUPND): To address the challenge of social media profiles, they introduce SUPND (Social UPND). This heuristic restricts comparisons between web pages originating from the same social networking service, acknowledging that different profiles on the same platform are likely to belong to different individuals. However, it allows these pages to be clustered with non-social pages, enabling transitive clustering.

Datasets

The authors evaluate their approach on several benchmark datasets:

  • WePS-1, WePS-2, and WePS-3: These are competitive evaluation campaigns for Web People Search, providing annotated datasets of English person names and search engine results. These datasets contain relatively few social media profiles.
  • ECIR2012 corpus: This dataset is highlighted as a more realistic scenario as it includes a significant number of social media profiles (Facebook, Twitter, LinkedIn) for Dutch person names, retrieved from various search engines.

Key Insights and Ideas

  • Capitalized n-grams are powerful features: The paper strongly supports the hypothesis that capitalized n-grams are more effective than standard Named Entities for person name disambiguation, leading to better performance.
  • Unsupervised learning is viable: The success of UPND demonstrates that high-quality person name disambiguation can be achieved without the need for labeled training data, reducing the effort and cost associated with supervised methods.
  • Social media requires special handling: The prevalence of social media profiles poses a significant challenge. The proposed SUPND heuristic offers a promising way to integrate these noisy but important data sources into the clustering process.
  • Dynamic thresholding is crucial: The proposed threshold mechanism, which considers shared n-grams and page size, is key to robust clustering.

Key Results

  • UPND outperforms unsupervised baselines: Across the WePS datasets, UPND consistently outperforms unsupervised baselines like ALL_IN_ONE, ONE_IN_ONE, and Fast AP.
  • Competitive with supervised methods: UPND achieves results comparable to, and in some cases better than, supervised systems, particularly on the WePS-1 dataset.
  • Z-score and Cosine combination is optimal: The combination of the z-score weighting function and cosine similarity function yielded the best overall performance (F0.5) across the WePS corpora.
  • SUPND improves performance on social media data: For the ECIR2012 corpus, which heavily features social media profiles, SUPND significantly outperforms UPND and the UvA system (which uses supervision), demonstrating the effectiveness of its social media heuristic.

Limitations and Future Work

  • Recall improvement: While precision is high, the authors aim to improve recall without sacrificing precision.
  • Filtering noisy n-grams: Further work is needed to filter out potentially noisy capitalized n-grams.
  • Handling lowercase content: The current approach primarily focuses on capitalized tokens. An alternative representation for web pages with predominantly lowercase content is suggested.
  • Multilingual disambiguation: The current work is monolingual (English and Dutch). Extending the approach to handle multilingual scenarios is a significant future direction.

Summary updated

BibTeX
@inproceedings{delgado2014data,
  author = {Agustín D. Delgado and Raquel Martínez and Víctor Fresno and Soto Montalvo},
  title = {{A Data Driven Approach for Person Name Disambiguation in Web Search Results}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
}
Figure