ClusType: Effective Entity Recognition and Typing by Relation Phrase-Based Clustering.

Area of Science:

• Natural Language Processing
• Information Extraction
• Machine Learning

Background:

• Entity recognition (ER) is crucial but challenging, especially in dynamic or emerging domains due to increased name ambiguity and context sparsity.
• Existing methods struggle with domain-specific data, necessitating domain-agnostic entity detection.
• Distant supervision offers a viable approach for training ER systems with limited labeled data.

Purpose of the Study:

• To propose a novel relation phrase-based entity recognition framework, ClusType, designed for domain-agnostic ER.
• To develop a data-driven approach for generating entity mention candidates and relation phrases.
• To enhance entity type prediction by leveraging soft clustering of relation phrases and type signatures.

Main Methods:

• ClusType employs data-driven phrase mining to identify entity mentions and relation phrases.
• It enforces soft clustering of relation phrases to propagate type information between argument entities.
• A joint optimization problem is formulated for type propagation and multi-view relation phrase clustering.

Main Results:

• Experiments on news, Yelp reviews, and tweets demonstrate ClusType's effectiveness and robustness.
• The ClusType framework achieved an average of 37% improvement in F1 score over the best compared methods.
• The approach successfully predicts entity types based on relation phrase type signatures and surface name indicators.

Conclusions:

• ClusType offers a robust and effective solution for domain-agnostic entity recognition.
• The proposed relation phrase-based framework with soft clustering significantly improves ER performance.
• This method addresses challenges posed by domain-specific, dynamic, and emerging text collections.