[Paper Review] Building a Fine-Grained Entity Typing System Overnight for a New X (X = Language, Domain, Genre)

Lifu Huang, Jonathan May+2 more|arXiv (Cornell University)|Mar 10, 2016

Topic ModelingComputer Science47 references21 citations

TL;DR

This paper proposes an unsupervised, zero-shot fine-grained entity typing framework that combines general entity embeddings, context-specific representations via linguistic structures (e.g., AMR, dependency trees), and knowledge base linking to discover domain- and language-specific types without labeled data or predefined schemas. It achieves state-of-the-art performance comparable to supervised systems across diverse languages (English, Chinese, Japanese, Hausa, Yoruba) and domains (news, biomedical), demonstrating high portability and adaptability to new genres, languages, and domains.

ABSTRACT

Recent research has shown great progress on fine-grained entity typing. Most existing methods require pre-defining a set of types and training a multi-class classifier from a large labeled data set based on multi-level linguistic features. They are thus limited to certain domains, genres and languages. In this paper, we propose a novel unsupervised entity typing framework by combining symbolic and distributional semantics. We start from learning general embeddings for each entity mention, compose the embeddings of specific contexts using linguistic structures, link the mention to knowledge bases and learn its related knowledge representations. Then we develop a novel joint hierarchical clustering and linking algorithm to type all mentions using these representations. This framework doesn't rely on any annotated data, predefined typing schema, or hand-crafted features, therefore it can be quickly adapted to a new domain, genre and language. Furthermore, it has great flexibility at incorporating linguistic structures (e.g., Abstract Meaning Representation (AMR), dependency relations) to improve specific context representation. Experiments on genres (news and discussion forum) show comparable performance with state-of-the-art supervised typing systems trained from a large amount of labeled data. Results on various languages (English, Chinese, Japanese, Hausa, and Yoruba) and domains (general and biomedical) demonstrate the portability of our framework.

Motivation & Objective

• Address the challenge of fine-grained entity typing in new domains, genres, and languages where annotated data and predefined type schemas are unavailable.
• Overcome limitations of existing supervised methods that rely on noisy distant supervision and pre-defined type sets.
• Develop a flexible, unsupervised framework that discovers fine-grained types through general semantics, context-specific representations, and domain knowledge.
• Enable rapid deployment across low-resource languages and specialized domains without retraining or manual feature engineering.

Proposed method

• Learn general entity embeddings from large-scale Wikipedia data to capture general semantics of entity mentions.
• Construct context-specific representations using linguistic structures such as Abstract Meaning Representation (AMR) and dependency relations to model local context semantics.
• Integrate knowledge base linking to enrich entity representations with domain-specific knowledge, particularly for biomedical and rare entities.
• Apply a joint hierarchical clustering and linking algorithm to group mentions into types without prior type definitions or labeled data.
• Use a weighted fusion of three representations—general semantics, context-specific features, and knowledge base links—with learned weights to optimize clustering.
• Incorporate an unsupervised entity linking system to improve clustering quality and generate corpus-specific typing schemas automatically.

Experimental results

Research questions

• RQ1 Can an unsupervised entity typing framework achieve performance comparable to supervised systems without any labeled data?
• RQ2 How effective is the integration of general semantics, context-specific linguistic structures, and knowledge base representations in discovering fine-grained types?
• RQ3 To what extent can the framework be adapted to new domains and languages without retraining or manual intervention?
• RQ4 How do different linguistic structures (AMR, dependency, bag-of-words) impact context representation quality and typing accuracy?
• RQ5 Can the framework generate meaningful, human-interpretable type names without supervision?

Key findings

• The framework achieves competitive performance on news and biomedical domains, with typing accuracy comparable to state-of-the-art supervised systems trained on large labeled datasets.
• On English, Chinese, and Japanese news, the framework achieves high accuracy, with the lowest ambiguity (below 0.05) in Chinese and Japanese, contributing to strong performance.
• For low-resource languages like Hausa and Yoruba, the framework achieves 85.42% and 72.26% mention-level typing accuracy, respectively, even without pre-trained embeddings or KBs.
• AMR and dependency-based context representations outperform bag-of-words methods by capturing semantically meaningful relations (e.g., 'carrying arms' for 'Kuwolsan' as 'Ship').
• The optimal fusion weights for representations are 0.45 (general), 0.05 (context), and 0.50 (knowledge) in the biomedical domain, and 0.45 (general), 0.20 (context), and 0.35 (knowledge) in news, showing context’s varying importance.
• Human evaluation shows that type naming performance reaches ~90% accuracy when the threshold λ is set to 0.8, indicating the framework generates semantically meaningful and interpretable type labels.