[Paper Review] Towards Zero-resource Cross-lingual Entity Linking

Shuyan Zhou, Shruti Rijhwani+1 more|arXiv (Cornell University)|Sep 29, 2019

Topic Modeling25 references4 citations

TL;DR

This paper proposes a zero-resource approach to cross-lingual entity linking (XEL) that improves performance in low-resource settings by combining neural and lookup-based candidate generation, using language-agnostic disambiguation features, and applying non-linear feature combination. It achieves 6–23% gains in end-to-end XEL accuracy across four low-resource languages (Oromo, Tigrinya, Kinyarwanda, Sinhala), demonstrating that performance degradation in real low-resource scenarios can be significantly mitigated with minimal resource assumptions.

ABSTRACT

Cross-lingual entity linking (XEL) grounds named entities in a source language to an English Knowledge Base (KB), such as Wikipedia. XEL is challenging for most languages because of limited availability of requisite resources. However, much previous work on XEL has been on simulated settings that actually use significant resources (e.g. source language Wikipedia, bilingual entity maps, multilingual embeddings) that are unavailable in truly low-resource languages. In this work, we first examine the effect of these resource assumptions and quantify how much the availability of these resource affects overall quality of existing XEL systems. Next, we propose three improvements to both entity candidate generation and disambiguation that make better use of the limited data we do have in resource-scarce scenarios. With experiments on four extremely low-resource languages, we show that our model results in gains of 6-23% in end-to-end linking accuracy.

Motivation & Objective

• To empirically evaluate the impact of resource scarcity on existing XEL systems in real low-resource settings, rather than simulated ones.
• To address the significant performance drop in XEL when key resources like bilingual maps, multilingual embeddings, or source-language Wikipedia are unavailable.
• To develop methods that maximize performance with minimal resources, focusing on candidate generation and disambiguation in truly low-resource scenarios.
• To demonstrate that language-agnostic features and non-linear feature combination can improve disambiguation transfer to low-resource languages.

Proposed method

• Proposes a hybrid candidate generation method that combines lookup-based (WikiMention) and pivot-based (Pivoting) approaches to improve recall by 9–24% across low-resource languages.
• Introduces language-agnostic disambiguation features trained on English Wikipedia and directly transferred to low-resource languages without language-specific adaptation.
• Employs a non-linear feature combination method (Burn) that iteratively combines features to better model complex interactions in low-resource settings.
• Uses a two-stage pipeline: first generating diverse candidate entities via hybrid methods, then disambiguating using a model trained on English data but applied to low-resource languages.
• Validates the approach on four low-resource languages—Oromo, Tigrinya, Kinyarwanda, and Sinhala—using English Wikipedia as the target KB.
• Evaluates performance using gold candidate recall and end-to-end linking accuracy, comparing baseline, hybrid, and improved disambiguation models.

Experimental results

Research questions

• RQ1 How do standard XEL systems perform when resources like source-language Wikipedia, bilingual entity maps, and multilingual embeddings are scarce or absent in real low-resource settings?
• RQ2 Can a hybrid candidate generation method combining lookup and neural approaches improve recall in low-resource languages where individual methods fail?
• RQ3 To what extent can language-agnostic disambiguation features trained on English data be effectively transferred to low-resource languages without fine-tuning?
• RQ4 Does non-linear feature combination outperform linear or greedy feature combination in low-resource XEL scenarios?
• RQ5 What is the relative contribution of improved candidate generation versus disambiguation in reducing end-to-end performance degradation in low-resource XEL?

Key findings

• The performance of existing XEL systems degrades significantly in real low-resource settings due to the absence of key resources like bilingual maps and multilingual embeddings.
• The hybrid candidate generation method increases gold candidate recall by 9–24% across four low-resource languages, especially benefiting languages with sparse Wikipedia coverage like Oromo and Kinyarwanda.
• The proposed non-linear feature combination method (Burn) improves disambiguation accuracy by 0.2–3.3% over greedy combination, demonstrating the value of flexible feature interaction.
• The combination of hybrid candidate generation and improved disambiguation yields 6–23% end-to-end accuracy gains over the baseline system, with the largest improvements (13–23%) in the most resource-scarce languages.
• Despite strong performance on English, the language-agnostic feature set does not consistently outperform the baseline on low-resource languages, suggesting a domain shift between English and low-resource data that limits transfer.
• The results show that candidate generation is a major bottleneck in low-resource XEL, and improving it can significantly increase the headroom for disambiguation to improve overall accuracy.