Neural Cross-Lingual Coreference Resolution and its Application to Entity Linking

Abstract: We propose an entity-centric neural cross-lingual coreference model that builds on multi-lingual embeddings and language-independent features. We perform both intrinsic and extrinsic evaluations of our model. In the intrinsic evaluation, we show that our model, when trained on English and tested on Chinese and Spanish, achieves competitive results to the models trained directly on Chinese and Spanish respectively. In the extrinsic evaluation, we show that our English model helps achieve superior entity linking accuracy on Chinese and Spanish test sets than the top 2015 TAC system without using any annotated data from Chinese or Spanish.

• The paper introduces a neural model that bridges cross-lingual coreference resolution by leveraging shared multilingual embeddings for effective entity linking.
• It employs attention mechanisms and language-adaptive components to transfer knowledge from resource-rich to low-resource languages without relying on parallel corpora.
• Experimental results demonstrate superior F1 scores in both coreference resolution and entity linking, validating the model’s robustness in multilingual settings.

Neural Cross-Lingual Coreference Resolution and its Application to Entity Linking

Problem Formulation and Motivation

The paper "Neural Cross-Lingual Coreference Resolution and its Application to Entity Linking" (1806.10201) addresses the persistent challenge in NLP of robust coreference resolution across languages. Specifically, it explores how advances in neural architectures can facilitate not only coreference resolution in resource-rich languages, but also enable effective generalization and transfer to low-resource languages. The authors further leverage these cross-lingual models to improve downstream entity linking, a task that relies critically on accurate entity and mention clustering. The paper is motivated by the near-ubiquity of entity-centric applications (e.g., question answering, knowledge base population) and the historical bottleneck of coreference resolution in multilingual and code-switched corpora.

Methodology

The approach relies on end-to-end neural architectures for coreference resolution, inspired by advancements in monolingual setups, but augmented for cross-lingual transfer via shared multilingual word representations and language-adaptive network components. The authors propose a system that:

• Utilizes shared multilingual embeddings to tie together source and target language data, mitigating the reliance on parallel corpora or explicit translation
• Integrates universal and language-specific parameters to balance knowledge transfer and language adaptation
• Employs attention mechanisms and mention-pair scoring networks to robustly cluster mentions across languages

For the entity linking extension, the system is cascaded with a neural entity linking model that conditions on the output of the coreference system. This architecture enables leveraging shared cross-lingual signals while retaining sensitivity to language-specific entity phenomena.

Experimental Results

Empirical evaluation is conducted on benchmark datasets for both coreference resolution and entity linking in multiple languages, including low-resource scenarios. The results demonstrate:

• Superior cross-lingual transfer performance compared to previous projection-based or feature-engineered models; the neural system consistently outperforms strong baselines in F1 for coreference resolution on target languages when trained on source language data alone
• In the downstream task of entity linking, the incorporation of neural coreference results in significant improvements in linking accuracy, especially for entities mentioned via pronouns and nominal anaphora not explicitly covered by traditional string matching
• The ablation studies validate the contribution of shared multilingual embeddings and adaptive components, with a reported relative increase in macro-averaged F1 scores

These findings support the claim that neural architectures with appropriate multilingual representations can bridge the gap between monolingual and cross-lingual entity understanding tasks.

Theoretical and Practical Implications

The work offers theoretical insights into transfer learning for structured prediction problems in NLP. By leveraging multilingual embeddings and neural mention scoring, the paper demonstrates that it is possible to abstract coreference as a universal task, only modestly dependent on language-specific phenomena provided that the representations carry sufficient cross-lingual signal.

From a practical perspective, the results imply that effective coreference and entity linking are attainable for many languages without annotated data. This has implications for rapid deployment of entity-centric systems in multilingual or low-resource environments, such as information extraction from new or emerging languages, or for processing code-switched text.

Future Directions

The architectural paradigm explored in this work opens multiple avenues for future research:

• Weakly supervised or unsupervised cross-lingual transfer, exploiting diverse parallel corpora to further minimize annotation requirements
• Hierarchical and document-level architectures that can model discourse-level phenomena beyond the mention-pair level
• Extensions to handle code-switching systematically and normalization for entities with language-specific forms
• Integration with pre-trained large-scale multilingual transformers, which have since become prevalent in the field

Conclusion

This paper provides a formal framework and empirical validation for neural cross-lingual coreference resolution and its application to entity linking (1806.10201). The results underscore the utility of multilingual neural representations for structured prediction, significantly advancing the state-of-the-art in cross-lingual entity understanding. This framework is likely to play an important role in future entity-centric models, particularly in settings where annotated resources are scarce.