Abstract
Yi Chen, Jiayang Cheng, Haiyun Jiang, Lemao Liu, Haisong Zhang, Shuming Shi, Ruifeng Xu. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
2022.
Introduction
Fine-Grained Entity Typing (FGET) is the task of assigning specific labels (e.g., /organization/company/tech_company) to an entity mention based on its context. While current state-of-the-art models perform well on average, they often stumble on "hard" mentions. These failures typically stem from complex sentence structures that obscure features or ambiguous contexts where the local information is insufficient to distinguish between types. For example, in the sentence "Sharp just signed a deal with Qualcomm," the mention "Sharp" could be a person or a company without additional information.
This paper introduces a novel approach that leverages "sibling mentions"—other mentions in a dataset that share similar semantic types—to resolve these ambiguities. By modeling these relationships in a scalable, heterogeneous graph neural network, the authors significantly improve performance on difficult cases where traditional models struggle.
Methods
The authors propose a Sibling Enhanced Heterogeneous Graph Model. The core idea is that an entity's type isn't just defined by its current sentence, but by the "collective intelligence" found in how similar entities are treated across the entire corpus.
Graph Construction and Sibling Selection
The model builds a graph where nodes represent both mentions and types. It incorporates three edge types: Sibling Edges connecting semantically similar mentions, Hierarchical Edges connecting types (e.g., company is a child of organization), and Label Edges used during training to connect mentions to their ground-truth types.
To find effective "siblings," the authors utilized a typing distribution-based metric. By using a base model to predict a probability distribution over types, they could identify mentions with similar predicted distributions, which proved more effective than simple word-based similarity metrics.
Attentive GNN and Scalable Inference
An attentive Graph Neural Network (GNN) updates representations by aggregating information from siblings and type neighbors. A Type Dropout mechanism is used during training to force the model to learn from siblings rather than relying solely on ground-truth labels. For inference, a new mention is added to the graph and connected to its $K$ nearest siblings from the training set, ensuring the model remains scalable to unseen data without retraining the entire graph.
Key Results
The model was evaluated on three standard benchmarks: OntoNotes (Original), OntoNotes (Augmented), and BBN. It consistently outperformed ten strong baselines across all datasets.
On the standard OntoNotes dataset, the model achieved a Strict Accuracy of 59.2%, outperforming the previous best base model by nearly 4%. On the Augmented OntoNotes dataset, it reached 65.7% accuracy. Most importantly, on the "hardest" 5% of mentions (defined by high prediction entropy in the base model), the model improved accuracy from 51.0% to 57.2%, demonstrating that sibling information effectively clarifies ambiguity.
Conclusion and Future Work
The research demonstrates that leveraging cross-corpus information is a powerful tool for fine-grained entity typing. The typing-based similarity metric suggests that latent semantic meaning is a better bridge between mentions than surface vocabulary.
However, limitations remain. A "Gold" sibling selection experiment suggests a performance ceiling of 80.0% Mi-F1, indicating that the current bottleneck is the similarity metric itself. Future work will explore more sophisticated sibling selection methods and deeper relational graph structures to further enhance performance.
Summary updated