Abstract
Distant supervision (DS) has been widely used to automatically construct (noisy) labeled data for relation extraction (RE). Given two entities, distant supervision exploits sentences that directly mention them for predicting their semantic relation. We refer to this strategy as 1-hop DS, which unfortunately may not work well for long-tail entities with few supporting sentences. In this paper, we introduce a new strategy named 2-hop DS to enhance distantly supervised RE, based on the observation that there exist a large number of relational tables on the Web which contain entity pairs that share common relations. We refer to such entity pairs as anchors for each other, and collect all sentences that mention the anchor entity pairs of a given target entity pair to help relation prediction. We develop a new neural RE method REDS2 in the multi-instance learning paradigm, which adopts a hierarchical model structure to fuse information respectively from 1-hop DS and 2-hop DS. Extensive experimental results on a benchmark dataset show that REDS2 can consistently outperform various baselines across different settings by a substantial margin.
Introduction
Relation Extraction (RE) is the task of identifying semantic links between entities in text (e.g., recognizing that "LeBron James" and "Akron" share a born_in relation). To avoid the high cost of manual labeling, researchers use Distant Supervision (DS), which automatically aligns Knowledge Bases with raw text.
However, standard DS (referred to as "1-hop DS") only looks at sentences that directly mention the target entity pair. This creates a massive bottleneck for "long-tail" entities: in the popular New York Times (NYT) dataset, nearly 75% of entity pairs appear in only one sentence. If that single sentence is noisy or vague, the model fails. The REDS2 framework seeks to bridge this gap by finding "extra" evidence from similar entities found in structured tables.
Methodology: The REDS2 Framework
The authors introduce 2-hop Distant Supervision and a corresponding model, REDS2 (Relation Extraction with 2-hop DS). The process works in three main stages:
- Table-aided Instance Expansion: The model identifies "anchor" pairs from Wikipedia tables—other entities in the same columns that likely share the same relation (e.g., using Mark Aguirre/Chicago as an anchor for LeBron James/Akron). It then collects all sentences mentioning these anchors to create a "2-hop bag" of supporting evidence.
- Sentence Encoding: Every sentence is processed using a Piecewise Convolutional Neural Network (PCNN) to create a vector representation.
- Hierarchical Bag Aggregation: This core innovation uses sentence-level attention to de-emphasize noise and bag-level fusion to balance direct (1-hop) and indirect (2-hop) sources.
Key Results and Insights
REDS2 achieved an AUC of 44.7, significantly outperforming top baselines like PCNN+HATT (42.0) and RESIDE (41.5). Key findings include:
- Solving the Long-Tail: In the "SINGLE" test setting (where entity pairs have only one supporting sentence), REDS2 showed a 25% improvement in AUC over standard models.
- Zero-Shot Potential: In experiments on entity pairs with zero direct sentences, relying only on 2-hop anchor sentences yielded an AUC of 54.7, proving the model can extract relations for entities it has never seen together in text.
- Tables as Relational Shortcuts: Web tables allow the model to "triangulate" relations that are poorly represented in prose, leveraging the high probability that entities in the same columns share semantic relationships.
Conclusion and Future Work
While REDS2 significantly improves performance for long-tail entities, it introduces higher computational overhead due to larger bag sizes. Future research directions include using Reinforcement Learning to better filter noisy sentences and incorporating table headers to further refine anchor selection.
Summary updated