TL;DR
The paper introduces 2-hop Distant Supervision (DS), a novel method to improve relation extraction by leveraging the structured nature of Web tables. By identifying "anchor" entity pairs in tables that share similar relations, the authors harvest extra training sentences for "long-tail" entities that rarely appear together in text, significantly outperforming existing state-of-the-art models.
What Problem Does This Paper Address?
Relation Extraction (RE) is the task of identifying the semantic relationship between two entities (e.g., "LeBron James" and "Akron") from raw text. Because manually labeling millions of sentences is impractical, researchers use Distant Supervision (DS). This technique assumes that if a knowledge base says a relationship exists between two entities, any sentence containing both entities likely expresses that relationship.
However, traditional DS (which the authors call "1-hop DS") suffers from the long-tail problem. Many entity pairs are rarely mentioned together in the same sentence. In the popular New York Times (NYT) dataset, nearly 75% of entity pairs have only a single supporting sentence. Without enough data, models struggle to learn accurately. The authors seek to expand the available evidence for these rare pairs without manual labeling.
Methods
The authors propose a new strategy called 2-hop Distant Supervision and a corresponding neural model named REDS2.
- Anchor Entity Pairs: The core idea is to use Web tables (like those on Wikipedia) as a bridge. If "LeBron James" and "Akron" appear in the same row of a table about basketball players and their hometowns, they are a "target pair." The model then looks for other pairs in the same columns (e.g., "Mark Aguirre" and "Chicago"). These are called anchors.
- 2-hop Sentence Collection: Instead of only looking for sentences containing the target pair (1-hop), the model also collects sentences containing the anchor pairs (2-hop). The intuition is that sentences describing the relationship between anchors will use similar language and patterns that apply to the target pair.
- REDS2 Model Architecture:
- Sentence Encoder: Uses a Piecewise Convolutional Neural Network (PCNN) to turn sentences into vector representations.
- Hierarchical Bag Aggregation: Since 2-hop data is noisier than 1-hop data, the model doesn't treat them equally. It uses a sentence-level attention mechanism to find the most relevant sentences in each "bag" (1-hop and 2-hop) and then uses a learned weighting factor to fuse the two bags into a final representation for classification.
Key Insights and Ideas
- Tables as Relational Blueprints: Tables are "semi-structured." Unlike a random sentence, a table column explicitly groups entities that share a property. This makes them a high-quality source for finding "analogous" entity pairs.
- Bridging the Data Gap: 2-hop DS effectively "borrows" data from well-known entities to help understand obscure ones. If you know how the media describes Mark Aguirre's hometown, you can use that knowledge to identify LeBron James's hometown, even if LeBron is rarely mentioned alongside Akron in your specific text corpus.
- Noise Management: The hierarchical structure of REDS2—separating the 1-hop and 2-hop bags before fusion—is what prevents the extra data from overwhelming the signal with noise.
Key Results
- State-of-the-Art Performance: REDS2 achieved an Area Under the Curve (AUC) of 44.7, significantly higher than previous models like PCNN+HATT (42.0) and RESIDE (41.5).
- 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 the baseline.
- Predicting the "Unseen": Most impressively, the authors tested the model on entity pairs that had zero direct sentences in the original corpus. By using only 2-hop data from tables, the model was still able to predict relations with high accuracy (AUC of 54.7).
Limitations and Future Work
- Computational Cost: Inference with 2-hop DS is slower. Because the 2-hop bags are much larger (averaging ~100x more sentences than 1-hop bags), processing takes more time—about 37 seconds for a test pass compared to 12 seconds for the baseline.
- Table Quality: The method relies on the quality of entity linking in tables. Errors in identifying which entity is in which table cell can lead to "false anchors" and incorrect relations.
- Future Directions: The authors suggest exploring smarter sampling methods to reduce the number of 2-hop sentences processed and incorporating table metadata (like headers and captions) to further refine the anchor selection process.
Summary updated