Open-World Knowledge Graph Completion

Baoxu Shi, Tim WeningerAAAI 2018 · DOI

TL;DR

This paper introduces "ConMask," a novel approach to Knowledge Graph Completion (KGC) that tackles the challenge of incomplete and rapidly growing knowledge graphs by embracing an "open-world" assumption. ConMask leverages entity descriptions and relationship-dependent content masking to predict missing relationships, even for entities not previously seen in the graph, outperforming existing methods on both open-world and closed-world KGC tasks.

What Problem Does This Paper Address?

Knowledge Graphs (KGs) are powerful tools for organizing information, but they are often incomplete and constantly expanding with new entities and relationships. Traditional Knowledge Graph Completion (KGC) methods operate under a "closed-world assumption," meaning they assume the KG is fixed and can only discover existing relationships. This approach struggles with new or sparsely connected entities, which are common in real-world, dynamic KGs. The paper addresses this limitation by proposing an "open-world KGC" task and a model, ConMask, designed to predict relationships involving unseen entities by utilizing their textual descriptions.

Methods

ConMask is built around three core components:

  • Relationship-Dependent Content Masking: This technique aims to filter out irrelevant information from an entity's description. It uses a similarity score between words in the entity's description and the words in the relationship name to highlight relevant snippets. This is achieved through a novel "Maximal Context-Relationship Weights" (MCRW) function, which adjusts word weights based on their context and similarity to the relationship.
  • Target Fusion: This component uses a fully convolutional neural network (FCN) to extract a fixed-size embedding from the masked content. Unlike traditional FCNs used in computer vision, ConMask's FCNs are adapted for text processing and focus on extracting relevant features without de-convolution.
  • Target Entity Resolution: This final step involves comparing the extracted entity embedding with candidate entities in the KG to predict the most likely target. The paper also introduces a partial list-wise ranking loss function to optimize the model's performance.

Datasets

The paper uses the following datasets:

  • Freebase 15K (FB15k): A standard dataset for KGC, but noted to have issues with reversed/synonym triples and limited textual information.
  • DBPedia50k and DBPedia500k: Two new datasets introduced by the authors, constructed from DBPedia and Wikipedia. These datasets are designed to provide richer textual information for content-based KGC methods and are intended for both open-world and closed-world KGC evaluation.

Key Insights and Ideas

  • Open-World KGC is Crucial for Dynamic KGs: The paper highlights that real-world KGs evolve rapidly, making closed-world assumptions impractical. An open-world approach is necessary to handle the continuous influx of new information.
  • Textual Descriptions are a Powerful Substitute for Topology: When an entity lacks topological connections in a KG, its textual description becomes a vital source of information for inferring relationships.
  • Relationship-Dependent Masking is Key to Mitigating Noise: Simply using raw text descriptions can introduce noise. ConMask's relationship-dependent masking effectively filters out irrelevant words, focusing on information pertinent to the specific relationship being predicted.
  • FCNs are Effective for Feature Extraction from Text: The paper demonstrates that FCNs, when adapted for text processing, can effectively capture salient features from masked textual content to generate meaningful entity embeddings.

Key Results

  • Strong Performance on Open-World KGC: ConMask significantly outperforms baseline methods on the open-world KGC task, showing substantial improvements in metrics like HITS@10 and MRR on the new DBPedia datasets.
  • Competitive Performance on Closed-World KGC: ConMask also demonstrates competitive, and in some cases superior, performance compared to existing closed-world KGC models on standard benchmarks like FB15k and the new DBPedia datasets.
  • DBPedia Datasets are More Challenging: The new DBPedia datasets, particularly DBPedia500k, show more significant performance gains for ConMask, suggesting that the textual information they provide is crucial for its effectiveness.

Limitations and Future Work

  • Sensitivity to Similar Names: The paper acknowledges that ConMask can sometimes rank entities with names similar to the relationship itself very highly, potentially hindering performance. The authors suggest rearranging such entities as a future improvement.
  • Extraction-Based Prediction: ConMask currently relies on information explicitly present in entity descriptions. Future work aims to extend the model to find new or implicit relationships that are not directly stated in the text.

Summary updated

BibTeX
@inproceedings{shi2018open,
  author = {Baoxu Shi and Tim Weninger},
  title = {{Open-World Knowledge Graph Completion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11535},
}
Figure