- Abstract
- Task: Text-based methods still lag behind graph embedding-based methods like TransE, RotatE. They identify the key issue is efficient contrastive learning.
- Contribution 1: Introduce three types of negatives (more negatives than before): in-batch negatives, pre-batch negatives, and self negatives
- Contribution 2: Combined with InfoNCE loss, our model SimKGC outperforms other embedding-based methods on several benchmark datasets
- Intro
- Limitation of the current models:
- Results on popular benchmarks tell a different story that text-based methods still lag behind even with pre-trained language models
- Motivation: inspired by the recent progress on contrastive learning
- Limitation of the current models:
- Pipeline
- Bi-encoder architecture. Two encoders are initialized with the same pre-trained language model but do not share parameters.
- Negative sampling: for knowledge graph completion the training data only consists of positive triples. Therefore, negative sampling needs to sample one or more negative triples to train discriminative models.
- In-batch negatives; pre-batch negatives; self-negatives
- loss function: InfoNCE loss with additive margin
- Experiments:
- Comparison between embedding-based methods, text-based methods, and their models with ablation studies.
- Results show that self-negatives make the model much more effective
- Limitations:
- SimKGC does not enable easy-to-understand interpretations
- dealing with false negatives
- Are there any loss functions that perform better than the infoNCE loss?
- Takeaways
- Analysis to get further insights is important to guide future research.
- Ideas may be simple, but strong and solid.