Go for a Walk and Arrive at the Answer: Reasoning Over Paths in Knowledge Bases using Reinforcement Learning
About
Knowledge bases (KB), both automatically and manually constructed, are often incomplete --- many valid facts can be inferred from the KB by synthesizing existing information. A popular approach to KB completion is to infer new relations by combinatory reasoning over the information found along other paths connecting a pair of entities. Given the enormous size of KBs and the exponential number of paths, previous path-based models have considered only the problem of predicting a missing relation given two entities or evaluating the truth of a proposed triple. Additionally, these methods have traditionally used random paths between fixed entity pairs or more recently learned to pick paths between them. We propose a new algorithm MINERVA, which addresses the much more difficult and practical task of answering questions where the relation is known, but only one entity. Since random walks are impractical in a setting with combinatorially many destinations from a start node, we present a neural reinforcement learning approach which learns how to navigate the graph conditioned on the input query to find predictive paths. Empirically, this approach obtains state-of-the-art results on several datasets, significantly outperforming prior methods.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Link Prediction | FB15k-237 (test) | Hits@1045.6 | 419 | |
| Link Prediction | WN18RR (test) | Hits@1051.3 | 380 | |
| Link Prediction | FB15k-237 | MRR20.5 | 280 | |
| Knowledge Graph Completion | WN18RR | Hits@10.351 | 165 | |
| Temporal Knowledge Graph reasoning | ICEWS 18 | Hits@100.33 | 60 | |
| Link Prediction | UMLS | Hits@1096.8 | 56 | |
| Temporal Knowledge Graph reasoning | ICEWS 14 | Hits@125.7 | 48 | |
| Link Prediction | Kinship | MRR0.72 | 36 | |
| Link Prediction | NELL-995 (test) | MRR20.1 | 27 | |
| Temporal Knowledge Graph reasoning | GDELT | MRR12.1 | 25 |