Logic Embeddings for Complex Query Answering (2021)

huyi / December 2022

0. Abstract

task and challenges

task: answering logical queries over incomplete knowledge bases

challenges:

• imcomplete KG –> call for link prediction
• brute force answering of existential first-order logic queries is exponential in the number of existential variables

previous work

that doesn’t support negation

Recent work of query embeddings provides fast querying, but most approaches model set logic with closed regions, so lack negation.

that support negation

use densities, triggering drawbacks:

1. only improvise logic
2. use expensive distributions
3. poorly model answer uncertainty

this work

• support negation
• improve on density approaches: 1) integrates well-studied t-norm logic and directly evaluates satisfiability, 2) simplifies modeling with truth values, 3) models uncertainty with truth bounds

contribution

• competitively fast and accurate in query answering over large, incomplete knowledge graphs
• outperform on negation queries
• provide improved modeling of answer uncertainty

1. Introduction