ResqLP: Relation Sequences for Link Prediction in Knowledge Graphs
Author
Bækdahl, Emil Stenderup
Term
4. term
Education
Publication year
2021
Pages
17
Abstract
Vidensgrafer repræsenterer domæneviden som entiteter og relationer mellem dem og bruges i mange informationssystemer. En almindelig opgave er link prediction: at udlede manglende fakta ved at forudsige nye forbindelser. Selvom forskningen skrider frem, kan få modeller forklare, hvorfor de laver en forudsigelse. Vi introducerer ResqLP, en model til link prediction, der lærer probabilistiske førsteordens logikregler—menneskeligt læsbare hvis–så-regler med usikkerhed—ud fra både entiteters semantik og sekvenser af relationer. Modellen indeholder et feature-ekstraktionstrin, der bruger indesluttede delgrafer til at reducere det ellers svære problem med at opregne alle mulige stier mellem entiteter. Vi præsenterer også en generel måde at indfange entiteters semantik på, baseret på relationer i et entitets nabolag. Vores resultater viser, at indesluttede delgrafer fremskynder feature-ekstraktion, selvom meget store vidensgrafer stadig kan give problemer i nogle tilfælde. ResqLP kan lære ikke-trivielle logiske regler, om end kvaliteten varierer, og modellen klarer sig bedst på relativt sparsomt forbundne vidensgrafer med mange forskellige relationstyper.
Knowledge graphs represent domain knowledge as entities and the relations between them, and they power many information systems. A common task is link prediction: inferring missing facts by predicting new connections. Although research is advancing, few models can explain why they make a prediction. We introduce ResqLP, a link prediction model that learns probabilistic first-order logic rules—human-readable if–then rules with uncertainty—from both entity semantics and sequences of relations. The model includes a feature extraction step that uses enclosing subgraphs to reduce the otherwise hard problem of enumerating all possible paths between entities. We also present a general way to capture entity semantics based on relations found in an entity’s neighborhood. Our results show that enclosing subgraphs speed up feature extraction, although very large knowledge graphs can still be problematic in some cases. ResqLP can learn non-trivial logic rules, though their quality varies, and it performs best on relatively sparsely connected knowledge graphs with many different relation types.
[This abstract was generated with the help of AI]
Documents