This is a combined framework integrating embedding method and rule-based reasonging to implement better link prediction.
-
Before running the reasoning, you must prepare the data well. There are some simple constraints about the data to be used.
-
A raw data set should consist of six files, w.r.t entity2id, relation2id, train, test, rule and a valid set is not a must.
-
The rule file contains the rules that are essential in reasoning. We get these rules throw some open work, here we use amie. The format of a rule is like:
?f /film/director/film ?a ?b /influence/influence_node/peers./influence/peer_relationship/peers ?f => ?a /film/film/edited_by ?b 1
rule.txt is not the source file of an official data set, we shoule use amie algorithm to generate them with: java -jar amie_plus.jar test.txt > ans.txt. Before running amie_plus.jar, the data in test.txt should meet the form of tsv and each line a triplet of (h,r,t), where h and t are both entities and r is a relation.
-
-
Reasoning will be operate by means of embedding model to get the embedding of entities and relations. The work of thu can be used https://github.com/thunlp/Fast-TransX . Three examples have been existing in the folders of transE, transH and transR respectively. These examples are about FB15k data set.
-
After the first two steps, some other pre-parations follow. And these codes are contained in preprocess.py.
- pyhon preprocess.py 1 simplify_rules
- pyhon preprocess.py 2 construct_knowledge_base
- pyhon preprocess.py 3 rating rules
- pyhon preprocess.py 4 simplify test set
- pyhon preprocess.py 5 filter test set
-
compile reasoning with g++ -std=c++11 -O2 reasoning.cpp, and run with ./a.out X
- ./a.out E denotes the embedding model is transE
- ./a.out H
- ./a.out R
note that it will cost a long time with nearly three days, so you can use ./a.out X t for a small test.
LSB Version: core-4.1-amd64:core-4.1-noarch
Distributor ID: CentOS
Description: CentOS Linux release 7.7.1908 (Core)
Release: 7.7.1908
Codename: Core
cpu cores : 16
total used free shared buff/cache available
Mem: 125G 10G 106G 85M 9.0G 114G
| Method | MR | MRR | Hits@10 | Hits@5 | Hits@3 | Hits@1 |
|---|---|---|---|---|---|---|
| TransE | 70.30 | 45.77 | 74.27 | 64.44 | 55.79 | 29.98 |
| EM-RBR(E) | 68.36(up1.94) | 50.01(+4.24) | 76.23(+1.96) | 67.84(+3.40) | 60.62(+4.83) | 34.44(+4.46) |
| TransH | 72.56 | 45.81 | 74.01 | 64.09 | 55.53 | 30.37 |
| TransH(EM-RBR) | 70.72(up1.84) | 52.39(+6.58) | 76.52(+2.51) | 68.28(+4.19) | 61.13(+5.60) | 38.82(+8.45) |
| TransR | 55.98 | 47.88 | 77.04 | 68.01 | 59.47 | 31.10 |
| TransR(EM-RBR) | 55.47(up0.51) | 51.93(+4.05) | 78.35(+1.31) | 70.54(+2.53) | 63.42(+3.95) | 35.86(+4.76) |