Knowledge reasoning with multiple relational paths

Figures & data

Figure 1. Path example.

Figure 2. RNN overall structure.

Figure 3. LSTM unit internal structure.

Figure 4. Bi-LSTM total structure.

Figure 5. NSPN model process flowchart.

Table 1. Dataset details.

	Relations	Entities	Train triples	Valid triples	Test triples
FB15K	1345	14951	483142	50000	59071
FB40K	1336	39528	370648	67946	96678
Google + (NIPS)	853	25525	515348	81254	85179
CN3lFamily	42	3206	17422	2177	2178
Kinships	25	104	8549	1068	1069

Table 2. Experiment comparison of relation prediction tasks on FB15 K and FB40 K.

	FB15K				FB40K
	MRR		MAP(%)		MRR		MAP(%)
Methods	Raw	Filtered	Raw	Filtered	Raw	Filtered	Raw	Filtered
PTransE	0.714	0.825	74.83	85.58	0.781	0.864	85.39	90.35
PRA	0.458	0.536	40.52	46.63	0.509	0.617	50.18	55.64
RNN + PRA	0.586	0.689	51.89	60.75	0.655	0.733	59.88	67.92
NSPN	0.796	0.879	79.48	91.84	0.821	0.919	87.55	94.14

Table 3. Experimental results of linking prediction tasks on FB15 K and FB40 K.

	FB15K				FB40K
	MRR		MAP(%)		MRR		MAP(%)
Methods	Raw	Filtered	Raw	Filtered	Raw	Filtered	Raw	Filtered
PTransE	0.509	0.587	58.44	65.79	0.577	0.604	64.62	69.25
PRA	0.297	0.336	38.43	44.95	0.327	0.399	39.50	45.15
RNN + PRA	0.399	0.548	47.02	54.71	0.433	0.518	51.18	59.01
NSPN	0.545	0.618	66.23	74.49	0.583	0.652	68.48	76.57

Table 4. Experiment results on 4 datasets in the filtered case.

	FB15K		CN3lFamily		Kinships		Google + (NIPS)
Methods	MRR	MAP(%)	MRR	MAP(%)	MRR	MAP(%)	MRR	MAP(%)
PTransE	0.587	65.79	0.574	58.65	0.461	34.61	0.586	70.59
PRA	0.336	44.95	0.322	40.37	0.326	29.23	0.447	51.09
RNN + PRA	0.548	54.71	0.486	51.22	0.459	33.87	0.577	65.81
R-GCN	0.591	70.13	0.533	62.84	0.473	35.13	0.618	73.52
Transformer	0.532	68.69	0.592	61.59	0.571	35.22	0.605	74.61
TACT	0.589	73.32	0.540	64.12	0.585	38.53	0.609	74.43
NSPN	0.618	74.49	0.583	64.57	0.561	37.58	0.625	77.26

Table 5. Experimental results on relation prediction tasks with different model configurations on FB15 K.

	FB15K
	MRR		MAP(%)
Methods	Raw	Filtered	Raw	Filtered
NoAttentionNoMixture	0.556	0.625	60.26	69.72
NoMixture	0.688	0.803	72.77	85.33
NSPN	0.796	0.879	79.48	91.84

Table 6. Experimental results on link prediction tasks with different model configurations on FB15 K.

	FB15K
	MRR		MAP(%)
Methods	Raw	Filtered	Raw	Filtered
NoAttentionNoMixture	0.392	0.459	50.04	57.70
NoMixture	0.510	0.586	62.22	68.41
NSPN	0.545	0.618	66.23	74.49

Figure 6. Time Performance Comparison on FB15 K.

Data availability statement

The data used to support the findings of this work is included within this article.