Background
Problem Formulation
知识图谱 G \mathcal{G} G是由一个三元组集合 { ( h , r , t ) } ⊆ E × R × E \{(h, r, t)\} \subseteq \mathcal{E} \times \mathcal{R} \times \mathcal{E} { (h,r,t)}⊆E×R×E表示,其中, E \mathcal{E} E 表示entity set, R \mathcal{R} R表示relation set。知识图谱的实现能够预测两个现有实体 ( h , ? , t ) (h, ?, t) (h,?,t)之间的不可见关系 r,或者给定头实体和查询关系 ( h , r , ? ) (h, r, ?) (h,r,?),预测尾部实体 t 。这一篇文章在只给定一个example triplet ( h 0 , r , t 0 ) \left(h_{0}, r, t_{0}\right) (h0,r,t0)的条件,使true tail entity t true t_{\text {true}} ttrue的序列度高于其他candidate entities t ∈ C h , r t \in C_{h, r} t∈Ch,r。
One-Shot Learning Settings
根据标准one-shot学习的设置,文章假设获取了training tasks。每个training tasks 都对应一个KG relation r ∈ R r \in \mathcal{R} r∈R, 并拥有training/testing triples: { D r t r a i n , D r t e s t } \left\{D_{r}^{t r a i n}, D_{r}^{t e s t}\right\} {
Drtrain,Drtest}。这一tasks set就是 the meta-training set, T meta − train \mathbb{T}_{\text {meta}-\text {train}} Tmeta−train。
根据one-shot的任务设定,在每个 D r t r a i n D_{r}^{t r a i n} Drtrain中,采用一个三元组 ( h 0 , r , t 0 ) \left(h_{0}, r,t_{0}\right) (h0,r,t0)。
The D r test = { ( h i , r , t i , C h i , r ) } D_{r}^{\text {test}}=\left\{\left(h_{i}, r, t_{i}, \mathcal{C}_{h_{i}, r}\right)\right\} Drtest={
(hi,r,ti,Chi,r)} consists of the testing triples of r with ground-truth tail entities ti for each query ( h i , r ) \left(h_{i}, r\right) (hi,r), and the corresponding tail entity candidates C h i , r = { t i j } \mathcal{C}_{h_{i}, r}=\left\{t_{i j}\right\} Chi,r={
tij} where each t i j t_{i j} tij is an entity in G \mathcal{G} G。
给定test query ( h i , r ) \left(h_{i}, r\right) (hi,r)和 D r t r a i n D_{r}^{t r a i n} Drtrain中the labeled triple,可对the candidate set C h i , r \mathcal{C}_{h_{i}, r} Chi,r进行排列测试the metric model。
The meta-training objective: min θ E T r [ ∑ ( h i , r , t i , C h i , r ) ∈ D r t e s t ℓ θ ( h i , r , t i ∣ C h i , r , D r t r a i n ) ∣ D r t e s t ∣ ] \min _{\theta} \mathbb{E}_{T_{r}}\left[\sum_{\left(h_{i}, r, t_{i}, \mathcal{C}_{h_{i}, r}\right) \in D_{r}^{t e s t}} \frac{\ell_{\theta}\left(h_{i}, r, t_{i} | \mathcal{C}_{h_{i}, r}, D_{r}^{t r a i n}\right)}{\left|D_{r}^{t e s t}\right|}\right] minθETr[∑