用于知识表示学习的对抗式负样本生成

知识表示学习目的是将知识图谱中符号化表示的关系与实体嵌入到低维连续向量空间。知识表示模型在训练过程中需要大量负样本，但多数知识图谱只以三元组的形式存储正样本。传统知识表示学习方法中通常使用负采样方法，这种方法生成的负样本很容易被模型判别，随着训练的进行对性能提升的贡献也会越来越小。为了解决这个问题，提出了对抗式负样本生成器（ANG）模型。生成器采用编码-解码架构，编码器读入头或尾实体被替换的正样本作为上下文信息，然后解码器利用编码器提供的编码信息为三元组填充被替换的实体，从而构建负样本。训练过程采用已有的知识表示学习模型与生成器进行对抗训练以优化知识表示向量。在链接预测和三元组分类任务上评估了该方法，实验结果表明该方法对已有知识表示学习模型在FB15K237、WN18和WN18RR数据集上的链接预测平均排名与三元组分类准确度都有提升。

Adversarial negative sample generation for knowledge representation learning

Knowledge graph embedding is to embed symbolic relations and entities of the knowledge graph into low dimensional continuous vector space. Despite the requirement of negative samples for training knowledge graph embedding models, only positive examples are stored in the form of triplets in most knowledge graphs. Moreover, negative samples generated by negative sampling of conventional knowledge graph embedding methods are easy to be discriminated by the model and contribute less and less as the training going on. To address this problem, an Adversarial Negative Generator (ANG) model was proposed. The generator applied the encoder-decoder pipeline, the encoder readed in positive triplets whose head or tail entities were replaced as context information, and then the decoder filled the replaced entity with the triplet using the encoding information provided by the encoder, so as to generate negative samples. Several existing knowledge graph embedding models were used to play an adversarial game with the proposed generator to optimize the knowledge representation vectors. By comparing with existing knowledge graph embedding models, it can be seen that the proposed method has better mean ranking of link prediction and more accurate triple classification result on FB15K237, WN18 and WN18RR datasets.