| 英文摘要 | Knowledge graphs are a huge semantic web composed of nodes and edges, in which nodes represent concepts and entities in the real world, and edges represent topological links and semantic relations between nodes. In recent years, knowledge graphs have become the focus of researchers as a key technology of intelligent application. Knowledge graphs can provide a solid foundation for intelligent applications such as intelligent search, intelligent question answering systems, and recommendation systems. However, in the real world, knowledge is constantly changing, and people's description of the world is constantly updated and revised. Therefore, to better meet the needs of system application, we must constantly expand the knowledge graphs. Most of the early knowledge graphs were constructed and extended manually,
which was not only inefficient but also costly. Therefore, the automatic expansion method of knowledge graphs has high research and application value. This paper focuses on how to mine corresponding knowledge from the knowledge graphs itself, additional knowledge graphs, and text to expand the knowledge graphs. The main contributions and innovations of this paper are summarized as follows:
1. Proposing One-Shot Relation Learning Method via Neighborhood Aggregation and Path Encoding
Knowledge reasoning is a simple and easy-to-deploy automatic expansion method of knowledge graphs. The traditional data-driven knowledge reasoning models are difficult to deal with the reasoning problems of relations and entities covering only a small amount of knowledge in large knowledge graphs. To solve the problem of relation and entity reasoning with few samples, this paper proposes a relation and entity prediction
method based on neighborhood aggregation and path encoding. This method uses the relation average attention network to aggregate the neighborhood information of entities so that the aggregated entities contain both potential entity type features and neighborhood entity features. The path encoding module is used to aggregate the path information between entities to enhance the representation of relations and reduce the impact of invisible entities on the representation of relations. The training task construction method can make the model predict the relations and entities in the same framework. Experiments on three one-shot relation learning datasets show that the relation and entity
reasoning performance of the proposed method significantly outperforms the strong baseline system based on representation learning.
2. Proposing Dual Attention Network for Cross-lingual Entity Alignment
In the method of using entity alignment to realize the automatic expansion of
knowledge graphs, the neighborhood information at the same level between peer entities is often inconsistent, and the number of neighborhood entities at different levels is also quite different. These two differences bring difficulties to the representation learning of peer entities. This paper proposes an entity alignment method based on the
dual attention mechanism. This method uses the relation-aware graph attention network to iteratively aggregate multi-layer neighborhood information to solve the problem of inconsistent information at the same level between peer entities and then uses the hierarchical attention network to selectively aggregate low-level and high-level information
to solve the problem of information imbalance between different levels. Experiments on three cross-language entity alignment datasets show that the proposed method can effectively reduce the structural differences between neighborhoods of peer entities and significantly improve the performance of entity alignment.
3. Proposing Bidirectional Entity Linking Method Based Cloze Test
The existing studies show that the entity linking method based on sequential decisions can expand the knowledge graphs efficiently. However, the current entity linking models based on fixed sequence ignore the decision order, which leads to the model can not make reasonable use of the linked entity information. To solve this problem,
this paper first proposes a method of dynamically constructing mention sequences. This method uses the reinforcement learning algorithm to continuously interact with previously linked entities and dynamically select a target to be linked that can make reasonable use of previous information. This method can provide a reasonable basis for all entity linking models based on sequential decisions. In addition, the entity linking methods based on unidirectional decisions have the problem of insufficient utilization of global information and potential wrong links can not be corrected. Because of this, inspired by the behavior of human beings when completing the cloze test, this paper proposes an entity linking method with the function of checking and correction. This
method uses the checking module to check whether the currently linked entity is correct. If correct, the entity will participate in the next decision as evidence; If incorrect, the correction module will be used to make a new decision on the mention. At the same time, the strategy of repeating the above checking and correction steps for the secondary linking can effectively solve the problem of insufficient utilization of information. Experiments show that the proposed method can make full and reasonable use of global information and significantly improve the performance of entity linking. |
修改评论