南通大学学报（自然科学版）

传统的命名实体消歧技术通常依靠丰富的上下文语境和外部实体知识库,而很多新兴实体缺乏知识库且包含实体的文本长度较短,这些局限性使得传统算法不能够充分利用上下文的语义信息。另外,由于受有效样本数量的限制,算法最终应用的场景十分有限。基于上述问题,提出一种基于深度学习的结合BERT(bidirectional encoder representation from transformers)模型和长短期记忆神经网络的实体消歧方法。该方法主要包含以下几个部分:1)设计了一种基于BERT模型的词向量,通过较少的数据样本仍然可以获取较多的信息;2)为了让长短期记忆神经网络保留较多的有用信息和验证短文本以适用该方法,对句子样本进行切分;3)结合微软公司提出的NNI(neural network intelligence)技术,高效地获取较优的神经网络超参数。通过与其他不同类型的词向量和神经网络技术进行比较,验证了使用文中基于深度学习的实体消歧技术在F-Measure值评测指标上效果更好。

The traditional named entity disambiguation technology usually relies on rich context and knowledge of external entities. However, many emerging entities lack knowledge bases and the text containing entities is short. These limitations make traditional algorithms unable to make full use of contextual semantic information. At the same time,due to the limitation of the number of effective samples, the final application scenarios of the algorithm are very limited. Based on the above defects, this paper proposes a deep learning-based entity disambiguation method combining bidirectional encoder representation from transformers(BERT) model and long short-term memory neural network.The main work are the following parts: 1) A word vector based on the BERT model is designed to obtain more information through fewer data samples. 2) In order to allow the long short-term memory neural networks to retain useful information and verify that the short text applies to the method of this article, this method segments the sentence samples. 3) This article uses the neural network intelligence(NNI) technology proposed by Microsoft, which makes it possible to quickly and efficiently obtain the optimal neural network hyperparameter. This study compares other different types of word vectors and neural network technology, confirming that the F-Measure value of the entity disambiguation technology based on deep learning used in this paper is higher.