Search

Purpose: In order to use a popular supervised learning algorithm such as BERT to extract the relationships of concepts (triple relationship extraction), it is necessary to label the relationship types manually. If some relation words are not been labeled in the training stag, they cannot be recognized probably in the test stage and the corresponding entities cannot been recognized accordingly. This paper proposes a new unsupervised algorithm to extract as many relation words as possible of two entities, especially those that are easily overlooked. Methods: The disease-cause relationship was taken as an example, and 10204 effective sentences of disease and corresponding causes were extracted by web crawler. According to the constraints of syntactic, semantic and lexical features, the relationship words were extracted with an unsupervised manner, and the automatic extracted results were summarized. Results: Some specific relation words that are ignored in manual labeling stage are found; the conjoining relation words often appeared together in the texts are recognized; some types and features of relation words are obtained. These types and features can be used to help the relation labeling in the supervised learning stage, and to help expanding the relevant knowledge graphs and improving the accuracy of information retrieval.

The knowledge graph of treatment-disease relationship can be a benefit not only to understand, inquire, and learn the relations between treatments and diseases from a macro level, but also to obtain the differences between treatments to the same disease through the comparison of different treatments; with the aid of commonalities of some treatments, a treatment to a disease that has not been discovered may be recognized; and with the aid of the commonalities of some diseases, a treatment to a disease that has not been discovered may be recognized too. At present, the therapy-disease knowledge graph has become the focus of medical informatics fields. International and domestic Internet giants have deployed some products in this area. The basic method of knowledge graph research is relation extraction in natural language processing. This paper uses semantic technology to discover several sentence patterns expressed in the therapy-disease sentence, and uses a method similar to BERT, that is, using the known part of the sentence to learn the unknown part of the sentence. An unsupervised learning algorithm is designed that extracts the relationship between therapies and diseases. By extracting the relationship between diseases and therapies entities from the abstracts of 300844 literature on diseases’ treatments, 203,238 entity relationships were identified, of which 180,675 entity relationships were valid. The results show that the sentence patterns expressing entity relationships in the literature on disease-treatment are relatively fixed. The entity relationship extraction is performed by dividing the sentence structure by subject, predicate and object. And the accuracy of extraction is 88.98%, and the recall rate is 60.05%.

Up to now, scholars have proposed some algorithms for learning synonyms of laboratory indicators. However, there is no report on learning multiple synonym sets at same time for outliers of laboratory indicators. Among the outliers of laboratory indicators of diseases, there are multiple synonym sets with various semantics such as {high, higher, increase}, {low, lower, decrease}, {very low, significantly low}. In this paper, a synonym set learning algorithm for different semantics of outliers of laboratory indicators is proposed. Firstly, using the sentence pattern laboratory indicators, outliers, and disease entities from medical texts are extracted; secondly, a set of synonyms for different semantics of laboratory indicators and outliers are learned; thirdly, the correspondences relation of laboratory indicators, outliers, and diseases are established. In the process of the first task, the algorithm is designed like a BERT model, that is, masking any one component and learning it with the other unmasked three parts, then updating the component sets that will include the learned components, and then with the new three unmasked components that includes the updated component sets to learn a new component that is included in the previous unmasked components. These steps run iteratively until every component sets do not varied. The experimental results show that the proposed algorithms is effective.

Entity relation extraction is one of the basic steps of knowledge Graph. It identifies the relations between entities. A BERT-Bidirectional gated recurrent units-Attention mechanism (BERT-BiGRU-Attention) model has been proposed, but it is based on the single Chinese character based masking. Due to the complexity of Chinese grammar structure and the semantic diversity, a BERT(wwm-ext) was proposed based on the whole Chinese word masking. In this paper we propose a BERT(wwm-ext)-BiGRU-Attention model. The experimental result shows that for the purpose of entity relation extraction the precision is 93.60%, recall rate is 91.90%, and F1 value 92.53%, which are higher than the BERT-BiGRU-Attention and its precision is 91.80%, recall rate is 90.16%, and F1 value is 90.97%. Since BERT(wwm-ext)-BIGRU-Attention gets higher precision, F1 value, and higher recall rate, it has better effects on the Chinese entity relation extraction tasks.

In medical fields rule based classifiers have an advantage over black box classifiers, because they are understandable and can be integrated into human's knowledge base to assist clinicians in decision-making. This paper proposes a new classification rule inducing algorithm. In comparison with standard rough sets theory it calculates value core without attribute reduction in advance and does not remove examples covered by the newly generated rule. An experiment on 28 medical data sets is executed in comparison with other 14 algorithms, and experimental results show that the proposed method achieves good classification performance.

Rough sets theory has taken an important role in data mining. This paper introduces a new rough set based classification rule generation algorithm. It has three features: the first is that the new algorithm can be used in inconsistent systems. The second is its ability to calculate the core value without attributes reduction before. The third is that every example gives a rule and the core values are added first in rule generation process. Experimental results indicate that the classification performanceismuch better than the standard rough set, its variants andJRIPPER, a little better thanCBA and KNN,andcompetive to C4.5in terms of 8 measures. The higher performance of the new algorithm may get benefit from its enough higher accuracy rules and having some properties like KNN.