Extracting Relations from Traditional Chinese Medicine Literature via Heterogeneous Entity Networks

we study the problem of extracting meaningful relations from large-scale traditional Chinese medicine (TCM) literature. Instead of separately extracting each relation from one single sentence or document, we propose collectively and globally extracting all types of relations (e.g., herb-syndrome, herb-disease, formula-syndrome, formula-disease, and syndrome-disease relations) from the entire TCM corpus, from the perspective of network mining in the context of heterogeneous TCM entity networks.

Firstly, we construct heterogeneous entity networks from the TCM literature. Specifically, we take all types of TCM named entities occurring in the literature as nodes, and create an edge for each pair of heterogeneous entities co-occurring in the same document. All the edges are treated as candidate relations to be identified. The following figure gives a simple example of a heterogeneous TCM network.

Then, we propose a unified graphical model, so called heterogeneous factor graph model, to simultaneously infer the labels of all the candidate relations in the context of heterogeneous TCM networks, by employing the idea of collective inference. A series of sophisticated feature functions are described to build the HFGM.

To evaluate the performance of our proposed method, we collect a dataset consisting of more than 100,000 paper abstracts from a Chinese publication database and randomly annotate a sample of relations. We train and evaluate our HFGM on the partially labelled dataset in a semi-supervised way. The experimental results demonstrate that our proposed method performs very well on the extraction of multiple types of relations, including herb-syndrome, herb-disease, formula-syndrome, formula-disease and syndrome-disease relations.

We propose a novel heterogeneous factor graph model (HFGM) to incorporate all the information within a single model for better identifying the TCM relations.

Probabilistic graphical models use a graph-based representation as the foundation for compactly encoding a complex distribution over a high-dimensional space. In this graphical representation, a node corresponds to a random variable while an edge corresponds to the direct probabilistic dependence between two variables. The graph structure denes the factorization of a joint probabilistic distribution over all variables. A factor graph is a particular type of graphical model, which takes a bipartite graph to represent the factorization of a joint distribution and enables ecient computation of marginal distributions through the sum-product algorithm.

The figure below gives the graphical representation of a HFGM. The dotted ellipse at the bottom is the input heterogeneous TCM network, in which a node represents a TCM entity of a certain type. The dotted ellipse in the middle is the set of candidate relations, each one of which corresponds to an edge in the input network. The dotted ellipse on the top is the factor graph generated from the input network, in which the colored ovals are variables (labels) corresponding to the candidate relations and the squares are factors. The green ovals represent the known labels that are taken as supervised information, while the red ones represent the unknown labels to be predicted. The black squares represent the evidence factors between variables and their attributes, while the blue ones represent the compatibility factors of triadic closures.

We train and test our proposed HFGM on the dataset we collected and annotated. Five types of relations, respectively herb-syndrome (HS), herb-disease (HD), formula-syndrome (FS), formula-disease (FD) and syndrome-disease (SD), are extracted from the TCM literature.

A traditional classification approach is employed as the baseline, in which the co-occurring frequency, lexical context and semantic distance mentioned before are taken as the classification features. Since the number of features is very large, we use a SVM (www.csie.ntu.edu.tw/~cjlin/libsvm/) as the basic classifier.

In order to let the basic classifier utilize the compatibility information among the relations, we also perform an iterative SVM algorithm on our relation extraction task. The algorithm takes the results of the basic SVM classifier as the initiate values of all the labels, and then iteratively updates the labels by simultaneously using the observed features and the inferred neighboring labels of each relation.

The experimental results show that our HFGM approach greatly improves the efficiency of relation extraction from the TCM literature over the basic SVM classifier (respectively increases the precision by around 10-12%, the recall by around 12-15% and the F1-measure by around 11-14% for different types of relations) and the iterative SVM classifier (respectively increases the precision by around 6-8%, the recall by around 7-10% and the F1-measure by around 7-11% for different types of relations). We performed a t-test between the performances of the three approaches with regard to the extraction of different types of relations, and found that the difference between them is extremely statistically significant (P < 0.001 for all precision, recall and F1-measure).

To further observe the effectiveness of our approach more legibly, we plot the receiver operating characteristic (ROC) curves of the basic SVM and HFGM approaches (as shown in Figure 4), where the y-axis represents the rate of predicated positive labels in all the positive samples, and the x-axis represents the rate of predicted positive labels in all the negative samples. It is obvious that the HFGM significantly outperforms the basic SVM classifier on extracting all types of relations.

Terminology Dictionaries

Extracting the relations from text first requires recognizing instances of the named entities. However, we do not study the named entity recognition problem in this paper, since our main focus is relation extraction. In fact, some relatively completed TCM terminology dictionaries have been published in several online TCM databases, such as TCMonline and TCMID. TCMonline (http://www.cintcm.com/) is the earliest online TCM database system built by the Institute of Information on Traditional Chinese Medicine, China Academy of Chinese Medical Science since 1984; TCMID (http://www.megabionet.org/tcmid/) is a comprehensive database to provide information and bridge the gap between TCM and modern life sciences built by the Institute of Biomedical Sciences, East China Normal University in recent years. We can directly collect dictionaries from these databases to detect the instances of entities in literature.

We collected four TCM entity dictionaries simply by copying and merging the names of the four types of entities from both the TCMonline and TCMID databases: an herb dictionary which contains 8,082 herbs, a formula dictionary which contains 39,932 formulae, a syndrome dictionary which contains 2,209 syndromes, and a disease dictionary which contains 3,316 MeSH (http://www.nlm.nih.gov/mesh/) diseases.

TCM Literature

We collected a corpus from CNKI (http://www.cnki.net/), which is one of the largest online Chinese publication databases and free searching is available on its homepage. The collected corpus contains the abstracts of all the 106,150 papers published in the most popular 114 Chinese TCM journals in recent 5 years, which almost cover all the research aspects of TCM. We used the four terminology dictionaries mentioned above to detect the entities occurring in the corpus, and finally 3,024 herbs, 4,957 formulae, 1,126 syndromes and 1,650 diseases were detected. Then we generated the candidate relations between the co-occurring heterogeneous entity pairs, and 11,197 herb-syndrome candidates, 11,755 herb-disease candidates, 9,659 formula-syndrome candidates, 7,882 formula-disease candidates and 9,645 syndrome-disease candidates were generated. It is noted that, we did not extract the herb-formula relations, because all the formulae in the dictionary are classical TCM formulae and their relations with herbs have been well defined.

Data Annotation

For training and quantitatively evaluating our proposed model, we randomly labeled a small fraction (i.e., 10%) of each type of candidate relations. We asked three TCM experts (respectively denoted as Kang, Tang and Zhan according to their family names), who major in Chinese materia medica and clinical TCM, to annotate the data. For each candidate relation to be labelled, domain experts read its corresponding paper abstracts to identify if the two end-entities are really related or just co-occurring by chance. The statistics of our dataset are summarized in the table below.

We used the Kappa statistics to measure the inter-annotator agreements. As shown in the following table, the Cohen’s Kappa scores between any two annotators and the Fleiss’ Kappa scores among the three annotators are mostly above 0.7, indicating a substantial agreement of them. After that, we employed the majority rule to decide the final label of each candidate relation.

Data Description

This dataset consists of 11 files:

FileName	Content	Size
dic_herb	term dictionary of herbs	8,082
dic_formula	term dictionary of formulae	39,932
dic_syndrome	term dictionary of syndromes	2,209
dic_disease	term dictionary of diseases	3,316
dic_formula2herb	formula-syndrome relation dictionary	354,176
candicate-relation_herb2syndrome	herb-syndrome relation candidates	11,197
candicate-relation_herb2disease	herb-disease relation candidates	11,755
candicate-relation_formula2syndrome	formula-syndrome relation candidates	9,659
candicate-relation_formula2disease	formula-disease relation candidates	7,882
candicate-relation_syndrome2disease	syndrome-disease relation candidates	9,645
tcm_literature	TCM documents in Chinese	106,150

Download

TCMRelationExtraction.zip

Huaiyu Wan, Marie-Francine Moens, Walter Luyten, Xuezhong Zhou, Qiaozhu Mei, Lu Liu, Jie Tang. Extracting Relations from Traditional Chinese Medicine Literature via Heterogeneous Entity Networks. Journal of the American Medical Informatics Association (JAMIA). 2016, 23(2): 356-365.​