An Empirical Investigation on Fine-Grained Syndrome Segmentation in TCM by Learning a CRF from a Noisy Labeled Data

Transcription

1 An Empirial Investigation on Fine-Grained Syndrome Segmentation in TCM by Learning a CRF from a Noisy Labeled Data Yaqiang Wang, Dan Tang, and Hongping Shu College of Software Engineering, Chengdu University of Information Tehnology, Chengdu, China {yaqwang, tangdan, shp}@uit.edu.n Chen Su Sihuan Aademy of Chinese Mediine Sienes, Chengdu, China oin_0308@foxmail.om 感 ( 表 证 ) (exogenous (syndrome)), 风 ( 证 ) (wind (syndrome)), and 热(证) (hot (syndrome)). These finegrained syndromes are inferred based on many parts of the medial problem desription, aording to the diagnosis theory of TCM. To feth the fine-grained information, oarse-grained syndromes need to be segmented into their fine-grained forms [6]. Most of researhes on knowledge disovery related to TCM syndromes did not onsider above fine-grained segmentation problem. Consequently, in this paper, we fous our attention on automati fine-grained TCM syndrome segmentation, FGSS for short. The results will support other preision medial researhes of TCM [2], [7], suh as TCM syndrome differentiation and syndrome-disease relationship mining, syndrome-gene network analysis, et. FGSS in TCM an be naturally treated as a Chinese word segmentation task, and fine-grained segments of TCM syndromes an be naturally regarded as Chinese words. Then this segmentation task an be solved by utilizing supervised sequene labeling models, e.g. Hidden Markov Models [8], Maximum Entropy Markov Models [9], [10], Conditional Random Fields (CRFs) [11], [12]. However, supervised models suffer from a laborintensive problem to build a labeled training data. To address this defiieny, semi-supervised models [13]-[15] are proposed in general domain. Although these methods redue muh ost of manual labor, labeled data is still a neessity. It is still a hallenging task to manually build a domain-speifially used dataset due to a requirement of ross-domain knowledge. In this paper, we propose a distantly supervised method for training CRFs for FGSS in TCM. The method takes advantages of naming and translating onventions of TCM syndromes. It firstly builds a noisy labeled training data through applying a probabilisti alignment model, whih is used to exploit relations between Chinese haraters and English words in a Chinese-English parallel orpus of TCM syndromes. This alignment model an infer high onfident fine-grained segments in TCM syndromes by heuristially employing a forward Abstrat Syndrome is an important omponent in Traditional Chinese Mediine (TCM), and it is also a distintive onept in TCM ompared with Western Mediine (WM). Clearly understand the TCM syndrome help researhers digest TCM regularities and bridge TCM and WM. Syndromes are often used in oarse-grained forms, however fine-grained medial information buried in the oarse-grained TCM syndromes would not be onsidered. In this paper, we empirially investigate Fine-Grained Syndrome Segmentation (FGSS) in TCM by a distantly supervised method to build a noisy labeled data for training CRFs for FGSS in TCM. The feasibility and effetiveness of the method are demonstrated based on a series of elaborate experiments. The best F1-sore an reah To the best of our knowledge, our work is the first to fous on finegrained information extration in Chinese medial texts. Index Terms information extration, distant supervision, biomedial natural language proessing, traditional Chinese mediine I. INTRODUCTION Syndrome in Traditional Chinese Mediine (TCM) onludes the etiology, loation of the disease, pathology as well as relationship between right and evil pathogen. It is a distintive onept in TCM omparing to Western Mediine (WM). A learly understanding of TCM syndromes will help researhers digest TCM regularities and bridge the TCM and WM [1], [2]. Reently, many researhers work on knowledge disovery of syndrome differentiation in TCM [3]-[5]. Syndromes in TCM literature and linial reords are in oarse-grained forms. For example, syndrome 外感风 热 (exogenous wind hot) is a syndrome onluded when a TCM pratitioner observes a medial problem desription 发热 (fever), 恶寒 (aversion to old), 头痛 (headahe), 脉浮数 (rapid floating pulse), 苔薄白 (moss thin and white), 舌红 (red tongue), 口渴 (thirst). It is omposed of three fine-grained segments, inluding 外 Manusript reeived January 12, 2018; revised May 17, J. Adv. Inf. Tehnol. doi: /jait

2 diretion alignment and a reversed diretion hek. Then, CRFs an be adopted and trained on this noisy labeled data. Experimental results demonstrate the feasibility and the effetiveness of the proposed method. The highest F1sore an reah , whih is a ompetitive result omparing to the result (0.939) obtained by a CRF trained based on a manually segmented data. II. alignment probability is measured by a ombination of the forward diretion aligning probability and the reversed diretion heking probability, and it is defined as follows. aˆi arg max ai 1 p i ea i (3) i where λ is the parameter that is used to balane the ontributions of p(i ej) and p(ej i). If there is a tie in the inferring proess, a heuristi strategy would be used that âi will be set by the same value as previously adjaent alignment result âi-1, and if urrent position is the first plae of in the TCM syndrome, âi will be empirially set to 1. THE PROBABILISTIC ALIGNMENT METHOD In this setion, we propose a simple and effetive probabilisti alignment method to onfidently align Chinese haraters in TCM syndromes to English words in the orresponding translations. The Chinese harater(s) aligned to the same English word would form a finegrained segment in a TCM syndrome. The alignment results are onfident, beause they passed a bidiretional onfirmation proess, i.e. a forward diretion aligning and a reversed diretion heking, whih would be introdued in following setions, and the results of these two methods are heuristially ombined together in order to infer the fine-grained segments. III. CRFS FOR FGSS IN TCM In Setion II, we introdued a heuristi method, whih help us reognize optimal fine-grained segments of TCM syndromes with the help of natural separators between English words in the translations of the TCM syndromes. The resulted segments ould, onsequently, make up a noisy labeled data for training CRFs for FGSS in TCM. In general, there are three key points with applying CRFs to a segmentation task, inluding segment representation [17], feature definition [18], and the parameter settings (or its implementation). The detailed information of these key points in this paper is present bellow. A. Forward Diretion Aligning Method Forward diretion aligning method ompletes a proess that probabilistially aligning Chinese haraters i in a TCM syndrome with English words ej in its translation e. The alignment probability of ej given i an be diretly defined by a onditional probability with a Laplae smoothing [16] that is e j,i (1) p e j i i V where (ej,i) is the o-ourrene frequeny of ej and i, (i) is the ount of i, and V is the amount of unique Chinese haraters in the orpus. α >= 0 is a smoothing parameter. A higher p(ej i) indiates that ompared to other English words in e, ej has a higher strength aligning to i. A. The Segment Representation The task of FGSS in TCM an be naturally defined as a sequene labeling proess that labels eah Chinese harater with a label, whih is used to represent whih part of a fine-grained segment the Chinese harater belongs to. There are five representative types of labels that are ommonly used for indiating a speifi part of a segment, inluding B (beginning of a segment), I (inside of a segment), O (outside of segments), E (end of a segment), and S (the single Chinese harater type segment). Speifially, in this paper, segment representation would like to inlude B, I and S. Firstly, there is no redundant ontent in TCM syndromes. Therefore, the segment representation O would be exluded in this paper. Seondly, owing to the onise style of desriptions of TCM syndromes, lengths of fine-grained segments in TCM syndromes are less than 3, and in most ases they are less than 2. Consequently, identifying the end of fine-grained segments individually would make little sense to improve the labeling performane, and sometimes, it would redue the labeling performane due to the diffiulty of rare lass identifiation. Moreover, it has been verified that individually identifying single Chinese harater type segments is helpful to lift labeling performane [17], [18], espeially in our task many single Chinese harater type segments are ontained. B. Reversed Diretion Cheking Method Reversed diretion heking method is designed to inspet results of the forward diretion aligning method, whether ej aligning to i with high onfidene. It means that ompared to other Chinese haraters in, i should also have a higher strength aligned with ej. It an be defined as i,e j (2) p i e j e j Ve In the equation, (i,ej) has the same meaning as (ej,i) in equation (1), (ej) is the ount of ej, and Ve is the amount of unique English words in the orpus. It is also smoothed by the Laplae, and β >= 0 is the smoothing parameter. C. The Heuristi Inferee Method The heuristi inferene method is used to find the best fine-grained segmentation ŝ of with the highest onfidene. ŝ onsists of ŝk, and ŝk is omposed by Chinese harater(s), whih are aligned with the same ej in e with the highest onfident alignment probability. The 2018 J. Adv. Inf. Tehnol. p ea i B. Feature Definition The CRFs is a sophistiated disriminative model for sequene labeling tasks. Its good labeling performane is dependent on a well-tailored feature definition [19]. In this paper we investigated two types of features, 46

3 inluding n-gram features, speifially the unigram, bigram and trigram features, and n-gram features with a gap (i.e. the skip n-gram features). N-gram features: It is not a trivial task to segment short and brief texts into words. Thus n-grams are usually used as their alternative to be the features. In this paper, we investigate the unigram, bigram and trigram features. Their usefulness has been demonstrated in the named entity reognition task in TCM [19]. For onveniene, in this paper, we set an exat window size 3 to extrat the features, beause lengths of TCM syndromes are often short, usually shorter than 6. Jumping n-gram features: Intuitively, bigram and trigram features with a gap may be helpful to FGSS in TCM. For example, a trigram feature with a gap in it 肾 * 虚 ( kidney * defiieny ) ould be used to determine that the Chinese harater at the position of * would be a fine-grained segment, suh as the fine-grained segments 阴 (yin) and 阳 (yang) in 肾阴虚 (kidney yin defiieny) and 肾阳虚 (kidney yang defiieny), respetively. The templates of above introdued features are listed in Table I. TABLE I. Feature Type DEFINITIONS OF FEATURE TEMPLATES Template Unigram Feature (U) i-3; i-2; i-1; i; i+1; i+2; i+3 Bigram Feature (B) Trigram Feature (T) i-3 i-2; i-2 i-1; i-1 i; i i+1; i+1 i+2; i+2 i+3 i-3 i-2 i-1; i-2 i-1 i; i-1 i i+1; i i+1 i+2; i+1 i+2 i+3 Skip Bigram Feature (JB) i-3 i-1; i-2 i; i-1 i+1; i i+2; i+1 i+3 Skip Trigram Feature (JT) C. CRFs Implementation i-3 i-2 i; i-2 i-1 i+1; i-1 i i+2; i i+1 i+3; i-3 i-1 i; i-2 i i+1; i-1 i+1 i+2; i i+2 i+3; In our experiments, the CRF++ tool is adopted, whih provides an effiient implementation for CRFs by using the limited-memory quasi-newton algorithm for training the models [11], [20]. Our aim is to demonstrate the feasibility and effetiveness of the proposed method to build a noisy labeled data for training CRFs for FGSS in TCM, and the usefulness of the segment representations and features desribed previously. So the default settings of the CRF++ tool are diretly used for simpliity. IV. EXPERIMENTS In this setion, the proposed method for building a noisy labeled data for training a CRF for FGSS in TCM is verified. The results would present the performane of the proposed probabilisti alignment method and how noisy the built data is. Moreover, the performane of CRFs trained on the noisy labeled training data for FGSS in TCM under different settings is also validated. The results would illustrate the appliability of the noisy labeled training data and, at the same time, investigate the effetiveness of the announed segment representation and feature definition. A. Experimental Datasets In this paper, a standard Chinese-English parallel orpus of TCM syndromes is used as the basis data for building the noisy labeled training data for CRFs for FGSS in TCM. It ontains 606 Chinese-English translation pairs of TCM syndromes. In order to evaluate the proposed probabilisti alignment method, the Chinese haraters and English words in these translation pairs are manually aligned in advane. This dataset is named as AD (i.e., the alignment data). TABLE II. DEFINITIONS OF FEATURE TEMPLATES AD SD Size Number of Unique Chinese Charathers Number of Unique Fine- Grained Segments Number of Common Unique Fine-Grained Segments 80 To illustrate the appliability of the noisy labeled training data and investigate the effetiveness of the aforementioned segment representation and feature definition, 260 linial TCM syndromes are extrated from a linial reord dataset. The linial reord dataset ontains 6722 linial reords, whih are aumulated during the routine diagnosti work of TCM pratitioners. None of these extrated linial TCM syndromes exists in the parallel orpus. Moreover, TCM experts are hired to pre-segment these linial TCM syndromes into their fine-grained forms manually. This dataset is named as SD (i.e., segmentation data). More detailed information of the experimental datasets is listed in Table II. B. Evaluation Metris We define three types of metris for evaluating the proposed methods as follows. Alignment acuray (AC): this metri is used for evaluating the alignment performane of the proposed probabilisti alignment method. In this paper, an alignment is orret, if and only if the aligned Chinese harater is or is a part of a segment that the English word in the translation of the TCM syndrome orresponds to. Higher AC is ahieved; better alignment performane would be obtained. The AC is defined as follows. a AC (4) a where a is the ount of the orretly aligned Chinese harater and English word pairs, and a is the amount of aligned Chinese harater and English word pairs. Segmentation Preision (SP), Reall (SR) and F1- sore (SF): these metris are used to validate the appliability of the resulted noisy labeled data for training CRFs for FGSS in TCM by the proposed probabilisti alignment method and the appropriateness of the segment representation and the defined features. If and only if the boundaries of a fine-grained segment are both reognized 2018 J. Adv. Inf. Tehnol. 47

4 aurately, the fine-grained segment is onsidered to be a orret result. Higher SP and higher SR are ahieved; better global segmentation performane would be gotten. SP, SR and SF are formulated below. SP = s (5) SR (6) s s results. The results obtained by the HIM are listed in Table IV. It shows that a better result an be obtained through balaning the ontributions of p(i ej) and p(ej i) by hanging the value of λ, and the best result an reah 0.932, when λ = 0.1. Moreover, the results illustrate that RDC is important to the alignment task in this paper, and these results also onveys that the heuristially inferred segment data ontains noisy. sr SF 2 SP SR SP SR TABLE III. RESULTS OBTAINED BY FDA AND RDC METHODS RESPECTIVELY UNDER DIFFERENT AND SETTINGS (7) values where s is the ount of segments segmented orretly, s is the ount of segments segmented, and sr is the ount of gold segments in the test data. Labeling Preision (LP), Reall (LR) and F1-sore (LF): these metris are utilized to assess the labeling performane of CRFs based FGSS in TCM in order to give a detailed analysis for investigating the appropriateness of the segment representations and the defined features. Higher LP and LR are ahieved; better global labeling performane LF would be gotten. LP, LR and LF are defined by equations (8), (9) and (10), respetively. LPt LRt LFt l,t LPt LRt values (9) (10) where the subsript t indiates whih lass label it is, e.g. B, I, E, or S. l,t is the ount of Chinese haraters orretly labeled with the lass label t, l,t is the ount of Chinese haraters labeled with the lass label t, and lr,t is the ount of gold lass label t in the test data. AC of HIM D. Evaluation of the Alignment Results The FGSS results obtained by CRFs trained on the noisy labeled data under different segment representation and feature definition settings are shown in Fig. 1. The best SF an reah , when BIS segment representation and UB features are used. It shows that the noisy labeled data an be used for training CRFs for FGSS in TCM. C. Evaluation of the Alignment Results In order to investigate the sensitivity of the proposed probabilisti alignment method to the smoothing parameter, we evaluate the forward diretion aligning (FDA) results and the reversed diretion heking (RDC) results separately under different smoothing parameter settings. The results are listed in Table III. It shows learly that FDA prefers a larger smoothing parameter value than RDC, and the best parameter values are 1.0 and 0.1, respetively. Furthermore, the results demonstrate that the smoothing would be helpful to improve the auray of the pro-posed probabilisti alignment method. To validate the performane of the proposed probabilisti alignment method, we verify the alignment auray of the heuristi inferene method (HIM) desribed in Setion II under different λ settings when the smoothing parameters α and β are set to 1.0 and 0.1, respetively, whih are the parameter settings let FDA and RDC methods ahieve the best alignment auray 2018 J. Adv. Inf. Tehnol. AC of RDC SET TO 1.0 AND 0.1 (8) lr,t 2 LPt LRt values TABLE IV. COPYRIGHT FORMS AND RE RESULTS OBTAINED BY HIM UNDER DIFFERENT SETTINGS RESPECTIVELY WHEN AND ARE l,t l,t AC of FDA Figure 1. Results of SP, SR and SF obtained by CRFs trained on the noisy labeled training data under different segment representation and feature definition settings. 48

5 In Fig. 1, we an also see that BIS is the most suitable segment representation to CRFs based FGSS in TCM. Under the same feature definition setting, BIS an ahieve better results than other type of segment representations in most ases, exept under the U (worse than BIES ) and UBT (equal to BIES ) feature definitions. Figure 2. Results of SP, SR and SF obtained by CRFs trained on the labeled training data generated by the gold aligned data under different segment representation and feature definition settings. It is also shown in Fig. 1 that, to the CRFs based FGSS in TCM, it is enough to just using the unigram and bigram features in the disriminative proess. Under the same segment representation, when ombining unigram and bigram features, the results are always better than other onditions. In order to further validate the usability of the noisy labeled training data to learn a CRF for FGSS in TCM, we run the same experiment again based on the labeled training data generated by the gold aligned data, and the results are shown in Fig. 2. We an see that, under the same settings, the results obtained by the CRFs trained on the noisy labeled data an ahieve ompetitive SFs with the results obtained by the CRFs trained on the labeled data generated by the gold aligned data. More importantly, our method an learn a CRF for FGSS in TCM without additional manual labor for building a training data. E. Detailed Analysis of the FGSS Results Due to the spae limitation, we just show the detailed labeling results, whih are obtained by the CRFs trained on the noisy labeled data under the BIS segment representation and different feature definition settings, in Fig. 3. The onlusions on these results obtained under other types of the segment representations are similar to the results shown by these figures. Figure 3. Results of LP, LR and LF of B, I and S obtained by CRFs trained on the noisy labeled training data under the ondition of BIS as the segment representation and different feature definition used. It learly shows in Fig. 3 that the results of LP, LR and LF of S are all better than the performane ahieved by B and I. It may be due to the ause that long finegrained segments are greatly less than the single Chinese harater type of fine-grained segments. Thus the distribution of labels is skewed, and it will result in the diffiulty of reognizing B and I. Furthermore, in Fig. 3 we an also see that LP results of B and I are always higher than the LR results under the same ondition. In ontrast, the results of S have an opposite onlusion. It reminds that we should pay more attention on improving LR results for reognizing B and I and, at the same time, enhaning LR results for S in future work. V. EXPERIMENTS In this paper, we empirially investigated the problem of fine-grained segmentation of TCM syndromes. A distantly supervised method is proposed to learn CRFs for FGSS in TCM. A series of experiments have done, and the feasibility and effetiveness of the proposed method were preliminarily proved in this paper. In future work, in order to further verify its pratiability, the method would be applied on a larger sale real world data in different situations. Moreover, if we do not have an available standard parallel orpus, ould we use largesale translation software generated noisy data as the input and how to deal with it? Many other new researh problems would produe when fine-grained information extrated, and how to generate linial syndromes of TCM based on the fine-grained syndrome segments in TCM ould be explored in future. ACKNOWLEDGMENT Authors are pleased to aknowledge the National Natural Siene Foundation of China (Grant No J. Adv. Inf. Tehnol. 49

6 ), the Sientifi Researh Foundation of Siene and Tehnology Department of Sihuan Provine (Grant No. 2016JY0240), the Young Talent Researh Start-up Foundation of Chengdu University and Information Tehnology (Grant No. KYTZ201638) and the Young Aademi Leader Researh Foundation of Chengdu University and Information Tehnology (Grant No. J201705); Corresponding author: Yaqiang Wang, uit.edu.n. REFERENCES [1] W. Wang, H. Zhao, J. Chen, J. Chen, and G. Xi, Bridge the gap between syndrome in traditional Chinese mediine and proteome in western mediine by unsupervised pattern disovery algorithm, in Pro. IEEE International Conferene on Networking, Sensing and Control, 2008, pp [2] P. Gu and H. Chen, Modern bioinformatis meets traditional Chinese mediine, Briefings in Bioinformatis, vol. 15, no. 6, pp , [3] X. Zhou, Y. Peng, and B. Liu, Text mining for traditional Chinese medial knowledge disovery: A survey, Journal of Biomedial Informatis, vol. 43, no. 4, pp , [4] Y. Wang, Z. Yu, Y. Jiang, Y. Liu, L. Chen, and Y. Liu, A framework and its empirial study of automati diagnosis of traditional Chinese mediine utilizing raw free-text linial reords, Journal of Biomedial Informatis, vol. 45, no. 2, pp , [5] H. Wang, X. Liu, B. Lv, F. Yang, and Y. Hong, Reliable multilabel learning via onformal preditor and random forest for syndrome differentiation of hroni fatigue in traditional Chinese mediine, PloS ONE, vol. 9, no. 6, p. e99565, [6] M. Jiang, C. Lua, C. Zhang, J. Yang, Y. Tan, A. Lu, and K. Chan, Syndrome differentiation in modern researh of traditional Chinese mediine, Journal of Ethnopharmaology, vol. 140, no. 3, pp , [7] Y. Feng, Z. Wu, X. Zhou, Z. Zhou, and W. Fan, Knowledge disovery in traditional Chinese mediine: State of the art and perspetives, Artifiial Intelligene in Mediine, vol. 38, no. 3, pp , [8] W. J. Teahan, Y. Wen, R. MNab, and I. H. Witten, A ompression-based algorithm for Chinese word segmentation, Computational Linguistis, vol. 26, no. 3, pp , [9] A. MCallum, D. Freitag, and F. Pereira, Maximum entropy markov models for information extration and segmentation, in Pro. 17th International Conferene on Mahine Learning, [10] N. Xue and L. Shen, Chinese word segmentation as LMR tagging, in Pro. Seond SIGHAN Workshop on Chinese Language Proessing, 2003, pp [11] J. Lafferty, A. MCallum, and F. Pereira, Conditional random fields: Probabilisti models for segmenting and labeling sequene data, in Pro. International Conferene on Mahine Learning, 2001, pp [12] H. Zhao, C. N. Huang, and M. Li, An improved Chinese word segmentation system with onditional random field, in Pro. Fifth SIGHAN Workshop on Chinese Language Proessing, 2006, pp [13] F. Jiao, S. Wang, C. H. Lee, R. Greiner, and D. Shuurmans, Semisupervised onditional random fields for improved sequene segmentation and labeling, in Pro. 21st International Conferene on Computational Linguistis and 44th Annual Meeting of the Assoiation for Computational Linguistis, 2006, pp [14] W. Sun and J. Xu, Enhaning Chinese word segmentation using unlabeled data, in Pro. Conferene on Empirial Methods in Natural Language Proessing, 2011, pp [15] X. Zeng, D. F. Wong, L. S. Chao, and I. Tranoso, Graph-based semi-supervised model for joint hinese word segmentation and part-of-speeh tagging, in Pro. 51st Annual Meeting of the Assoiation for Computational Linguistis, 2013, pp [16] Q. Yuan, G. Cong, and N. M. Thalmann, Enhaning naive Bayes with various smoothing methods for short text lassifiation, in Pro. 21st International Conferene on World Wide Web, 2012, pp [17] H. C. Cho, N. Okazaki, M. Miwa, and J. Tsujii, Named entity reognition with multiple segment representations, Information Proessing and Management, vol. 49, no. 4, pp , [18] H. Zhao, C. N. Huang, and B. L. Lu, A unified harater-based tagging framework for Chinese word segmentation, ACM Transations on Asian Language Information Proessing, vol. 9, no. 2, pp. 1 32, [19] Y. Wang, Y. Liu, Z. Yu, L. Chen, and Y. Jiang, A preliminary work on symptom name reognition from free-text linial reords of traditional Chinese mediine using onditional random fields and reasonable features, in Pro. Workshop on Biomedial Natural Language Proessing, 2012, pp [20] F. Sha and F. Pereira, Shallow parsing with onditional random fields, in Pro. Conferene of the North Amerian Chapter of the Assoiation for Computational Linguistis on Human Language Tehnology, 2003, pp Yaqiang Wang reeived his PhD degree in 2013 from College of Computer Siene, Sihuan University. His MS and BS degree were obtained from Sihuan University in 2010 and CUIT in 2007, respetively. Yaqiang is now a Leture of Chengdu University of Information Tehnology. He eleted as a Master advisor of Chengdu University of Information Tehnology in He is a member of ACL and CCF. He reeived Outstanding Reviewer award twie from Journal of Biomedial Informatis in 2014 and 2017, respetively. Currently, his researh interests are natural language proessing, mahine learning, and biomedial informatis. Dan Tang is a Professor of Chengdu University of Information Tehnology. Currently, his researh interests are data mining and information seurity. Hongping Shu is a Professor of Chengdu University of Information Tehnology. Currently, his researh interests are big data mining and software engineering. Chen Su is an Assoiate Professor of Sihuan Aademy of Chinese Mediine Sienes. Currently, her researh interests are big data mining and biomedial informatis J. Adv. Inf. Tehnol. 50