Classifying Biomedical Text Abstracts based on Hierarchical Concept Structure

Transcription

1 Classifying Biomeical Text Abstracts base on Hierarchical Concept Structure Rozilawati Binti Dollah an Masai Aono Abstract Classifying biomeical literature is a ifficult an challenging tas, especially when a large number of biomeical articles shoul be organize into a hierarchical structure. In this paper, we present an approach for classifying a collection of biomeical text abstracts ownloae from Meline atabase with the help of ontology alignment. To accomplish our goal, we construct two types of hierarchies, the OHSUMED isease hierarchy an the Meline abstract isease hierarchies from the OHSUMED ataset an the Meline abstracts, respectively. Then, we enrich the OHSUMED isease hierarchy before aapting it to ontology alignment process for fining probable concepts or categories. Subsequently, we compute the cosine similarity between the vector in probable concepts (in the enriche OHSUMED isease hierarchy) an the vector in Meline abstract isease hierarchies. Finally, we assign category to the new Meline abstracts base on the similarity score. The results obtaine from the experiments show the performance of our propose approach for hierarchical classification is slightly better than the performance of the multi-class flat classification. Keywors Biomeical literature, hierarchical text classification, ontology alignment, text mining. I. INTRODUCTION EXT classification system on biomeical literature aims to T select relevant articles to a specific issue from large corpora [1]. However, classifying biomeical literature becomes one of the challenging tass ue to the fact that a large number of biomeical articles are ivie into quite a few subgroups in a hierarchy. Many researchers have attempte to fin more applicable ways for classifying biomeical literature in orer to help users fin relevant articles on the web. However, most approaches use in text classification tas have applie flat classifiers that ignore the hierarchical structure an treat each concept separately. Generally, text classification can be consiere as a flat classification technique, where the ocuments are classifie into a preefine set of flat categories an no relationship specifie between the categories. Singh an Naata [] state that the flat classification approach was suitable when a small number of categories were efine. However, ue to the increasing number of publishe biomeical articles on the web, the tas of fining the most relevant category for a ocument becomes much more ifficult. Consequently, flat classification turns out to be inefficient, while hierarchical classification is much preferre. Contrary to flat classification, hierarchical classification can be efine as a process of classifying ocuments into a hierarchical organization of classes or categories base on the relationship between terms or categories. Recently, the use of hierarchies for text classification has been wiely investigate an applie by many researchers. However, only little attention has been pai to the classification of biomeical literature. Therefore, in this paper, we propose a hierarchical classification metho where we employ the hierarchical concept structure for classifying biomeical text abstracts with the help of ontology alignment. Our propose metho is ifferent compare to the previous wors because we construct two types of hierarchies, which are the enriche OHSUMED isease hierarchy as our training ontology an the Meline abstract isease hierarchy as testing ontology. Next, we employ the ontology alignment process to match the concepts an relations between the enriche OHSUMED isease hierarchy an the Meline abstract isease hierarchy for exploring an searching the aligne pairs. We select the aligne pairs as a set of probable relevant categories for classification purpose. Then, we evaluate the more specific concepts by calculating the cosine similarity score between the new Meline abstract an each probable relevant category. Finally, we classify this abstract base on the similarity score. In orer to evaluate our approach, we conucte the experiments of the multi-class flat classification as our baseline. Then, we compare the results of our propose metho with the baseline. The experimental evaluation inicates that our propose approach performs slightly better than the performance of the baseline. We organize the paper as follows. In section II, we summarize the relate wor on text classification. The propose hierarchical classification metho is explaine in section III. Section IV contains experiments an results. Finally, we conclue this paper with a summary an suggestions for future wor in section V. R. B. Dollah is a PHD stuent in the Grauate School of Engineering, Department of Computer Science an Engineering, Toyohashi University of Technology, Aichi, Japan an is with the Faculty of Computer Science an Information Systems, Universiti Tenologi Malaysia, Johor, Malaysia ( rozeela@e.cs.tut.ac.p, rozilawati@utm.my. M. Aono is a professor in the Department of Computer Science an Engineering, Toyohashi University of Technology, Aichi, Japan ( aono@tut.ac.p). II. RELATED WORK Text classification is the process of using automate techniques to assign text samples into one or more set of preefine classes [3]. In flat classification, the classifier will assign a new ocuments to a category base on training 178

2 examples of preefine ocuments. Meanwhile, in hierarchical classification, a new ocument woul be assigne to a specific category base on the concepts an relationships within the hierarchy of preefine classes. A few hierarchical classification methos have been propose. For example, Puliala an Gauch [4] an Gauch, et al. [5] reporte that they classify ocuments uring inexing which can be retrieve by using a combination of eywor an conceptual match. Ruiz an Srinivasan [6] have propose a text categorization metho base on the Hierarchical Mixture of Expert (HME) moel using neural networs. Li et al. [7] have propose another approach of hierarchical ocument classification using linear iscriminant proection to generate topic hierarchies. Several statistical classification methos an machine learning techniques have been applie to text an web pages classification incluing techniques base on ecision tree, neural networ [6] an support vector machine (SVM) [8], [9], [10]. SVM has been prominently an wiely use for ifferent classification tas, in particular for ocument classification. For instance, Nenaic et al. [8] have use SVM for classifying the gene names from the molecular biology literature. Meanwhile, Wang an Gong [9] have use SVM to istinguish between any two sub-categories uner the same concept or category for web page hierarchical classification. They were use the voting score from all category-to-category classifier for assigning a web ocument to a sub-category. An they reporte that their metho can improve the performance of imbalance ata. In [10], Dumais an Chen reporte that they use the hierarchical structures for classifying web content. Accoringly, they were employe SVM to train secon-level category moels using ifferent contrast sets. Eventually, they classifie a web content base on the scores that were combine from the top-level an secon-level moel. Fig. 1 A metho for hierarchical classification of biomeical text abstracts Our approach is ifferent from the previous wors, where we explore the use of hierarchical concept structure with the help of ontology alignment for searching an ientifying the probable categories in orer to classify biomeical text abstracts. III. HIERARCHICAL CLASSIFICATION METHOD The large number of biomeical articles that publishe on web has mae the process of classification becomes challenging an ifficult. Lately, various classification methos are propose for classifying biomeical literature. However, in our research, we explore the use of hierarchical structure or ontology for classifying biomeical text abstracts. The features or concepts in ontology can be use to inex the biomeical text abstracts for improving the accuracy of classification performance an also the result of searching relevant ocuments. Fig. 1 illustrates our propose hierarchical classification metho that implemente in our research. A. Datasets For our experiments, we use two ifferent atasets, which are a subset of the OHSUMED ataset as training ocuments an Meline abstracts as test ocuments. The OHSUMED ataset [11] is a subset of clinical paper abstracts from the Meline atabase, from year 1987 through We select 400 recors (ocuments) from 43 ifferent categories from the subset of OHSUMED ataset for enriching the OHSUMED isease hierarchy. While, for classification purpose, we have selecte ranomly a subset of Meline abstracts from the Meline atabase. We retrieve this ataset with the query terms, such as arrhythmia, "heart bloc", etc. Then, we inexe this ataset belonging to 1 categories of isease as shown in Table 1. A total number of Meline abstracts are

3 TABLE 1 THE NUMBER AND CATEGORIES OF MEDLINE ABSTRACTS Category Category Name No. of Documents No. 1 Arrhythmia 10 Heart Bloc 5 3 Coronary Disease 10 4 Angina Pectoris 5 5 Heart Neoplasms 10 6 Heart Valve Diseases 10 7 Aortic Valve Stenosis 5 8 Myocarial Diseases 10 9 Myocaritis Pericaritis 5 11 Cancer 10 1 Human Disease 10 Total 100 B. Text Preprocessing an Feature Selection In our research, we perform text preprocessing for extracting the features (or noun phrases) from the OHSUMED ataset an the Meline abstracts, respectively, by performing part-of-speech (POS) tagging an phrase chuning. POS tagging is a tas of assigning POS categories to terms from a preefine set of categories. Meanwhile, phrase chuning is the process of recovering the noun phrases constructe by the part-of-speech tags. Then, we employe these noun phrases (as concepts) for constructing the Meline abstract isease hierarchy an enriching the OHSUMED isease hierarchy. In aition, we also perform feature selection in orer to reuce the original features to a small number of features by removing the rare an irrelevant features. In feature selection process, we employ the ocument frequency an the chi-square (χ ) techniques to istinguish between relevant an irrelevant features. Firstly, we use the ocument frequency as a feature reuction technique for eliminating rare features. Therefore, we compute the ocument frequency for each unique feature in both atasets. Next, we eliminate the features with the highest an lowest frequencies. The purpose of this process is to reuce the feature space into a small number of important features. Subsequently, a χ test is use to measure the inepenence between feature (t) an category (c) in orer to istinguish between relevant an irrelevant features. Then, we select the relevant features by assigning features to specific categories. We measure the relationship between features (t) an categories (c) using the following equation. ( ) = oi e, i, χ ; (1) i, ei, where o i, is the observe frequency for each cell, while e i, is the expecte frequency for each cell. For χ test, we use the x contingency table to compare the χ istribution with one egree of freeom. Then, we ran the features accoring to the χ score. For our experiments, we only select the features with the χ score greater than as the relevant features. We use all of selecte features for enriching the OHSUMED isease hierarchy an also for constructing the Meline abstract isease hierarchies. Table below shows the statistic of the relevant features that prouce in the feature selection process. TABLE II THE STATISTIC OF SELECTED FEATURES Number of Features Features (after extract noun phrases) 3,145 After remove DF< an DF>0 1,130 After remove Chi-square score < ,081 Each ocument is represente by the TF-IDF weighting. Therefore, we calculate each term or feature weight using the equation given below. n n + 1 w = i tfi log ; (),, fi f i + 1 where term frequency tf i, is the frequency of term i occurs in ocument an = 1,, m. Document frequency f i is the total number of ocuments that contain term i an n is the total number of ocuments. C. Enriche OHSUMED Disease Hierarchy Fig. An approach for constructing an enriching the OHSUMED isease hierarchy In our research, we construct the OHSUMED isease hierarchy by referring to the OHSUMED isease irectory. Subsequently, we enrich this hierarchy by assigning the features that extracte from the OHSUMED ataset to each noe of the hierarchy as shown in Fig.. Fig. 3 An example of the enriche OHSUMED isease hierarchy The important tas in enriching the OHSUMED isease hierarchy is to select meaningful an relevant features from the 180

4 OHSUMED ataset. Enrichment ha been one by mapping each concept or noe in the OHSUMED isease hierarchy with a set of relate features that extracte an selecte from the OHSUMED ataset. Fig. 3 escribes the example of the enriche OHSUMED isease hierarchy. D. Meline Abstract Disease Hierarchy The Meline abstract isease hierarchies were constructe using the selecte features that extracte from a collection of biomeical text abstracts that ownloae from the Meline atabase. The escription of text preprocessing an feature selection has been explaine in subsection B. In our propose approach, we create the Meline abstract isease hierarchy using Protégé. For this purpose, we refer to the Meical Subect Heaings (MeSH) for inexing our features. We map all selecte features to the concepts in the MeSH tree structure [1] for ientifying heaing an subheaing of hierarchical grouping. Finally, we construct the Meline abstract isease hierarchy by using the heaing an subheaing of hierarchical grouping that are suggeste by the MeSH tree structure. Fig. 4 epicts a part of the Meline abstract isease hierarchy. Tachycaria Arrhythmia Braycaria Cariovascular Disease Heart Disease... Bunle-Branch Bloc Sinoatrial Bloc Fig. 4 An example of a part of the Meline abstract isease hierarchy E. Ontology Alignment The purpose of ontology alignment in our research is to match the concepts an relations between the enriche OHSUMED isease hierarchy an the Meline abstract isease hierarchy. For that reason, we perform ontology alignment using the Anchor-Floo algorithm (AFA) [13]. During ontology alignment process, AFA woul explore an search for the similarity among the neighboring concepts in both hierarchies base on terminological alignment an structural alignment. Then, AFA woul narrow own the enriche OHSUMED isease hierarchy an the Meline abstract isease hierarchy for proucing the aligne pairs (aligne entities across hierarchies or ontologies). These aligne pairs are obtaine by measuring similarity values, which consier textual contents, structure an semantics (available in the hierarchies) between pairs of entities. We select all of the aligne pairs as our probable concepts for classification purpose. F. Ontology-base Hierarchical In our propose approach, we construct two types of hierarchies, which are the OHSUMED isease hierarchy (as... Heart Bloc... our training ontology) an the Meline abstract isease hierarchies (as testing ontology). Then, we perform ontology alignment in orer to match both hierarchies for proucing the aligne pairs. We consier the aligne pairs as a set of probable relevant categories for classifying biomeical text abstracts. Afterwars, we evaluate the more specific concepts base on the similarity between the new Meline abstract an probable relevant category. We compute the cosine similarity score between the vector of unnown new abstract in each Meline abstract isease hierarchy an the vector of each probable category in the enriche OHSUMED isease hierarchy for ientifying an preicting more specific category. The cosine similarity score woul be calculate using the following equation. r r n w i = i w 1, i, sim (, ) = r r = (3) n n w w where the vector of = (w 11, w 1,, w 1n ) an the vector of = (w 1, w,, w n ). We ran all the probable categories base on their cosine similarity score. Eventually, we classify the new Meline abstracts into the first ran of similarity score. IV. EXPERIMENTS AND RESULTS For our experiments, we use 400 recors (ocuments) from 43 ifferent categories of the subset of OHSUMED ataset for enriching our ontology learning. While, for classification purpose, we ranomly ownloae 100 biomeical text abstracts that relate to human iseases from Meline atabase. Then, we conucte the multi-class flat classification experiments using LIBSVM [14], which ignore hierarchical structure as our baseline for evaluating the performance of our propose metho for hierarchical classification of biomeical text abstracts. The results an the performance of the flat an hierarchical classification are shown in Table 3 an Fig. 5. Category No. TABLE III THE RESULTS OF THE FLAT AND HIERARCHICAL CLASSIFICATION Flat Flat with chi-square Hierarchical (% Accuracy) Hierarchical with chisquare Average Confusion matrix in Table 4 emonstrates the etails of the results of the hierarchical classification, where we use the features that are selecte from the feature selection process. i= 1 i, i= 1 i, 181

5 TABLE IV CONFUSION MATRIX FOR THE HIERARCHICAL CLASSIFICATION (WITH CHI-SQUARE) PREDICTED CATEGORY C1 C C3 C4 C5 C6 C7 C8 C9 C10 C11 C1 UNKNOWN Total C C % 17 of 100 correct C % 83 of 100 incorrect or unnown C C * C is Category C ACTUAL CLASS C C C C C C UNKNOWN Total Overall, our propose approach performs slightly better that the baseline. We foun the experimental results inicate that our propose approach performs slightly better than the flat classifier using SVM, whereby the performance of the flat an hierarchical classification show on average 13% an 17% accuracy, respectively. Base on the results that obtaine from the experiments, overall, we foun the classification accuracies in some of categories in the hierarchical classification experiments, such as in category 1 (arrhythmia), 7 (bunle-branch bloc), 8 (coronary thrombosis) an 10 (aortic coarctation) perform better than in the flat classification experiments. Fig. 5 The performance of classification accuracies In aition, the classification accuracies of the category 1 (arrhythmia), 5 (heart valve iseases), 8 (coronary thrombosis) an 10 (aortic coarctation) perform better in both flat an hierarchical classification experiments compare to other iseases categories. This might be cause by some of the Meline abstracts too short an woul extract very few relevant features. Consequently, we coul construct a small hierarchy for ontology alignment purpose, which may prouce a small number of aligne pairs (probable category). Besies that, the small number of ocuments that are represente in a particular class in the ataset may also affect the ecrease of the classification accuracy. Even though the percentage of performance improvement in the hierarchical classification is small compare to the performance of the flat classification, we believe that our propose approach suggests that by using hierarchical concept structure with the help of ontology alignment coul improve the classification performance. V. CONCLUSION In this paper, we propose a hierarchical classification metho that utilize the ontology alignment, namely Anchor-Floo algorithm to search the probable categories for given biomeical text abstracts. Then, we evaluate the performance of our approach by conucting the hierarchical classification experiments using the features that extracte from the OHSUMED ataset an Meline abstracts. As our baseline, we performe the multi-class flat classification experiments using SVM. Other than that, we also performe feature selection by employing the ocument frequency an 18

6 chi-square techniques for selecting relevant features. Then, we conucte some experiments of the flat an hierarchical classification using the features that are selecte from feature selection process. Generally, the experimental results inicate that the use of hierarchical concept structure with the help of Anchor-Floo algorithm can improve classification performance. Although our propose approach is still naïve in achieving the goo classification accuracies, we believe that we coul moify our propose approach to prouce more relevant concept an preict more specific category for classifying biomeical text abstracts. Our future target is to see an propose more accurate approaches for selecting relevant an meaningful features in orer to enrich an expan the OHSUMED isease hierarchy an Meline abstract isease hierarchies. By increasing the total number of ocuments that are represente in each class in the ataset may lea to performance increase of our propose approach for hierarchical text classification. ACKNOWLEDGMENT This wor was supporte in part by Global COE Program Frontiers of Intelligent Sensing from the Ministry of Eucation, Culture, Sports, Science an Technology, Japan. REFERENCES [1] F. M. Couto, B. Martins an M. J. Silva, Classifying biological articles using web sources, In Proceeings of the 004 ACM symposium on Applie Computing, 004, pp [] A. Singh an K. Naata, Hierarchical classification of web search results using personalize ontologies, In Proceeings of the 3r International Conference on Universal Access in Human-Computer Interaction, HCI International 005, 005. [3] A. M. Cohen, An effective general purpose approach for automate biomeical ocument classification, AMIA 006 Symposium Proceeing, 006, pp [4] A. K. Puliala an S. Gauch, Hierarchical text classification, 004, URL: [5] S. Gauch, A. Chanramouli an S. Ranganathan, Training a hierarchical classifier using inter-ocument relationships, Technical Report, ITTC-FY007-TR , August 006. [6] M. E. Ruiz an P. Srinivasan, Hierarchical neural networs for text categorization, Information Retrieval, 5, 00, pp [7] T. Li, S. Zhu an M. Ogihara, Hierarchical ocument classification using automatically generate hierarchy, Journal of Intelligent Information Systems, 9(), 007, pp [8] G. Nenaic, S. Rice, I. Spasic, S. Ananiaou an B. Stapley, Selecting text features for gene name classification: from ocuments to terms, In Proceeings of the ACL 003 worshop on Natural language processing in biomeicine, Vol. 13, 003, pp [9] Y. Wang an Z. Gong, Hierarchical classification of web pages using support vector machine, In Proceeings of 11 th International Conference on Asian Digital Libraries, ICADL 008, Bali, Inonesia. Proceeings, Lecture Notes in Computer Science 536, Springer, 008, pp [10] S. Dumais an H. Chen, Hierarchical classification of web content, In Proceeing of the SIGIR000, Athens, GR, 000, pp [11] OHSUMED ataset, URL: ohsume.html. [1] Meical Subect Heaing (MeSH) tree structures, URL: [13] M.H. Seiqui an M. Aono, An efficient an scalable algorithm for segmente alignment of ontologies of arbitrary size, Web Semantics: Science, Services an Agents on the Worl Wie Web, (7), 009, pp [14] C.-C. Chang an C.-J. Lin, LIBSVM: a library for support vector machines, 007, URL: 183