{ "paper_id": "O08-6003", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:02:34.519871Z" }, "title": "Hierarchical Taxonomy Integration Using Semantic Feature Expansion on Category-Specific Terms", "authors": [ { "first": "Cheng-Zen", "middle": [], "last": "Yang", "suffix": "", "affiliation": { "laboratory": "", "institution": "Yuan Ze University", "location": { "addrLine": "135 Yuan-Tung Rd", "postCode": "320", "settlement": "Chungli", "country": "Taiwan" } }, "email": "czyang@syslab.cse.yzu.edu.tw" }, { "first": "Ing-Xiang", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "Yuan Ze University", "location": { "addrLine": "135 Yuan-Tung Rd", "postCode": "320", "settlement": "Chungli", "country": "Taiwan" } }, "email": "" }, { "first": "Cheng-Tse", "middle": [], "last": "Hung", "suffix": "", "affiliation": { "laboratory": "", "institution": "Yuan Ze University", "location": { "addrLine": "135 Yuan-Tung Rd", "postCode": "320", "settlement": "Chungli", "country": "Taiwan" } }, "email": "" }, { "first": "Ping-Jung", "middle": [], "last": "Wu", "suffix": "", "affiliation": { "laboratory": "", "institution": "Yuan Ze University", "location": { "addrLine": "135 Yuan-Tung Rd", "postCode": "320", "settlement": "Chungli", "country": "Taiwan" } }, "email": "pjwu@syslab.cse.yzu.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In recent years, the hierarchical taxonomy integration problem has obtained considerable attention in many research studies. Many types of implicit information embedded in the source taxonomy are explored to improve the integration performance. The semantic information embedded in the source taxonomy, however, has not been discussed in previous research. In this paper, an enhanced integration approach called SFE (Semantic Feature Expansion) is proposed to exploit the semantic information of the category-specific terms. From our experiments on two hierarchical Web taxonomies, the results show that the integration performance can be further improved with the SFE scheme.", "pdf_parse": { "paper_id": "O08-6003", "_pdf_hash": "", "abstract": [ { "text": "In recent years, the hierarchical taxonomy integration problem has obtained considerable attention in many research studies. Many types of implicit information embedded in the source taxonomy are explored to improve the integration performance. The semantic information embedded in the source taxonomy, however, has not been discussed in previous research. In this paper, an enhanced integration approach called SFE (Semantic Feature Expansion) is proposed to exploit the semantic information of the category-specific terms. From our experiments on two hierarchical Web taxonomies, the results show that the integration performance can be further improved with the SFE scheme.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "In many daily information processing tasks, merging two classified information sources to create a larger taxonomy with abundant information is in great demand. For example, an e-commerce service provider may merge various catalogs from other vendors into its local catalog to provide customers with versatile contents. A Web user may also want to integrate different blog catalogs from Web 2.0 portals to organize a personal information management library. In these examples, people may need an efficient automatic integration approach to process the huge amount of information.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In recent years, the taxonomy integration problem has obtained much attention in many research studies (e.g. Agrawal & Srikan, 2001; Sarawagi, Chakrabarti, & Godbole, 2003; Zhang & Lee, 2004a; Zhang & Lee, 2004b; Zhu, Yang, & Lam, 2004; Chen, Ho, & Yang, 2005; Wu, Tsai, & Hsu, 2005; Ho, Chen, & Yang, 2006; Chen, Ho, & Yang, 2007; Cheng & Wei, 2008; Wu, Tsai, Lee, & Hsu, 2008) . As pointed out in these studies, the integration work is more subtle than traditional classification work because the integration accuracy can be further improved with different kinds of implicit information embedded in the source or destination taxonomy. A taxonomy, or catalog, usually contains a set of objects divided into several categories according to some classified characteristics. In the taxonomy integration problem, the objects in a taxonomy, the source taxonomy S, are integrated into another taxonomy, the destination taxonomy D. As shown in earlier research, this problem is more than a traditional document classification problem because different kinds of implicit information in the source taxonomy are explored to greatly help integrate source documents into the destination taxonomy. For example, a Naive Bayes classification approach (Agrawal & Srikan, 2001 ) with the classification relationship information implicitly existing in the source catalog can achieve integration accuracy improvement. Several SVM (Support Vector Machines) approaches (Chen, Ho, & Yang, 2005) can also have similar improvement with other implicit source information.", "cite_spans": [ { "start": 109, "end": 132, "text": "Agrawal & Srikan, 2001;", "ref_id": "BIBREF0" }, { "start": 133, "end": 172, "text": "Sarawagi, Chakrabarti, & Godbole, 2003;", "ref_id": "BIBREF12" }, { "start": 173, "end": 192, "text": "Zhang & Lee, 2004a;", "ref_id": "BIBREF19" }, { "start": 193, "end": 212, "text": "Zhang & Lee, 2004b;", "ref_id": "BIBREF20" }, { "start": 213, "end": 236, "text": "Zhu, Yang, & Lam, 2004;", "ref_id": "BIBREF21" }, { "start": 237, "end": 260, "text": "Chen, Ho, & Yang, 2005;", "ref_id": "BIBREF3" }, { "start": 261, "end": 283, "text": "Wu, Tsai, & Hsu, 2005;", "ref_id": "BIBREF14" }, { "start": 284, "end": 307, "text": "Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 308, "end": 331, "text": "Chen, Ho, & Yang, 2007;", "ref_id": "BIBREF4" }, { "start": 332, "end": 350, "text": "Cheng & Wei, 2008;", "ref_id": "BIBREF5" }, { "start": 351, "end": 378, "text": "Wu, Tsai, Lee, & Hsu, 2008)", "ref_id": "BIBREF15" }, { "start": 1237, "end": 1260, "text": "(Agrawal & Srikan, 2001", "ref_id": "BIBREF0" }, { "start": 1449, "end": 1473, "text": "(Chen, Ho, & Yang, 2005)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "The implicit source information studied in previous enhanced approaches generally includes the following features: (1) co-occurrence relationships of source objects (Agrawal & Srikan, 2001; Zhu, Yang, & Lam, 2004; Chen, Ho, & Yang, 2005) , (2) latent source-destination mappings (Sarawagi, Chakrabarti, & Godbole, 2003; Zhang & Lee, 2004b; Cheng & Wei, 2008) , (3) inter-category centroid information (Zhang & Lee, 2004a) , and (4) parent-children relationships in the source hierarchy (Wu, Tsai, & Hsu, 2005; Ho, Chen, & Yang, 2006; Wu, Tsai, Lee, & Hsu, 2008) . In our survey, however, the semantic information embedded in the source taxonomy has not been discussed. Since different applications have shown that the semantic information can benefit the task performance (Krikos, Stamou, Kokosis, Ntoulas, & Christodoulakis, 2005; Hsu, Tsai, & Chen, 2006) , such information should be able to achieve similar improvements for taxonomy integration. In addition, we further study the hierarchical taxonomy integration problem because many taxonomies, such as Web catalogs, existing in the real world are hierarchical.", "cite_spans": [ { "start": 165, "end": 189, "text": "(Agrawal & Srikan, 2001;", "ref_id": "BIBREF0" }, { "start": 190, "end": 213, "text": "Zhu, Yang, & Lam, 2004;", "ref_id": "BIBREF21" }, { "start": 214, "end": 237, "text": "Chen, Ho, & Yang, 2005)", "ref_id": "BIBREF3" }, { "start": 279, "end": 319, "text": "(Sarawagi, Chakrabarti, & Godbole, 2003;", "ref_id": "BIBREF12" }, { "start": 320, "end": 339, "text": "Zhang & Lee, 2004b;", "ref_id": "BIBREF20" }, { "start": 340, "end": 358, "text": "Cheng & Wei, 2008)", "ref_id": "BIBREF5" }, { "start": 401, "end": 421, "text": "(Zhang & Lee, 2004a)", "ref_id": "BIBREF19" }, { "start": 486, "end": 509, "text": "(Wu, Tsai, & Hsu, 2005;", "ref_id": "BIBREF14" }, { "start": 510, "end": 533, "text": "Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 534, "end": 561, "text": "Wu, Tsai, Lee, & Hsu, 2008)", "ref_id": "BIBREF15" }, { "start": 772, "end": 831, "text": "(Krikos, Stamou, Kokosis, Ntoulas, & Christodoulakis, 2005;", "ref_id": "BIBREF10" }, { "start": 832, "end": 856, "text": "Hsu, Tsai, & Chen, 2006)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In this paper, we propose an enhanced integration approach by exploiting the implicit semantic information in the source taxonomy with a semantic feature expansion (SFE) mechanism. The basic idea behind SFE is that some semantically related terms can be found to represent a source category, and these representative terms can be further viewed as the additional common category labels for all documents in the category. Augmented with these additional semantic category labels, the source documents should be more precisely integrated into the correct destination category. The semantic expanding scheme, however, needs to consider the polysemy situation to avoid introducing many topic-irrelevant features. Therefore,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Semantic Feature Expansion on Category-Specific Terms SFE employs an efficient correlation coefficient method to select representative semantically-related terms.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 423", "sec_num": null }, { "text": "To study the effectiveness of SFE, we implemented it based on a hierarchical taxonomy integration approach (EHCI) proposed in Ho et al. (2006) and Chen et al. (2007) with the Maximum Entropy (ME) model classifiers. We have conducted experiments with real-world Web catalogs from Yahoo! and Google, and measured the integration performance with precision, recall, and F 1 measures. The results show that the SFE mechanism consistently can improve the integration performance of the EHCI approach.", "cite_spans": [ { "start": 126, "end": 142, "text": "Ho et al. (2006)", "ref_id": "BIBREF8" }, { "start": 147, "end": 165, "text": "Chen et al. (2007)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 423", "sec_num": null }, { "text": "The rest of the paper is organized as follows. Section 2 describes the problem definition and Section 3 reviews previous related research. Section 4 elaborates the proposed semantic feature expansion approach and the hierarchical integration process. Section 5 presents the experimental results, and discusses the factors that influence the experiments. Section 6 concludes the paper and discusses some future directions of our work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 423", "sec_num": null }, { "text": "Following the definitions in Ho et al. (2006) , we assume that two homogeneous hierarchical taxonomies, the source taxonomy S and the destination taxonomy D, participate in the integration process. The taxonomies are said to be homogeneous if the topics of the two taxonomies are similar. In addition, the taxonomies under consideration are required to overlap with a significant number of common documents. For example, in our experimental data sets, 20.6% of the total documents (436/2117) in the Autos directory of Yahoo! also appear in the corresponding Google directory.", "cite_spans": [ { "start": 29, "end": 45, "text": "Ho et al. (2006)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "2." }, { "text": "The source taxonomy S has a set of m categories, or directories, S 1 , S 2 , \u2026, S m . These categories may have subcategories, such as S 1,1 and S 2,1 . Similarly, the destination catalog D has a set of n categories. The integration process is to directly decide the destination category", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "2." }, { "text": "S 1 S 1,1 S \u2026 S m D 1 D 1,1 D \u2026 D n D 2 D 2,2 D 2,1 Source Taxonomy Destination Taxonomy S 1,2 1 x d S 1 S 1,1 S \u2026 S m D 1 D 1,1 D \u2026 D n D 2 D 2,2 D 2,1 Source Taxonomy Destination Taxonomy S 1,2 1 x d", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "2." }, { "text": "in D for each document d x in S. In this study, we allow that d x can be integrated into multiple destination categories because a document commonly appears in several different directories in a real-world taxonomy. Figure 1 depicts a typical scenario of the integration process on two hierarchical taxonomies. For illustration, we assume that the source category S 1,1 has a significant number of overlapped documents with the destination categories D 1,1 and D 2,2 . This means that the documents appearing in S 1,1 should have similar descriptive information as the documents in D 1,1 and D 2,2 . Therefore, a non-overlapped document 1", "cite_spans": [], "ref_spans": [ { "start": 216, "end": 224, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Figure 1. A typical integration scenario for two hierarchical taxonomies.", "sec_num": null }, { "text": "x d in category S 1,1 should be intensively integrated into both two destination categories D 1,1 and D 2,2 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 1. A typical integration scenario for two hierarchical taxonomies.", "sec_num": null }, { "text": "In previous studies, different sorts of implicit information embedded in the source taxonomy are explored to help the integration process. These implicit source features can be mainly categorized into four types: (1) co-occurrence relationships of source objects, (2) latent source-destination mappings, (3) inter-category centroid information, and (4) parent-children relationships in the source hierarchy. The co-occurrence relationships of source objects are first studied to enhance a Naive Bayes classifier based on the concept that if two documents are in the same source category, they are more likely to be in the same destination category (Agrawal & Srikan, 2001 ). The enhanced Na\u00efve Bayes classifier (ENB) is shown to have more than 14% accuracy improvement on average. The work in Chen et al. (2005) also has the similar concept in its iterative pseudo relevance feedback approach. As reported in Chen et al. (2005) , the enhanced SVM classifiers consistently achieve improvement.", "cite_spans": [ { "start": 648, "end": 671, "text": "(Agrawal & Srikan, 2001", "ref_id": "BIBREF0" }, { "start": 793, "end": 811, "text": "Chen et al. (2005)", "ref_id": "BIBREF3" }, { "start": 909, "end": 927, "text": "Chen et al. (2005)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Integration Techniques", "sec_num": "3.1" }, { "text": "Latent source-destination mappings are explored in Sarawagi et al. (2003) and Zhang and Lee (2004b) . The cross-training (CT) approach (Sarawagi, Chakrabarti, & Godbole, 2003) extracts the mappings from the first semi-supervised classification phase using the source documents as the training sets. Then, the destination documents are augmented with the latent mappings for the second semi-supervised classification phase to complete the integration. The co-bootstrapping (CB) approach (Zhang & Lee, 2004b ) exploits the predicted source-destination mappings to repeatedly refine the classifiers. The experimental results show that both CT and CB outperform ENB (Sarawagi, Chakrabarti, & Godbole, 2003; Zhang & Lee, 2004b) .", "cite_spans": [ { "start": 51, "end": 73, "text": "Sarawagi et al. (2003)", "ref_id": "BIBREF12" }, { "start": 78, "end": 99, "text": "Zhang and Lee (2004b)", "ref_id": "BIBREF20" }, { "start": 135, "end": 175, "text": "(Sarawagi, Chakrabarti, & Godbole, 2003)", "ref_id": "BIBREF12" }, { "start": 486, "end": 505, "text": "(Zhang & Lee, 2004b", "ref_id": "BIBREF20" }, { "start": 662, "end": 702, "text": "(Sarawagi, Chakrabarti, & Godbole, 2003;", "ref_id": "BIBREF12" }, { "start": 703, "end": 722, "text": "Zhang & Lee, 2004b)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Integration Techniques", "sec_num": "3.1" }, { "text": "In Zhang and Lee (2004a) , a cluster shrinkage (CS) approach, in which the feature weights of all objects in a document category are shrunk toward the category centroid, is proposed. Therefore, the cluster-binding relationships among all documents of a category are strengthened. In Wu et al. (2005) and Ho et al. (2006) , the parent-children information embedded in hierarchical taxonomies is intentionally extracted. Based on the hierarchical characteristics, Wu et al. extend the CS and CB approach to improve the integration performance. In Ho et al. (2006) , an enhanced approach called EHCI is proposed to further extract the hierarchical relationships as a conceptual thesaurus. Their results show that the implicit hierarchical information can be effectively used to boost the accuracy performance.", "cite_spans": [ { "start": 3, "end": 24, "text": "Zhang and Lee (2004a)", "ref_id": "BIBREF19" }, { "start": 283, "end": 299, "text": "Wu et al. (2005)", "ref_id": "BIBREF14" }, { "start": 304, "end": 320, "text": "Ho et al. (2006)", "ref_id": "BIBREF8" }, { "start": 545, "end": 561, "text": "Ho et al. (2006)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Integration Techniques", "sec_num": "3.1" }, { "text": "The semantic information embedded in the source taxonomy has not been discussed in past studies. This observation motivates us to study the embedded taxonomical semantic information and its effectiveness.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration Techniques", "sec_num": "3.1" }, { "text": "In our proposed SFE scheme, we use the Maximum Entropy (ME) model classifiers to perform the main integration task. Here, we provide a brief overview of the ME model as the background of our work. More details can be found in Berger et al. (1996) . In ME, the entropy ( ) H p for a conditional distribution ( | ) p y x is used to measure the uniformity of ( | ) p y x , where y is an instance of all outcomes Y in a random process and x denotes a contextual environment of the contextual space X, or the history space. To express the relationship between x and y, we can have an indicator function ( , ) f x y (usually known as feature function) defined as:", "cite_spans": [ { "start": 226, "end": 246, "text": "Berger et al. (1996)", "ref_id": "BIBREF1" }, { "start": 598, "end": 603, "text": "( , )", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 if ( ) has the defined relationship ( , ) 0 else x, y f x y \u23a7 = \u23a8 \u23a9", "eq_num": "(1)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "The entropy ( ) H p is defined by:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( ) ( | )log ( | ) x X H p p y x p y x \u2208 = \u2212 \u2211", "eq_num": "(2)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "The Maximum Entropy Principle is to find a probability model * p C \u2208 such that:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "* arg max ( ) p C p H p \u2208 = (3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "where C is a set of allowed conditional probabilities. There are, however, two constraints: ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "{ } { } p p E f E f = (4) and ( | ) 1 y Y p y x \u2208 = \u2211", "eq_num": "(5)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "{ } ( , ) ( , ) p x y E f p x y f x y \u2261 \u2211 (6) , { } ( ) ( | ) ( , ) p x y E f p x p y x f x y \u2261 \u2211", "eq_num": "(7)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "As indicated in [10] , the conditional probability ( | ) p y x can be computed by:", "cite_spans": [ { "start": 16, "end": 20, "text": "[10]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 ( | ) exp ( , ) ( ) i i i p y x f x y z x \u03bb \u239b \u239e = \u239c \u239f \u239d \u23a0 \u2211", "eq_num": "(8)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "where i \u03bb is the Lagrange multiplier for feature i f , and ( ) z x is defined as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( ) exp ( , ) i i y i z x f x y \u03bb \u239b \u239e = \u239c \u239f \u239d \u23a0 \u2211 \u2211", "eq_num": "(9)" } ], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "With the improved iterative scaling (IIS) algorithm (Darroch & Ratcliff, 1972; Berger, Pietra, & Pietra, 1996) , the i \u03bb values can be estimated. Then, the classifiers are built according to the ME model and the training data.", "cite_spans": [ { "start": 52, "end": 78, "text": "(Darroch & Ratcliff, 1972;", "ref_id": "BIBREF6" }, { "start": 79, "end": 110, "text": "Berger, Pietra, & Pietra, 1996)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Overview of the Maximum Entropy Model Classifiers", "sec_num": "3.2" }, { "text": "Previous integration research for hierarchal taxonomy integration mainly can be classified into two categories: clustering-based (Cheng & Wei, 2008) and classification-based (Ho, Chen, & Yang, 2006; Zhu, Yang, & Lam, 2004; Chen, Ho, & Yang, 2007) . The clustering-based approach has the advantage in handling manifold taxonomies which may even have small overlaps and in performing integration without a priori training work. Therefore, the application of the clustering-based approach is much more general. The effectiveness of the clustering-based approach, however, depends on the clustering parameters. For inexperienced users, finding optimal clustering parameters will be very challenging.", "cite_spans": [ { "start": 129, "end": 148, "text": "(Cheng & Wei, 2008)", "ref_id": "BIBREF5" }, { "start": 174, "end": 198, "text": "(Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 199, "end": 222, "text": "Zhu, Yang, & Lam, 2004;", "ref_id": "BIBREF21" }, { "start": 223, "end": 246, "text": "Chen, Ho, & Yang, 2007)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration", "sec_num": "3.3" }, { "text": "Although the classification-based approach is more appropriate for handling taxonomies which have significant overlaps, it cannot handle the subtle relationships embedded in categories. For example, CatRelate uses five types of hierarchical relationships in a taxonomy to help catalog integration (Zhu, Yang, & Lam, 2004) , and an integration scheme called EHCI uses a hierarchical weighting mechanism to strengthen the integration effectiveness (Ho, Chen, & Yang, 2006; Chen, Ho, & Yang, 2007) . Nonetheless, CatRelate only discusses the hierarchical relationships on a category basis with a set of simple rules. It may suffer from complicated hierarchical relationships when handling large taxonomies. In contrast, EHCI's hierarchical weighting mechanism considers the influences of category labels of more comprehensive neighboring levels on a document basis. The experimental results reported in Ho et al. (2006) and Chen et al. (2007) also show that EHCI is effective for handling large taxonomies. Therefore, we use EHCI as our baseline to study the effectiveness of the proposed SFE approach. The following gives a brief overview for EHCI.", "cite_spans": [ { "start": 297, "end": 321, "text": "(Zhu, Yang, & Lam, 2004)", "ref_id": "BIBREF21" }, { "start": 446, "end": 470, "text": "(Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 471, "end": 494, "text": "Chen, Ho, & Yang, 2007)", "ref_id": "BIBREF4" }, { "start": 900, "end": 916, "text": "Ho et al. (2006)", "ref_id": "BIBREF8" }, { "start": 921, "end": 939, "text": "Chen et al. (2007)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration", "sec_num": "3.3" }, { "text": "In EHCI, the conceptual relationships (category labels) are first extracted from the hierarchical taxonomy structure as a thesaurus (Ho, Chen, & Yang, 2006; Chen, Ho, & Yang, 2007) . Then, the features of each document are extended with the thesaurus by adding the weighted label features. A weighting formula is designed to control the impact of the semantic concepts of each hierarchical level. Equation 10 calculates the EHCI feature weight", "cite_spans": [ { "start": 132, "end": 156, "text": "(Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 157, "end": 180, "text": "Chen, Ho, & Yang, 2007)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 427 Semantic Feature Expansion on Category-Specific Terms", "sec_num": null }, { "text": ", e x d f of each term x in document d, where L i is the relevant label weight assigned as 1 2 i with an i-level depth, , x d", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 427 Semantic Feature Expansion on Category-Specific Terms", "sec_num": null }, { "text": "f is the original weight, and \u03bb is used to control the magnitude relation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 427 Semantic Feature Expansion on Category-Specific Terms", "sec_num": null }, { "text": ", x d f is assigned by x i TF TF \u2211", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": ", where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": "x TF is the term frequency of x, and i denotes the number of the stemmed terms in each document. The label weight L i of each thesaurus is exponentially decreased and accumulated based on the increased levels. , , 0", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": "(1 ) Table 1 shows the label weights of different levels, where L 0 is the document level, L 1 is one level upper, and so on to L n for n levels upper. The label weighting scheme uses a power-law distribution to avoid over-emphasis on the least related hierarchical levels. To build the enhanced classifiers for destination categories, the same enhancement on hierarchical label information is also applied to the destination taxonomy to strengthen the discriminative power of the classifiers.", "cite_spans": [], "ref_spans": [ { "start": 5, "end": 12, "text": "Table 1", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "e x x d x d n i i L f f L \u03bb \u03bb = = \u00d7 + \u2212 \u00d7 \u2211", "eq_num": "(10)" } ], "section": "The weight", "sec_num": null }, { "text": "Although the EHCI approach employs only the embedded hierarchical information with a simple power-law distribution, the integration accuracy performance can be effectively improved. As reported in Chen et al. (2007) , the EHCI approach outperforms a straightforward classification scheme that does not employ any embedded information to help hierarchical taxonomy integration. ", "cite_spans": [ { "start": 197, "end": 215, "text": "Chen et al. (2007)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": "Hierarchical Level Label Weight Document Level (L 0 ) 1/2 0 One Level Upper (L 1 ) 1/2 1 Two Levels Upper (L 2 ) 1/2 2 \u2026 \u2026 n Levels Upper (L n ) 1/2 n", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The weight", "sec_num": null }, { "text": "The proposed semantic feature expansion (SFE) approach is to use extracted representative terms of a category as the implicit semantic information to help the corresponding integration process. In the following, the overall processing flow of SFE is presented first. Related approaches incorporated in the integration process are then described. Finally, the SFE approach is elaborated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration with Semantic Feature Expansion", "sec_num": "4." }, { "text": "To apply SFE to hierarchical taxonomies, a hierarchical taxonomy integration approach (EHCI) (Ho, Chen, & Yang, 2006; Chen, Ho, & Yang, 2007) is considered as the baseline. Currently, classifiers based on the Maximum Entropy (ME) model are used because of its prominent performance in many tasks, such as natural language processing (Berger, Pietra, & Pietra, 1996) and flattened taxonomy integration (Wu, Tsai, & Hsu, 2005) . Figure 2 shows the entire integration process flow of the SFE approach. ", "cite_spans": [ { "start": 93, "end": 117, "text": "(Ho, Chen, & Yang, 2006;", "ref_id": "BIBREF8" }, { "start": 118, "end": 141, "text": "Chen, Ho, & Yang, 2007)", "ref_id": "BIBREF4" }, { "start": 333, "end": 365, "text": "(Berger, Pietra, & Pietra, 1996)", "ref_id": "BIBREF1" }, { "start": 401, "end": 424, "text": "(Wu, Tsai, & Hsu, 2005)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 427, "end": 435, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Integration Process", "sec_num": "4.1" }, { "text": "To further improve the integration performance, the semantic information of inter-taxonomy documents is explored in the proposed approach to perform semantic feature expansion (SFE). The main idea is to augment the feature space of each document with representative topic words. As noted in Tseng et al. (2006) , the hypernyms of documents can be considered as the candidates of the representative topic words for the documents. Hereby, SFE adopts a similar approach to Tseng et al. (2006) to first select important term features from the documents and then decide the representative topic terms from hypernyms.", "cite_spans": [ { "start": 291, "end": 310, "text": "Tseng et al. (2006)", "ref_id": "BIBREF13" }, { "start": 470, "end": 489, "text": "Tseng et al. (2006)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "Since feature expansion with hypernyms intends to introduce features that are not related to the document topic, these irrelevant features need to be filtered out before the final integration work. From the aspect of improving integration accuracy, the expanded features that have little discriminative power among categories are considered to be removed. According to previous studies (Ng, Goh, & Low, 1997; Yang & Pedersen, 1997; Tseng, Lin, Chen, & Lin, 2006) , although the 2 \u03c7 -test (chi-square) method is very effective in feature selection for text classification, it cannot differentiate negatively related terms from positively related ones. For a term t and a category c, their 2 \u03c7 measure is defined as:", "cite_spans": [ { "start": 386, "end": 408, "text": "(Ng, Goh, & Low, 1997;", "ref_id": "BIBREF11" }, { "start": 409, "end": 431, "text": "Yang & Pedersen, 1997;", "ref_id": "BIBREF18" }, { "start": 432, "end": 462, "text": "Tseng, Lin, Chen, & Lin, 2006)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "( )( )( )( ) 2 2 ( ) ( , ) T T F F T F F T T F F T N N N N N t c N N N N N N N N \u03c7 + \u2212 + \u2212 + \u2212 + \u2212 + + \u2212 \u2212 \u00d7 \u00d7 \u2212 \u00d7 = + + + + (11)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "where N is the total number of the documents, T N + ( F N + ) is the number of the documents of category c (other categories) containing the term t, and T N \u2212 ( F N \u2212 ) is the number of the documents of category c (other categories) not containing the term t.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "Therefore, the correlation coefficient (CC) method is suggested to filter out the negatively related terms (Ng, Goh, & Low, 1997; Tseng, Lin, Chen, & Lin, 2006) . Since N is the same for each term, we can omit it and get the following equation to calculate the CC value for each term:", "cite_spans": [ { "start": 107, "end": 129, "text": "(Ng, Goh, & Low, 1997;", "ref_id": "BIBREF11" }, { "start": 130, "end": 160, "text": "Tseng, Lin, Chen, & Lin, 2006)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( )( )( )( ) ( ) ( , ) T T F F T F F T T F F T N N N N CC t c N N N N N N N N + \u2212 + \u2212 + \u2212 + \u2212 + + \u2212 \u2212 \u00d7 \u2212 \u00d7 = + + + +", "eq_num": "(12)" } ], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "Since the categories in a taxonomy are in a hierarchical relationship, SFE only considers the categories of the same parent in the CC method.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "Then, the five terms with the highest CC values are selected to perform semantic feature expansion. As indicated by (Ng, Goh, & Low, 1997; Tseng, Lin, Chen, & Lin, 2006) , the terms selected with CC are highly representative for a category. The category-specific terms of a source category, however, may not be topic-genetic to the corresponding destination category. Therefore, SFE uses them as the basis to find more topic-indicative terms for each category.", "cite_spans": [ { "start": 116, "end": 138, "text": "(Ng, Goh, & Low, 1997;", "ref_id": "BIBREF11" }, { "start": 139, "end": 169, "text": "Tseng, Lin, Chen, & Lin, 2006)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "Some lexical dictionaries, such as InfoMap (http://infomap.stanford.edu/) and WordNet (http://wordnet.princeton.edu/), can be used to extract the hypernyms of the category-specific terms to get the topic indicative features of a category. For example, if a category has the following five category-specific terms: output, signal, circuit, input, and frequency, SFE gets the following hypernyms from InfoMap: signal, signaling, sign, communication, abstraction, relation, etc . These hypernyms are more topic-generic than the category specific terms. Then, SFE calculates the weight ", "cite_spans": [ { "start": 399, "end": 474, "text": "InfoMap: signal, signaling, sign, communication, abstraction, relation, etc", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "x HF is the term frequency of x, and i denotes the number of the hypernyms in each category.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "For each document k d , its SFE feature vector k sf is changed by extending Equation 10 as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "[ ] (1 ) (1 ) k k k k \u03bb \u03bb \u03b1 \u03b1 = \u00d7 + \u2212 \u00d7 \u00d7 + \u2212 \u00d7 sf l h f", "eq_num": "(14)" } ], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "where k l denotes the feature vector of the hierarchical thesaurus information computed from the left term of Equation 10, k h denotes the feature vector of the topic-generic terms of the category computed from Equation 13, and k f denotes the original feature vector of the document derived from the right term of Equation 10.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Feature Expansion", "sec_num": "4.2" }, { "text": "We have conducted experiments with real-world catalogs from Yahoo! and Google to study the performance of the SFE scheme with a Maximum Entropy classification tool from Edinburgh University (ver. 20041229) (Zhang, 2004) . Two integration procedures were implemented. The baseline is ME with EHCI (EHCI-ME), and the other is ME with EHCI and SFE (SFE-ME). We measured three scores with different \u03bb and \u03b1 values: precision, recall, and F 1 measures. Both integration directions were evaluated: from Google to Yahoo! and from Yahoo! to Google. The experimental results show that SFE-ME can effectively improve the integration performance. For recall measures, SFE-ME outperforms EHCI-ME in more than 60% of all cases. For precision measures, SFE-ME outperforms EHCI-ME in more than 90% of all cases. SFE-ME can also achieve the best recall and precision performance. For F 1 measures, SFE-ME outperforms EHCI-ME in nearly 95% of all the cases. The experimental results are detailed in the following.", "cite_spans": [ { "start": 206, "end": 219, "text": "(Zhang, 2004)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Analysis", "sec_num": "5." }, { "text": "Semantic Feature Expansion on Category-Specific Terms", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Hierarchical Taxonomy Integration Using 431", "sec_num": null }, { "text": "In the experiments, five directories from Yahoo! and Google were extracted to form two experimental taxonomies (Y and G). Table 2 shows these directories and the number of the extracted documents after ignoring the documents that could not be retrieved. As in previous studies (Agrawal & Srikan, 2001; Sarawagi, Chakrabarti, & Godbole, 2003; Ho, Chen, & Yang, 2006) , the documents appearing in only one category were used as the training data (|Y-G| and |G-Y|), and the common documents were used as the testing data (|Y Test| and |G Test|). Since some documents may appear in more than one category in a taxonomy, |Y Test| is slightly different from |G Test|. For simplicity consideration, the level of each hierarchy was controlled to be at most three in the experiments. If the number of the documents of a certain subcategory was less than 10, the subcategory would be merged upward to its parent category. Before the integration, we used the stopword list in Frakes and Baeza-Yates (1992) to remove the stopwords, and the Porter algorithm (Porter, 1980) for stemming. In the integration process, we allow that each source document x d can be integrated into multiple destination categories (one-to-many) as we find in real-world taxonomies. Different \u03bb values from 0.1 to 1.0 were applied to the source taxonomy ( s \u03bb ) and the destination taxonomy ( d \u03bb ). To both taxonomies, the same \u03b1 value ranging from 0.1 to 1.0 was applied for semantic feature expansion. The lexical dictionary used in the experiments was InfoMap to get hypernyms. As reported in Tseng et al. (2006) , we believe that WordNet will result in similar hypernym performance.", "cite_spans": [ { "start": 277, "end": 301, "text": "(Agrawal & Srikan, 2001;", "ref_id": "BIBREF0" }, { "start": 302, "end": 341, "text": "Sarawagi, Chakrabarti, & Godbole, 2003;", "ref_id": "BIBREF12" }, { "start": 342, "end": 365, "text": "Ho, Chen, & Yang, 2006)", "ref_id": "BIBREF8" }, { "start": 965, "end": 994, "text": "Frakes and Baeza-Yates (1992)", "ref_id": "BIBREF7" }, { "start": 1045, "end": 1059, "text": "(Porter, 1980)", "ref_id": null }, { "start": 1561, "end": 1580, "text": "Tseng et al. (2006)", "ref_id": "BIBREF13" } ], "ref_spans": [ { "start": 122, "end": 129, "text": "Table 2", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Data Sets", "sec_num": "5.1" }, { "text": "In the experiments, we measured the integration performance of EHCI-ME and SFE-ME in six scores: macro-averaged recall (MaR), micro-averaged recall (MiR), macro-averaged precision (MaP), micro-averaged precision (MiP), macro-averaged F 1 measure (MaF), and micro-averaged F 1 measure (MiF). The standard F 1 measure is defined as the harmonic mean of recall and precision: 1 2 F rp r p = + , where recall is computed as correctly integrated documents all test documents r = and precision is computed as correctly integrated documents all predicted positive documents p =", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Sets", "sec_num": "5.1" }, { "text": ". The micro-averaged scores were measured by computing the scores globally over all categories in five directories. The macro-averaged scores were measured by first computing the scores for each individual category, then averaging these scores. The recall measures are used to reflect the traditional performance measurements on integration accuracy. The precision measures show the degrees of false integration. The standard F 1 measures show the compromised scores between recall and precision.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Sets", "sec_num": "5.1" }, { "text": "Although we have measured the integration performance with different \u03bb values, this paper only lists part of the results in five different d \u03bb values, which are 0.1, 0.3, 0.5, 0.7, and 0.9.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Results and Discussion", "sec_num": "5.2" }, { "text": "Considering \u03b1, we have also measured the integration performance with different values ranging from 0.1 to 1.0. When \u03b1 is between 0.1 and 0.4, SFE-ME is superior to EHCI-ME. For different integration directions, we found that the optimal \u03b1 value may be also different. Here, we only report two cases, \u03b1 = 0.4 for integrating documents from Google to Yahoo! and \u03b1 = 0.1 for integrating documents from Yahoo! to Google, in which the SFE approach can show its effectiveness. Table 3 and Table 4 show the macro-averaged and micro-averaged recall results of EHCI-ME and SFE-ME. The macro-averaged and micro-averaged precision results of EHCI-ME and SFE-ME are listed in Table 5 and Table 6 . In Table 7 and Table 8 , the macro-averaged and micro-averaged F 1 measure results of EHCI-ME and SFE-ME are listed, respectively.", "cite_spans": [], "ref_spans": [ { "start": 472, "end": 479, "text": "Table 3", "ref_id": null }, { "start": 484, "end": 491, "text": "Table 4", "ref_id": null }, { "start": 665, "end": 684, "text": "Table 5 and Table 6", "ref_id": "TABREF5" }, { "start": 690, "end": 709, "text": "Table 7 and Table 8", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Experimental Results and Discussion", "sec_num": "5.2" }, { "text": "From Table 3 (a), we can notice that SFE-ME is superior to EHCI-ME in more than 75% of all MaR scores for the integrations from Google to Yahoo!. Although Table 3 (b) shows that SFE-ME can only achieve nearly 40% improvements for the integration from Yahoo! to Google, SFE-ME has consistent MaR performance. Two reasons cause this lower-than-average MaR performance. First, the recall performance of SFE-ME is not as good as EHCI-ME for categories with few positive examples in the Y\u2192G integration process. This can be justified from the superior MiR performance of SFE-ME. Second, the d \u03bb weight increasingly mitigates the improvements of SFE in the MaR measures of SFE-ME in a consistent way in the Y\u2192G integration process. The MiR performance of SFE-ME also has the similar mitigation.", "cite_spans": [], "ref_spans": [ { "start": 5, "end": 12, "text": "Table 3", "ref_id": null }, { "start": 155, "end": 162, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Experimental Results and Discussion", "sec_num": "5.2" }, { "text": "From Table 3 , we can also notice that SFE-ME achieves the best MaR of 0.8935 when From table 4, we can notice that SFE-ME is superior to EHCI-ME in more than 60% of all MiR scores for the G\u2192Y integration process and in nearly 75% of all MiR scores for the Y\u2192G integration process. Among these cases, SFE-ME can achieve the best G\u2192Y MiR of Table 3 (b), the experimental results suggest that over-emphasizing the weight of Google's hierarchical thesaurus information will impair the effectiveness of SFE. From Table 5 , we can notice that SFE-ME is superior to EHCI-ME in more than 80% of all MaP for the G\u2192Y integration process, and in all cases for the Y\u2192G integration process. In addition, SFE-ME achieves the best G\u2192Y MaP of 0.6662 when s \u03bb = 1.0 and d \u03bb = 0.1, and the best Y\u2192G MaP of 0.4663 when s \u03bb = 0.7 and d \u03bb = 0.9. Table 6 , SFE-ME achieves the best G\u2192Y MiP of 0.6078 when s \u03bb = 0.9 and d \u03bb = 0.1, and the best Y\u2192G MiP of 0.2988 when s \u03bb = 0.7 and d \u03bb = 0.9. In addition, SFE-ME achieves MiP improvements in 90% of all cases for the G\u2192Y integration process and in all cases for the Y\u2192G integration process. These results show that the number of incorrectly integrated documents in SFE-ME is much lower. With high precision performance, SFE-ME may reduce a lot of time for users in manually verifying the integration correctness. For many applications, a compromised performance may be required with a high F 1 score. From Table 7 and Table 8 , we can notice that SFE-ME is superior to EHCI-ME in nearly 90% of all MaF and MiF scores for the G\u2192Y integration process, and it has consistent improvements in all cases for the Y\u2192G integration process. In our experiments with \u03b1 = 0.4, SFE-ME achieves the highest MaF (0.6839) Cheng-Zen Yang et al. We have also measured these six scores for the s \u03bb = 0.0, d \u03bb = 0.0, and \u03b1 = 0.0 cases, which means that the integration is performed by only ME without EHCI and SFE enhancements. In this configuration, for the G\u2192Y integration process, ME can achieve very prominent recall performance in MaR (0.9578) and MiR (0.9616) but with poor precision performance in MaP (0.0111) and MiP (0.0111). Its MaF and MiF are 0.022 and 0.0219, respectively. For the Y\u2192G integration process, ME has similar performance. Although ME can attain the best recall performance, these results show that it allows many documents of other categories to be incorrectly integrated. The experimental results show that SFE-ME can get more improved integration performance with the SFE scheme. Compared with EHCI-ME, SFE-ME shows that the semantic information of the hypernyms of the category-specific terms can be used to facilitate the integration process between two hierarchical taxonomies.", "cite_spans": [ { "start": 1719, "end": 1731, "text": "MaF (0.6839)", "ref_id": null }, { "start": 1732, "end": 1753, "text": "Cheng-Zen Yang et al.", "ref_id": null } ], "ref_spans": [ { "start": 5, "end": 12, "text": "Table 3", "ref_id": null }, { "start": 340, "end": 347, "text": "Table 3", "ref_id": null }, { "start": 509, "end": 516, "text": "Table 5", "ref_id": "TABREF5" }, { "start": 826, "end": 833, "text": "Table 6", "ref_id": "TABREF6" }, { "start": 1433, "end": 1452, "text": "Table 7 and Table 8", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Experimental Results and Discussion", "sec_num": "5.2" }, { "text": "In recent years, the taxonomy integration problem has been progressively studied for integrating two homogeneous hierarchical taxonomies. Many types of implicit information embedded in the source taxonomy are explored to improve the integration performance. The semantic information embedded in the source taxonomy, however, has not been discussed in previous research.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6." }, { "text": "In this paper, an enhanced integration approach (SFE) is proposed to exploit the semantic information of the hypernyms of the category-specific terms. Augmented with these additional semantic category features, the source documents can be more precisely integrated into the correct destination category in the experiments. The experimental results show that SFE-ME can achieve the best macro-averaged F 1 score and the best micro-averaged F 1 score. The results also show that the SFE scheme can get precision and recall enhancements in a significant portion of all cases.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6." }, { "text": "There are still some issues left for future study. For example, the effectiveness of SFE on other classification schemes, such as SVM and NB, may need to be investigated to decide which one has the best integration performance. In addition, deciding the optimal parameter configuration is a classical classification problem which is also important to the taxonomy integration problem. Although mining more valuable implicit information can be a tough challenge, we believe that the integration performance can be further improved with appropriate assistance of more effective auxiliary information and advanced classifiers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6." } ], "back_matter": [ { "text": "This work was supported in part by National Science Council of R.O.C. under grant NSC 96-2422-H-006-002 and NSC 96-2221-E-155-067. The authors would also like to express their gratitude to the anonymous reviewers for their precious comments.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgement", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "On Integrating Catalogs", "authors": [ { "first": "R", "middle": [], "last": "Agrawal", "suffix": "" }, { "first": "R", "middle": [], "last": "Srikan", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the 10 th International Conference on World Wide Web", "volume": "", "issue": "", "pages": "603--612", "other_ids": {}, "num": null, "urls": [], "raw_text": "Agrawal, R., & Srikan, R. (2001). On Integrating Catalogs. in Proceedings of the 10 th International Conference on World Wide Web, 603-612.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "A Maximum Entropy Approach to Natural Language Processing", "authors": [ { "first": "A", "middle": [ "L" ], "last": "Berger", "suffix": "" }, { "first": "V", "middle": [ "J D" ], "last": "Pietra", "suffix": "" }, { "first": "S", "middle": [ "A D" ], "last": "Pietra", "suffix": "" } ], "year": 1996, "venue": "Computational Linguistics", "volume": "", "issue": "", "pages": "39--71", "other_ids": {}, "num": null, "urls": [], "raw_text": "Berger, A. L., Pietra, V. J. D., & Pietra, S. A. D. (1996). A Maximum Entropy Approach to Natural Language Processing. Computational Linguistics, 39-71.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "An Iterative Approach for Web Catalog Integration with Support Vector Machines", "authors": [ { "first": "I.-X", "middle": [], "last": "Chen", "suffix": "" }, { "first": "J.-C", "middle": [], "last": "Ho", "suffix": "" }, { "first": "C.-Z", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 2 nd Asia Information Retrieval Symposium (AIRS 2005)", "volume": "", "issue": "", "pages": "703--708", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, I.-X., Ho, J.-C., & Yang, C.-Z. (2005). An Iterative Approach for Web Catalog Integration with Support Vector Machines. in Proceedings of the 2 nd Asia Information Retrieval Symposium (AIRS 2005), 703-708.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Hierarchical Web Catalog Integration with Conceptual Relationships in a Thesaurus", "authors": [ { "first": "I.-X", "middle": [], "last": "Chen", "suffix": "" }, { "first": "J.-C", "middle": [], "last": "Ho", "suffix": "" }, { "first": "C.-Z", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2007, "venue": "International Journal of Computational Linguistics and Chinese Language Processing", "volume": "12", "issue": "2", "pages": "155--174", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, I.-X., Ho, J.-C., & Yang, C.-Z. (2007). Hierarchical Web Catalog Integration with Conceptual Relationships in a Thesaurus. International Journal of Computational Linguistics and Chinese Language Processing, 12(2), 155-174.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "A Clustering-based Approach for Integrating Document-Category Hierarchies", "authors": [ { "first": "T.-H", "middle": [], "last": "Cheng", "suffix": "" }, { "first": "C.-P", "middle": [], "last": "Wei", "suffix": "" } ], "year": 2008, "venue": "IEEE Transactions on Systems, Man, and Cybernetics -Part A: Systems and Humans", "volume": "38", "issue": "", "pages": "410--424", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cheng, T.-H., & Wei, C.-P. (2008). A Clustering-based Approach for Integrating Document-Category Hierarchies. IEEE Transactions on Systems, Man, and Cybernetics -Part A: Systems and Humans, 38(2), 410-424.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Generalized Iterative Scaling for Log-linear Models", "authors": [ { "first": "J", "middle": [ "N" ], "last": "Darroch", "suffix": "" }, { "first": "D", "middle": [], "last": "Ratcliff", "suffix": "" } ], "year": 1972, "venue": "Annals of Mathematical Statistics", "volume": "43", "issue": "", "pages": "1470--1480", "other_ids": {}, "num": null, "urls": [], "raw_text": "Darroch, J. N., & Ratcliff, D. (1972). Generalized Iterative Scaling for Log-linear Models. Annals of Mathematical Statistics, 43, 1470-1480.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Information Retrieval: Data Structures and Algorithms", "authors": [ { "first": "W", "middle": [], "last": "Frakes", "suffix": "" }, { "first": "R", "middle": [], "last": "Baeza-Yates", "suffix": "" } ], "year": 1992, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Frakes, W., & Baeza-Yates, R. (1992). Information Retrieval: Data Structures and Algorithms, 1st edition, Prentice Hall, PTR.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Learning to Integrate Web Catalogs with Conceptual Relationships in Hierarchical Thesaurus", "authors": [ { "first": "J.-C", "middle": [], "last": "Ho", "suffix": "" }, { "first": "I.-X", "middle": [], "last": "Chen", "suffix": "" }, { "first": "C.-Z", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the 3 rd Asia Information Retrieval Symposium", "volume": "", "issue": "", "pages": "217--229", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ho, J.-C., Chen, I.-X., & Yang, C.-Z. (2006). Learning to Integrate Web Catalogs with Conceptual Relationships in Hierarchical Thesaurus. in Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006), 217-229.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Query Expansion with ConceptNet and WordNet: an Intrinsic Comparison", "authors": [ { "first": "M.-H", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "M.-F", "middle": [], "last": "Tsai", "suffix": "" }, { "first": "H.-H", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the 3 rd Asia Information Retrieval Symposium", "volume": "", "issue": "", "pages": "1--13", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hsu, M.-H., Tsai, M.-F., & Chen, H.-H. (2006). Query Expansion with ConceptNet and WordNet: an Intrinsic Comparison. in Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006), 1-13.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "DirectoryRank: Ordering Pages in Web Directories", "authors": [ { "first": "V", "middle": [], "last": "Krikos", "suffix": "" }, { "first": "S", "middle": [], "last": "Stamou", "suffix": "" }, { "first": "P", "middle": [], "last": "Kokosis", "suffix": "" }, { "first": "A", "middle": [], "last": "Ntoulas", "suffix": "" }, { "first": "D", "middle": [], "last": "Christodoulakis", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 7 th ACM International Workshop on Web Information and Data Management", "volume": "", "issue": "", "pages": "17--22", "other_ids": {}, "num": null, "urls": [], "raw_text": "Krikos, V., Stamou, S., Kokosis, P., Ntoulas, A., & Christodoulakis, D. (2005). DirectoryRank: Ordering Pages in Web Directories. in Proceedings of the 7 th ACM International Workshop on Web Information and Data Management (WIDM 2005), 17-22.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Feature selection, Perception Learning, and a Usability Case Study for Text Categorization", "authors": [ { "first": "H.-T", "middle": [], "last": "Ng", "suffix": "" }, { "first": "W.-B", "middle": [], "last": "Goh", "suffix": "" }, { "first": "K.-L", "middle": [], "last": "Low", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the 20 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "67--73", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ng, H.-T., Goh, W.-B., & Low, K.-L. (1997). Feature selection, Perception Learning, and a Usability Case Study for Text Categorization. in Proceedings of the 20 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 67-73.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Cross-training: Learning Probabilistic Mappings between Topics", "authors": [ { "first": "S", "middle": [], "last": "Sarawagi", "suffix": "" }, { "first": "S", "middle": [], "last": "Chakrabarti", "suffix": "" }, { "first": "S", "middle": [], "last": "Godbole", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the 9 th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining", "volume": "", "issue": "", "pages": "177--186", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sarawagi, S., Chakrabarti, S., & Godbole, S. (2003). Cross-training: Learning Probabilistic Mappings between Topics. in Proceedings of the 9 th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 177-186.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Toward Generic Title Generation for Clustered Documents", "authors": [ { "first": "Y.-H", "middle": [], "last": "Tseng", "suffix": "" }, { "first": "C.-J", "middle": [], "last": "Lin", "suffix": "" }, { "first": "H.-H", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Y.-I", "middle": [], "last": "Lin", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the 3 rd Asia Information Retrieval Symposium", "volume": "", "issue": "", "pages": "145--157", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tseng, Y.-H., Lin, C.-J., Chen, H.-H., & Lin, Y.-I. (2006). Toward Generic Title Generation for Clustered Documents. in Proceedings of the 3 rd Asia Information Retrieval Symposium (AIRS 2006), 145-157.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Learning to Integrate Web Taxonomies with Fine-Grained Relations: A Case Study Using Maximum Entropy Model", "authors": [ { "first": "C.-W", "middle": [], "last": "Wu", "suffix": "" }, { "first": "T.-H", "middle": [], "last": "Tsai", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "Hsu", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 2 nd Asia Information Retrieval Symposium (AIRS 2005)", "volume": "", "issue": "", "pages": "190--205", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wu, C.-W., Tsai, T.-H., & Hsu, W.-L. (2005). Learning to Integrate Web Taxonomies with Fine-Grained Relations: A Case Study Using Maximum Entropy Model. in Proceedings of the 2 nd Asia Information Retrieval Symposium (AIRS 2005), 190-205.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Web Taxonomy Integration with Hierarchical Shrinkage algorithm and Fine-Grained Relations", "authors": [ { "first": "C.-W", "middle": [], "last": "Wu", "suffix": "" }, { "first": "T.-H", "middle": [], "last": "Tsai", "suffix": "" }, { "first": "C.-W", "middle": [], "last": "Lee", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "Hsu", "suffix": "" } ], "year": 2008, "venue": "Expert Systems with Applications", "volume": "35", "issue": "4", "pages": "2123--2131", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wu, C.-W., Tsai, T.-H., Lee, C.-W., & Hsu, W.-L. (2008). Web Taxonomy Integration with Hierarchical Shrinkage algorithm and Fine-Grained Relations. Expert Systems with Applications, 35(4), 2123-2131.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Semantic Feature Expansion on Category-Specific Terms", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Semantic Feature Expansion on Category-Specific Terms", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "A Comparative Study on Feature Selection in Text Categorization", "authors": [ { "first": "Y", "middle": [], "last": "Yang", "suffix": "" }, { "first": "J", "middle": [ "O" ], "last": "Pedersen", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the 14 th International Conference on Machine Learning (ICML'97", "volume": "", "issue": "", "pages": "412--420", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yang, Y., & Pedersen, J. O. (1997). A Comparative Study on Feature Selection in Text Categorization. in Proceedings of the 14 th International Conference on Machine Learning (ICML'97), 412-420.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Web Taxonomy Integration using Support Vector Machines", "authors": [ { "first": "D", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "W.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 13 th International Conference on World Wide Web", "volume": "", "issue": "", "pages": "472--481", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhang, D., & Lee, W.-S. (2004a). Web Taxonomy Integration using Support Vector Machines. in Proceedings of the 13 th International Conference on World Wide Web, 472-481.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Web Taxonomy Integration Through Co-Bootstrapping", "authors": [ { "first": "D", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "W.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 27 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "410--417", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhang, D., & Lee, W.-S. (2004b). Web Taxonomy Integration Through Co-Bootstrapping. in Proceedings of the 27 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 410-417.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "CatRelate: A New Hierarchical Document Category Integration Algorithm by learning Category Relationships", "authors": [ { "first": "S", "middle": [], "last": "Zhu", "suffix": "" }, { "first": "C", "middle": [ "C" ], "last": "Yang", "suffix": "" }, { "first": "W", "middle": [], "last": "Lam", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 7 th International Conference on Asian Digital Libraries (ICADL 2004)", "volume": "", "issue": "", "pages": "280--289", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhu, S., Yang, C. C., & Lam, W. (2004). CatRelate: A New Hierarchical Document Category Integration Algorithm by learning Category Relationships. in Proceedings of the 7 th International Conference on Asian Digital Libraries (ICADL 2004), Shanghai, China, 280-289.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Online Resources Information Mapping Project", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Online Resources Information Mapping Project, Computational Semantics Laboratory, Stanford University. http://infomap.stanford.edu/.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "A lexical database for the English language", "authors": [ { "first": "", "middle": [], "last": "Wordnet", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "WordNet, A lexical database for the English language: Cognitive Science Laboratory, Princeton University, http://wordnet.princeton.edu/.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Maximum Entropy Modeling Toolkit for Python and C++", "authors": [ { "first": "L", "middle": [], "last": "Zhang", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhang, L., \"Maximum Entropy Modeling Toolkit for Python and C++,\" http://homepages.inf.ed.ac.uk/s0450736/maxent.html.", "links": null } }, "ref_entries": { "FIGREF1": { "type_str": "figure", "text": "= 0.1 and d \u03bb = 0.1 for the G\u2192Y integration process. Although Table 3 (b) shows that EHCI-ME achieves the best MaR for the Y\u2192G integration process, SFE-ME indeed achieves higher MaR of 0.9501 in our experiment while s \u03bb = 0.1, d \u03bb = 0.1, and \u03b1 = 0.4. 0.7317 0.7261 0.7262 0.7292 0.7207 0.7081 0.7284 0.7338 0.7338 0.7338 0.70 0.7262 0.7242 0.7233 0.7263 0.7191 0.6941 0.7227 0.7333 0.7338 0.7338 0.80 0.7231 0.7205 0.7232 0.7253 0.7235 0.6922 0.7208 0.7277 0.7304 0results of the integration from Yahoo! to Google", "num": null, "uris": null }, "FIGREF2": { "type_str": "figure", "text": "0.9301 and the best Y\u2192G MiR of 0.9055 when s \u03bb = 0.1 and d \u03bb = 0.1. When d \u03bb = EHCI-ME outperforms SFE-ME for both MaR and MiR in the G\u2192Y integration process. Considering the d \u03bb influences of Google's hierarchical thesaurus information shown in", "num": null, "uris": null }, "FIGREF3": { "type_str": "figure", "text": "The micro-averaged recall (MiR) measures of EHCI-ME and SFE-ME. 0.8562 0.8624 0.8627 0.8620 0.8552 0.8545 0.8463 0.8425 0.8418 (b) The results of the integration from Yahoo! to Google Hierarchical Taxonomy Integration Using 435 Semantic Feature Expansion on Category-Specific Terms", "num": null, "uris": null }, "FIGREF4": { "type_str": "figure", "text": "0.1012 0.1024 0.1019 0.1017 0.1661 0.1797 0.1847 0.1876 0.1881 (b) The results of the integration from Yahoo! to GoogleFrom", "num": null, "uris": null }, "FIGREF5": { "type_str": "figure", "text": "and the highest MiF (0.6764) when s \u03bb = 0.6 and d \u03bb = 0.1 for the G\u2192Y integration process. For the Y\u2192G integration process, SFE-ME achieves the highest MaF (0.5919) and the highest MiF (0.4413) when \u03b1 = 0.1, s \u03bb = 0.7, and d \u03bb = 0.9. These two tables show that the SFE scheme can mostly get more balanced improvements in both recall and precision considerations.", "num": null, "uris": null }, "FIGREF6": { "type_str": "figure", "text": "The macro-averaged F 1 (MaF) measures of EHCI-ME and SFE-ME. 0.2355 0.3223 0.3346 0.3345 0.3426 0.3574 0.5225 0.5459 0.5829 0.5904 0.70 0.2354 0.3234 0.3379 0.3352 0.3411 0.3629 0.5251 0.5472 0.5834 0.5919 0.80 0.2358 0.3247 0.3388 0.3356 0.3405 0.3672 0.5251 0.5484 0.5839 0The results of the integration from Yahoo! to Google", "num": null, "uris": null }, "FIGREF7": { "type_str": "figure", "text": "438", "num": null, "uris": null }, "FIGREF8": { "type_str": "figure", "text": "0.2021 0.1991 0.1982 0.1978 0.1975 0.2508 0.3413 0.3623 0.3973 0.4413 0.80 0.2025 0.2000 0.1980 0.1976 0.1974 0.2558 0.3439 0.3633 0.3964 0.4408 0.90 0.2027 0.1989 0.1972 0.1975 0.1964 0.2584 0.3441 0.3667 0.3952 0", "num": null, "uris": null }, "TABREF0": { "text": "The experimental results show that the CS-enhanced Transductive SVMs give", "num": null, "type_str": "table", "content": "
Hierarchical Taxonomy Integration Using425
Semantic Feature Expansion on Category-Specific Terms
significant improvement to the original T-SVMs and consistently outperform ENB.
", "html": null }, "TABREF2": { "text": "", "num": null, "type_str": "table", "content": "", "html": null }, "TABREF4": { "text": "", "num": null, "type_str": "table", "content": "
CategoryGoogle|G-Y| |G Class| |G Test|Yahoo!|Y-G| |Y Class| |Y Test|
Autos/autos/\u2026109612427 /automotive/\u2026 168124436
Movies /movies/\u20265188261422 /movies_Film/\u2026 7255271344
Outdoors /outdoors/\u2026239616208 /outdoors/\u2026157919210
Photo/photography/\u2026 6159235 /photography/\u2026 130423218
Software /software/\u2026582927641 /software/\u2026187625691
Total1512490293213695 1082918
", "html": null }, "TABREF5": { "text": "", "num": null, "type_str": "table", "content": "
. The macro-averaged precision (MaP) measures of EHCI-ME and SFE-ME.
EHCI-MESFE-ME (\u03b1 = 0.4)
\u03bb s\u03bb d0.100.300.500.700.900.100.300.500.700.90
0.100.1936 0.3273 0.3356 0.3426 0.3425 0.2122 0.2980 0.3139 0.3158 0.3557
0.200.3491 0.3482 0.3475 0.3459 0.3559 0.3664 0.3696 0.3572 0.3477 0.3510
0.300.3890 0.3537 0.3486 0.3460 0.3547 0.4707 0.3960 0.3793 0.3523 0.3486
0.400.4090 0.3613 0.3497 0.3482 0.3543 0.5794 0.4137 0.3797 0.3723 0.3531
0.500.4253 0.3657 0.3515 0.3521 0.3560 0.6279 0.4649 0.3971 0.3778 0.3552
0.600.4373 0.3734 0.3565 0.3588 0.3603 0.6613 0.4918 0.4192 0.3556 0.3624
0.700.4455 0.3811 0.3611 0.3681 0.3655 0.6600 0.5592 0.4397 0.3663 0.3916
0.800.4532 0.3876 0.3686 0.3735 0.3559 0.6607 0.6403 0.4872 0.3876 0.3333
0.900.4548 0.3904 0.3747 0.3853 0.3607 0.6636 0.6543 0.5738 0.4321 0.3548
1.000.4565 0.4125 0.3862 0.4070 0.3625 0.6662 0.6575 0.5955 0.5043 0.4304
", "html": null }, "TABREF6": { "text": "", "num": null, "type_str": "table", "content": "
. The micro-averaged precision (MiP) measures of EHCI-ME and SFE-ME.
EHCI-MESFE-ME (\u03b1 = 0.4)
\u03bb s\u03bb d0.100.300.500.700.900.100.300.500.700.90
0.100.1156 0.2504 0.2715 0.2740 0.2817 0.1205 0.2835 0.3099 0.2782 0.3687
0.200.2253 0.3018 0.2947 0.2822 0.3080 0.1569 0.3661 0.3570 0.3504 0.3629
0.300.2741 0.3170 0.2984 0.2858 0.3107 0.2737 0.3777 0.3946 0.3515 0.3642
0.400.3136 0.3329 0.3002 0.2897 0.3115 0.4721 0.3834 0.3776 0.3866 0.3688
0.500.3494 0.3390 0.3033 0.2695 0.3135 0.5581 0.4556 0.3862 0.3879 0.3666
0.600.3763 0.3475 0.3101 0.3101 0.3199 0.6061 0.4663 0.4032 0.3147 0.3700
0.700.3906 0.3583 0.3192 0.3336 0.3317 0.6041 0.4924 0.3952 0.3180 0.4151
0.800.3966 0.3759 0.3334 0.3485 0.3414 0.6016 0.5824 0.4335 0.3229 0.3452
0.900.3987 0.3826 0.3402 0.3734 0.3540 0.6078 0.5871 0.4726 0.3509 0.3800
1.000.3992 0.4332 0.3772 0.4198 0.3568 0.5999 0.5894 0.4937 0.3879 0.3974
(a) The results of the integration from Google to Yahoo!
EHCI-MESFE-ME (\u03b1 = 0.1)
\u03bb s\u03bb d0.100.300.500.700.900.100.300.500.700.90
0.100.0677 0.1269 0.1172 0.1118 0.1111 0.0943 0.1818 0.2089 0.2344 0.2686
0.200.0999 0.1226 0.1145 0.1121 0.1120 0.1183 0.1903 0.2184 0.2485 0.2829
0.300.1087 0.1186 0.1137 0.1123 0.1122 0.1269 0.1929 0.2278 0.2534 0.2860
0.400.1125 0.1158 0.1131 0.1124 0.1119 0.1318 0.1948 0.2299 0.2585 0.2909
0.500.1142 0.1152 0.1127 0.1122 0.1121 0.1365 0.1984 0.2303 0.2606 0.2943
0.600.1147 0.1131 0.1123 0.1121 0.1118 0.1418 0.2042 0.2304 0.2612 0.2983
0.700.1147 0.1128 0.1122 0.1119 0.1117 0.1469 0.2131 0.2301 0.2597 0.2988
0.800.1150 0.1134 0.1121 0.1118 0.1116 0.1503 0.2152 0.2312 0.2591 0.2985
0.900.1151 0.1127 0.1116 0.1117 0.1110 0.1522 0.2154 0.2340 0.2581 0.2988
1.000.0979 0.0867 0.0865 0.0870 0.0865 0.1190 0.1388 0.1417 0.1468 0.1573
(b) The results of the integration from Yahoo! to Google
", "html": null }, "TABREF7": { "text": "", "num": null, "type_str": "table", "content": "
. The micro-averaged F 1 (MiF) measures of EHCI-ME and SFE-ME.
", "html": null } } } }