{ "paper_id": "R13-1041", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T14:56:34.125985Z" }, "title": "A Boosting-based Algorithm for Classification of Semi-Structured Text using Frequency of Substructures", "authors": [ { "first": "Tomoya", "middle": [], "last": "Iwakura", "suffix": "", "affiliation": { "laboratory": "", "institution": "Fujitsu Laboratories Ltd", "location": {} }, "email": "iwakura.tomoya@jp.fujitsu.com" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Research in text classification currently focuses on challenging tasks such as sentiment classification, modality identification, and so on. In these tasks, approaches that use a structural representation, like a tree, have shown better performance rather than a bag-of-words representation. In this paper, we propose a boosting algorithm for classifying a text that is a set of sentences represented by tree. The algorithm learns rules represented by subtrees with their frequency information. Existing boostingbased algorithms use subtrees as features without considering their frequency because the existing algorithms targeted a sentence rather than a text. In contrast, our algorithm learns how the occurrence frequency of each subtree is important for classification. Experiments on topic identification of Japanese news articles and English sentiment classification shows the effectiveness of subtree features with their frequency.", "pdf_parse": { "paper_id": "R13-1041", "_pdf_hash": "", "abstract": [ { "text": "Research in text classification currently focuses on challenging tasks such as sentiment classification, modality identification, and so on. In these tasks, approaches that use a structural representation, like a tree, have shown better performance rather than a bag-of-words representation. In this paper, we propose a boosting algorithm for classifying a text that is a set of sentences represented by tree. The algorithm learns rules represented by subtrees with their frequency information. Existing boostingbased algorithms use subtrees as features without considering their frequency because the existing algorithms targeted a sentence rather than a text. In contrast, our algorithm learns how the occurrence frequency of each subtree is important for classification. Experiments on topic identification of Japanese news articles and English sentiment classification shows the effectiveness of subtree features with their frequency.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Text classification is used to classify texts such as news articles, E-mails, social media posts, and so on. A number of machine learning algorithms have been applied to text classification successfully. Text classification handles not only tasks to identify topics, such as politics, finance, sports or entertainment, but also challenging tasks such as categorization of customer E-mails and reviews by types of claims, subjectivity or sentiment (Wiebe, 2000; Banea et al., 2010; Bandyopadhyay and Okumura, 2011) . To identify difficult categories on challenging tasks, a traditional bag-of-words representation may not be sufficient. Therefore, a richer, structural representation is used rather than the traditional bag-of-words. A straightforward way to extend the traditional bag-of-words representation is to heuristically add new types of features such as fixed-length n-grams such as word bi-gram or tri-gram, or fixed-length syntactic relations. Instead of such approaches, learning algorithms that handle semi-structured data have become increasingly popular (Kudo and Matsumoto, 2004; Kudo et al., 2005; Ifrim et al., 2008; Okanohara and Tsujii, 2009) . This is due to the fact that these algorithms can learn better substructures for each task from semi-structured texts annotated with parts-of-speech, base-phrase information or syntactic relations.", "cite_spans": [ { "start": 447, "end": 460, "text": "(Wiebe, 2000;", "ref_id": null }, { "start": 461, "end": 480, "text": "Banea et al., 2010;", "ref_id": "BIBREF2" }, { "start": 481, "end": 513, "text": "Bandyopadhyay and Okumura, 2011)", "ref_id": null }, { "start": 1069, "end": 1095, "text": "(Kudo and Matsumoto, 2004;", "ref_id": "BIBREF11" }, { "start": 1096, "end": 1114, "text": "Kudo et al., 2005;", "ref_id": "BIBREF12" }, { "start": 1115, "end": 1134, "text": "Ifrim et al., 2008;", "ref_id": "BIBREF7" }, { "start": 1135, "end": 1162, "text": "Okanohara and Tsujii, 2009)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Among such learning algorithms, boostingbased algorithms have the following advantages: Boosting-based learning algorithms have been applied to Natural Language Processing problems successfully, including text classification (Kudo and Matsumoto, 2004) , English syntactic chunking (Kudo et al., 2005) , zero-anaphora resolution (Iida et al., 2006) , and so on. Furthermore, classifiers trained with boosting-based learners have shown faster classification speed (Kudo and Matsumoto, 2004) than Support Vector Machines with a tree kernel (Collins and Duffy, 2002) .", "cite_spans": [ { "start": 225, "end": 251, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" }, { "start": 281, "end": 300, "text": "(Kudo et al., 2005)", "ref_id": "BIBREF12" }, { "start": 328, "end": 347, "text": "(Iida et al., 2006)", "ref_id": "BIBREF8" }, { "start": 462, "end": 488, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" }, { "start": 537, "end": 562, "text": "(Collins and Duffy, 2002)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "However, existing boosting-based algorithms for semi-structured data, boosting algorithms for classification (Kudo and Matsumoto, 2004) and for ranking (Kudo et al., 2005) , have the following point that can be improved. The weak learners used in these algorithms learn classifiers which do not consider frequency of substructures. This is because these algorithms targeted a sentence as their input rather than a document or text consisting of two or more sentences. Therefore, even if crucial substructures appear several times in their target texts, these algorithms cannot reflect such frequency. For example, on sentiment classification, different types of negative expressions may be preferred to a positive expression which ap-pears several times. As a result, it may happen that a positive text using the same positive expression several times with some types of negative expressions is classified as a negative text because consideration of frequency is lacking.", "cite_spans": [ { "start": 109, "end": 135, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" }, { "start": 152, "end": 171, "text": "(Kudo et al., 2005)", "ref_id": "BIBREF12" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "This paper proposes a boosting-based algorithm for semi-structured data that considers the occurrence frequency of substructures. To simplify the problem, we first assume that a text to be classified is represented as a set of sentences represented by labeled ordered trees (Abe et al., 2002) . Word sequence, base-phrase annotation, dependency tree and an XML document can be modeled as a labeled ordered tree. Experiments on topic identification of news articles and sentiment classification confirm the effectiveness of subtree features with their frequency.", "cite_spans": [ { "start": 274, "end": 292, "text": "(Abe et al., 2002)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Prior boosting-based algorithms for semistructured data, such as boosting algorithms for classification (Kudo and Matsumoto, 2004) and for ranking (Kudo et al., 2005) , learns classifiers which do not consider frequency of substructures. Ifrim et. al (Ifrim et al., 2008) proposed a logistic regression model with variable-length N-gram features.", "cite_spans": [ { "start": 104, "end": 130, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" }, { "start": 147, "end": 166, "text": "(Kudo et al., 2005)", "ref_id": "BIBREF12" }, { "start": 251, "end": 271, "text": "(Ifrim et al., 2008)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "The logistic regression learns the weights of N-gram features. Compared with these two algorithms, our algorithm learns frequency thresholds to consider occurrence frequency of each subtree.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "Okanohara and Tsujii (Okanohara and Tsujii, 2009) proposed a document classification method using all substrings as features. The method uses Suffix arraies (Manber and Myers, 1990) for efficiently using all substrings. Therefore, the trees used in our method are not handled. Their method uses feature types of N-grams features, such as term frequency, inverted document frequency, and so on, in a logistic regression. In contrast, our algorithm differs in the learning of a threshold for feature values. Tree kernel (Collins and Duffy, 2002; Kashima and Koyanagi, 2002) implicitly maps the example represented in a labeled ordered tree into all subtree spaces, and Tree kernel can consider the frequency of subtrees. However, as discussed in the paper (Kudo and Matsumoto, 2004) , when Tree kernel is applied to sparse data, kernel dot products between similar instances become much larger than those between different instances. As a result, this sometimes leads to overfitting in training. In contrast, our boosting algo-rithm considers the frequency of subtrees by learning the frequency thresholds of subtrees. Therefore, we think the problems caused by Tree kernel do not tend to take place because of the difference presented in the boosting algorithm (Kudo and Matsumoto, 2004) .", "cite_spans": [ { "start": 21, "end": 49, "text": "(Okanohara and Tsujii, 2009)", "ref_id": "BIBREF14" }, { "start": 157, "end": 181, "text": "(Manber and Myers, 1990)", "ref_id": "BIBREF13" }, { "start": 518, "end": 543, "text": "(Collins and Duffy, 2002;", "ref_id": "BIBREF4" }, { "start": 544, "end": 571, "text": "Kashima and Koyanagi, 2002)", "ref_id": "BIBREF9" }, { "start": 754, "end": 780, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" }, { "start": 1260, "end": 1286, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "3 A Boosting-based Learning Algorithm for Classifying Trees", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "We describe the problem treated by our boostingbased learner as follows. Let X be all labeled ordered trees, or simply trees, and Y be a set of labels {\u22121, +1}. A labeled ordered tree is a tree where each node is associated with a label. Each node is also ordered among its siblings. Therefore, there are a first child, second child, third child, and so on (Abe et al., 2002) . Let S be a set of training samples", "cite_spans": [ { "start": 357, "end": 375, "text": "(Abe et al., 2002)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "{(x 1 , y 1 ), ..., (x m , y m )},", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "where each example x i \u2208 X is a set of labeled ordered trees, and y i \u2208 Y is a class label.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "The goal is to induce a mapping", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "F : X \u2192 Y from S.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "Then, we define subtrees (Abe et al., 2002) .", "cite_spans": [ { "start": 25, "end": 43, "text": "(Abe et al., 2002)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "Definition 1 Subtree Let u and t be labeled ordered trees. We call t a subtree of u, if there exists a one-to-one mapping \u03c6 between nodes in t to u, satisfying the conditions: (1) \u03c6 preserves the parent relation, (2) \u03c6 preserves the sibling relation, and (3) \u03c6 preserves the labels. We denote t as a subtree of u as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "t \u2286 u .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "If a tree t is not a sbutree of u, we denote it as t \u2288 u .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "We define the frequency of the subtree t in u as the number of times t occurs in u and denoted as |t \u2286 u| .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "The number of nodes in a tree t is referred as the size of the tree t and denote it as |t| .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "To represent a set of labeled ordered trees, we use a single tree created by connecting the trees with the root node of the single tree in this paper. Figure 1 is an example of subtrees of a tree consisting of two sentences \"a b c\" and \"a b\" connected with the root node R \u20dd. The trees in the right box are a portion of subtrees of the left tree. Let u be Figure 1 : A labeled ordered tree and its subtrees. the tree in the left side. For example, the size of the subtree a \u20ddb \u20dd (i.e. | a \u20ddb \u20dd|) is 2 and the frequency | a \u20ddb \u20dd \u2208 u| is also 2. For the subtree a \u20ddc \u20dd, the size | a \u20ddc \u20dd| is also 2, however, the frequency", "cite_spans": [], "ref_spans": [ { "start": 151, "end": 159, "text": "Figure 1", "ref_id": null }, { "start": 356, "end": 364, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "| a \u20dd- c \u20dd \u2208 u | is 1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "3.2 A Classifier for Trees with the Occurrence Frequency of a Subtree We define a classifier for trees -that is used as weak hypothesis in this paper. A boosting algorithm for classifying trees uses subtree-based decision stumps, and each decision stump learned by the boosting algorithm classifies trees whether a tree is a subtree of trees to be classified or not (Kudo and Matsumoto, 2004) . To consider the frequency of a subtree, we define the following decision stump.", "cite_spans": [ { "start": 366, "end": 392, "text": "(Kudo and Matsumoto, 2004)", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Preliminaries", "sec_num": "3.1" }, { "text": "Let t and u be trees, z be a positive integer, called frequency threshold, and a and b be a real number, called a confidence value, then a classifier for trees is defined as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 2 Classifier for trees", "sec_num": null }, { "text": "h \u27e8t,z,a,b\u27e9 (u) = \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 a t \u2286 u \u2227 z \u2264 |t \u2286 u| \u2212a t \u2286 u \u2227 |t \u2286 u| < z b otherwise .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 2 Classifier for trees", "sec_num": null }, { "text": "Each decision stump has a subtree t and its frequency threshold z as a condition of classification, and two scores, a and b. If t is a subtree of u (i.e. t \u2286 u), and the frequency of the subtree |t \u2286 u| is greater than or equal to the frequency threshold z, the score a is assigned to the tree. If u satisfies t \u2286 u and |t \u2286 u| is less than z, the score -a is assigned to the tree. If t is not a subtree of u (i.e. t \u2288 u), the score b is assigned to the tree. This classifier is an extension of decision trees learned by learning algorithms like C4.5 (Quinlan, 1993) for classifying trees. For example, C4.5 learns the thresholds for features that have continuous values, and C4.5 uses the thresholds for classifying samples including continuous values. In a similar way, each decision stump for trees uses a frequency threshold for classifying samples with a frequency of a subtree.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 2 Classifier for trees", "sec_num": null }, { "text": "Classifying Trees To induce accurate classifiers, a boosting algorithm is applied. Boosting is a method to create a final hypothesis by repeatedly generating a weak hypothesis in each training iteration with a given weak learner. These weak hypotheses are combined as the final hypothesis. We use real Ad-aBoost used in BoosTexter (Schapire and Singer, 2000) since real AdaBoost-based text classifiers show better performance than other algorithms, such as discrete AdaBoost (Freund and Schapire, 1997) .", "cite_spans": [ { "start": 331, "end": 358, "text": "(Schapire and Singer, 2000)", "ref_id": "BIBREF16" }, { "start": 475, "end": 502, "text": "(Freund and Schapire, 1997)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "Our boosting-based learner selects R types of rules for creating a final hypothesis F on several training iterations. The F is defined as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "F (u) = sign( \u2211 R r=1 h \u27e8tr,zrar,br\u27e9 (u)).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "We use a learning algorithm that learns a subtree and its frequency threshold as a rule from given training samples", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "S = {(x i , y i )} m i=1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "and weights over samples {w r,1 , ..., w r,m } as a weak learner. By training the learning algorithm R times with different weights of samples, we obtain R types of rules.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Boosting-based Rule Learning for", "sec_num": "3.3" }, { "text": "w r,i is the weight of sample number i after selecting r \u2212 1 types of rules, where 0", "html": null, "text": "is the weight of correctly classified samples that have +1 as their labels, and W r\u27e8\u22121,<,z\u27e9 (t) is the weight of correctly classified samples that have -1 as their labels.W \u00ac r\u27e8y\u27e9 (t) is the sum of the weights of samples that a rule is not applied to (i.e. t \u2288 x i ) and y", "type_str": "table" }, "TABREF1": { "num": null, "content": "
Mainichi Shimbun
CategoryTrainingDevelopmentTest
#P#N#W#P#N#W#P#N#W
business4,782 18,790 67,452597 2,348 29,023597 2,348 29,372
entertainment938 22,632 67,682117 2,829 29,330117 2,829 28,939
international4,693 18,879 67,705586 2,359 28,534586 2,359 29,315
sports12,687 10,884 67,592 1,586 1,360 28,658 1,585 1,360 29,024
technology473 23,097 67,51659 2,887 29,33759 2,887 28,571
", "html": null, "text": "Statistics of Mainichi Shimbun data set. #P, #N and #W relate to the number of positive samples, the number of negative samples, and the number of distinct words, respectively.", "type_str": "table" }, "TABREF2": { "num": null, "content": "
Amazon review data
CategoryTrainingDevelopmentTest
#N#P#W#N#P#W#N#P#W
books357,319 2,324,575 1,327,312 44,664 290,571 496,453 44,664 290,571 496,412
dvd52,674352,213446,6286,58444,026 157,4956,58444,026 155,468
electronics12,04740,58485,5431,5065,07326,9451,5055,07326,914
music35,050423,654571,3994,38152,956 180,2134,38152,956 179,787
video13,47988,189161,9201,68511,02361,3791,68411,02361,958
", "html": null, "text": "Statistics of Amazon data set. #N, #P and #W relate to the number of negative reviews, the number of positive reviews, and the number of distinct words, respectively.", "type_str": "table" }, "TABREF3": { "num": null, "content": "
BOWTree
Category+FQ-FQ+FQ-FQ
business88.7988.87 * 91.45 * 90.89
entertaiment95.07 * 94.2795.11 * 94.64
international85.2585.99 * 87.9188.28 *
sports98.1798.52 * 98.70 * 98.64
technology83.02 * 78.5079.2180.77 *
", "html": null, "text": "Experimental Results of the training on the Mainichi Shinbun. Results in bold show the best accuracy, and while an underline means the accuracy of a boosting is better than the booting algorithm with the same feature representation (e.g. Tree(-FQ) for Tree(+FQ)) on each category.", "type_str": "table" }, "TABREF4": { "num": null, "content": "
BOWN-gram
Category+FQ-FQ+FQ-FQ
books74.35
", "html": null, "text": "Experimental Results of the training on the Amazon data set. The meaning of results in bold and each underline are the same asFigure 3.74.13 87.33 * 87.20 dvd 83.18 * 82.96 93.35 93.66 * electronics 89.39 * 89.06 93.36 93.57 * music 77.85 * 77.57 91.65 * 91.30 video 95.09 * 95.04 97.10", "type_str": "table" }, "TABREF5": { "num": null, "content": "", "html": null, "text": "Janyce Wiebe. 2000. Learning subjective adjectives from corpora. In Proceedings of the Seventeenth National Conference on Artificial Intelligence and Twelfth Conference on Innovative Applications of Artificial Intelligence, pages 735-740. AAAI Press. Mohammed Javeed Zaki. 2002. Efficiently mining frequent trees in a forest. In Proc. of KDD'02, pages 71-80.", "type_str": "table" } } } }