| { |
| "paper_id": "Y02-1021", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T13:43:41.773573Z" |
| }, |
| "title": "Generating a Category Set of Words Using a Hierarchical Part-of-Speech System and Tagged Corpus", |
| "authors": [ |
| { |
| "first": "Takeyuki", |
| "middle": [], |
| "last": "Kojima", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Tokyo University of Agric. and Tech", |
| "location": { |
| "addrLine": "2-24-16 Nakacho", |
| "settlement": "Koganei, Tokyo", |
| "country": "Japan" |
| } |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Yoshiyuki", |
| "middle": [], |
| "last": "Kotani", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Tokyo University of Agric. and Tech", |
| "location": { |
| "addrLine": "2-24-16 Nakacho", |
| "settlement": "Koganei, Tokyo", |
| "country": "Japan" |
| } |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "In this paper, we propose a method of generating a proper categorization of morphemes by giving a hierarchical part-of-speech system and a corpus tagged using this part-ofspeech system. Our method use hierarchical information in the part-of-speech system and statistical information in the corpus to generate a category set. The statistical information is based on the context of occurrence of categories. First, we specify the format of given information. Then, we describe an algorithm to generate a proper categorization. Finally, we present the results of our experiments in applying this method. We obtaind a moderately proper categorization and found several candidates for improvement.", |
| "pdf_parse": { |
| "paper_id": "Y02-1021", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "In this paper, we propose a method of generating a proper categorization of morphemes by giving a hierarchical part-of-speech system and a corpus tagged using this part-ofspeech system. Our method use hierarchical information in the part-of-speech system and statistical information in the corpus to generate a category set. The statistical information is based on the context of occurrence of categories. First, we specify the format of given information. Then, we describe an algorithm to generate a proper categorization. Finally, we present the results of our experiments in applying this method. We obtaind a moderately proper categorization and found several candidates for improvement.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "In natural language processing, it is important to categorize words or morphemes properly. A proper categorization depends on, among other things, the kind of processing task, the domain of target corpus, and the size of the corpus. When categorization is too general, we can not use characters of individual categories. Because the characters of categories hide each other. On the contrary, when categorization is too specific, we can not use the characters of categories also. Because a low frequency of a category decreases the reliability of characters of the category.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Past researches have proposed categorizations and tagsets for differnt purposes: morphological analysis, syntactic analysis or information extraction, etc. While some of the categorizations are made by hand using linguistic knowledge, others are created from annotated corpora automatically or semiautomatically. Several researches have forcused on methods of modifing the existing categorization in order to improve the accuracy of their task with respect to their purpose (Brants, 1995) . Criteria of categorizations in these researches are the accuracies of their tasks. Other researches have proposed an criteria of categorizations of words based on linguistic quality and not processing quality (Dejean, 2000) .", |
| "cite_spans": [ |
| { |
| "start": 474, |
| "end": 488, |
| "text": "(Brants, 1995)", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 700, |
| "end": 714, |
| "text": "(Dejean, 2000)", |
| "ref_id": "BIBREF1" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "We propose here a method to decide a proper categorization of morphemes, giving a hierarchical part-of-speech system and a corpus tagged using this part-of-speech system. In other words, our method forcuses on reducing an existing category set using hierarchical information of part-of-speech system and statistical information of the corpus. We recursively subdivide the categories using topdown approach with subdivision score. The subdivision score, which indicates how significant it is to subivide a category, is based on difference between the context of the category and that of its parent category.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "We explain structure of given information, which consist of a part-of-speech system and a tagged corpus, and a generating category set in Section 2. Section 3 explains the method of generating a category set. Section 4 shows experiments that are performed in order to test the generating algorithm. In the next section, we disscuss the result of the experiments.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "2 Given Information and Generated Category Set", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The part-of-speech system we use in this research is a tree structure specified by a triplet U = (VM , Vp, P), where VM is a set of the leaf part-of-speech categories of this tree structure. Vp is a set of the part-of-speech categories occupying the intermediate (or non-leaf) nodes including the root node p. In other words, each element of VM is one of the most specific categories in the part-of-speech system. For instance, p-N-NO, p-N-PN-NT and p-SYM are elements of VM , while p-N, p-N-PN and p are elements of Vp 1 . A hyphen stands for a parent-child relationship in these category names.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Hierarchical Part-of-Speech System", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "The parent-child relationship between the nodes of a tree structure is defined by the parent function P : VM U Vp -4 Vp. A category P(x) is more general than a category x. For instance, P(p-N-PN) = p-N and P(p-SYM) = p. The parent of the root node is undefined.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Hierarchical Part-of-Speech System", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "Given a parent function P, the children function C is defined as follows :", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Hierarchical Part-of-Speech System", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "C (x) = {y IP(y) = x} .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Hierarchical Part-of-Speech System", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "C(x) is a set of all children of a category x.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Hierarchical Part-of-Speech System", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "The tagged corpus we use in this research is tagged with the part-of-speech system described in Section 2.1. The corpus is tagged with the most specific categories. In other words, each word in the original corpus is replaced with the leaf category of the part-of-speech system.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Tagged Corpus", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "Our objective is to generate a category set G from a part-of-speech system U and a tagged corpus, which is a subset of the category set of U; that is, G C VM U Vp . Figure 1 shows a hierarchical part-of-speech system U and a category set G. In this figure, a circle stands for category and a line stands for the parent-child relationship between categories.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 165, |
| "end": 173, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Since an input tagged corpus consists of leaf categories, all leaf categories must be replaced by corresponding categories that are elements of category set G. We put requirements on G, in order to ensure that all leaf categories have its ancestor category in G.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "We recursively define a function as follows:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "6U,G (x) { x, x E G Su (P(x)), otherwise This equation means that (5 U' G (X)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "is an ancestor of x and it is an element of G. We require that the generated category set G must satisfy the following two requirements:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "(1) 6 U'G (X) is defined for all x E VM , (2) P(x) sg' G for all x E G.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Category Set", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "We quantify the context of occurrence in a corpus H with a category set G by the terms pH (x; z) and plI R ' u' G (x; z) as follows:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "PL (X; Z) PH R '\" (X ; Z) fGrl'U (x z) fG ILU (z) II,U f G (Z \u2022 X) f G H'u CO \u2022", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "In these equations, fG li'u (x) is the frequency of the category x in the corpus obtained by replacing each morpheme by its nearest ancestor category that is an element of G2.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "Here, a ll 'u'G (x; z) is the conditional probability that category x precedes category z, whereas", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "H,U, p G R (x, z)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "is the conditional probability that category x follows category z. We call this pair of probability distributions the environment of category z.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "3 Generating a Category Set", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Environment of a Category", |
| "sec_num": "2.4" |
| }, |
| { |
| "text": "At the start of the algorithm, we set the category set to be a singleton containing only the root node. Then, we subdivide the category that must be subdivided, topdown. There are several possible methods of deciding whether we must subdivide a certain category or not. In this paper, we use a subdivision score based on Kullback-Leibler distance divergence. First, we quantify the difference, with respect to environments, between category z and its parent category v = P(z) by the following two terms:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "DL H,U,G (z) DR H,U,G (z) H,U,G x; z) EpL il'\" PL ( X; Z) log _ll,u,G xEG PL (X; Z) HU,G H,U,G pR (x,z) log ___ R P 4) II,UGi xEG PR,' Z)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "where E stands for summation over those instances in which the divisor is not zero. We express these terms simply as DL (z) and pE(x, z), if the suffixes are clear from the context. Here, Til LI ' U'G (x, z) is the probability distribution that a category x precedes the parent category v. Based on the requirement on the category set described in Section 2.3, the parent category v is not an element of G. It means f G ILU (V) = 0. Therefore, we use the summation of the frequency of the sibling category y E C(v) in place of the frequency of the parent category v.", |
| "cite_spans": [ |
| { |
| "start": 188, |
| "end": 207, |
| "text": "Til LI ' U'G (x, z)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "(x; z) _11,U, pR G (x; z) EYEC(P(z)) f r d'U (x y) EYEC(P(z)) (y) E YEC(P(z)) (Y \u2022 x)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "EyEc(P(z)) g' u (y) Now we define the difference, with respect to environments, between the category z and its parent category v, namely D ll 'U'G (Z) as the larger one of D L I-Lu'G (z) and DR IT ' u'G (z) . This is because if the environments on either side of category z differ sufficiently from the parent's environment, the category z and its parent category v are different with respect to these environments.", |
| "cite_spans": [ |
| { |
| "start": 16, |
| "end": 19, |
| "text": "(y)", |
| "ref_id": null |
| }, |
| { |
| "start": 194, |
| "end": 206, |
| "text": "IT ' u'G (z)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "Furthermore, we assign a value3 of -1 to D ll ' U'G (Z) as shown below, when the frequency of category z is less than a certain threshold fT T . This is because we cannot judge whether the category must be subdivided or not when the frequency is very low.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "D'' (z) J -1 fGHP(z) < .fT max(DE(z), DR (z)) , otherwise", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "Finally, we define a subdivision score Ell ' U'G (V) which indicates how significant it is to subdivide category v. If at least one child category is sufficiently different from the parent category, then the parent category may be subdivided. So, we define the subdivision score En ' tj'G (V) as the largest difference from among the differences between category v and each of its children.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "Ell,u,G (v) 11,U,GuC(v)-{v} max D z ec (v)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "It means that the subdivision score of a category v is the difference of the child which is most different from v. The subdivision score E\" (v) is 0, when a category v must not be subdivided. If a category v should be subdivided, E-(v) has a larger value. And, the subdivision score E-(v) has a value of -1, when we can not judge whether a category v must be subdivided or not.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Outline of Generating Algorithm", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "Here, we summarize the algorithm described above.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "(1) G\u00b0 {p} ,t 1 .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "(2) argmax", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "(v). veGt-l-VaU) (3) Et En,u,Gt -I (i)-t (4) if Et < ET then qu (5) Gt Gt-1 U Cebt)", |
| "eq_num": "(6)" |
| } |
| ], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "t t + 1 and goto", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "ET is the subdivision score threshold used to decide when the algorithm must be terminated. Unsimilar subcategories are gathered into a category, if ET is too low. 3This value, which indicates that the reliability of the category is low, need not be -1. We aim to distinguish the value from a divergence, which is nonnegative. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Algorithm", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "We used RWC newspaper corpus tagged with the hierarchical part-of-speech system. The partof-speech system had 509 leaf categories and 208 middle categories (or I Vm I = 509 and I U PI = 208). We used the whole corpus containing 18,672 sentences, called \"whole corpus\", and 10 of its subcorpora. Table 1 shows the size of each of these corpora used in the experiment. We set the frequency threshold fT that specifies the significance of the subdivision score as 100. We also set ET to be 0, because the proper value for ET is unknown and we would like to estimate the proper value with this experiment. Figure 3 shows the change in the subdivision score of the subdivided categories with time. The individual subdivided categories are shown in Tables 2, 3 and 4. Table 2 shows which Tables 3 and 4 show which categories are subdivided with \"whole corpus\". In in Tables 2, 3 and 4, the third column labeled nt indicates the name of the subdivided category at time t. A hyphen stands for a parent-child relationship in their category names. Figure 4 shows the change in the size of .the category set G with time. Figure 3 shows that all the subcorpora, are almost equivalent in the transition of the maximum subdivision score. It also means that, if we use such subcorpora, the generated category set has little noise.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 295, |
| "end": 302, |
| "text": "Table 1", |
| "ref_id": "TABREF1" |
| }, |
| { |
| "start": 602, |
| "end": 610, |
| "text": "Figure 3", |
| "ref_id": null |
| }, |
| { |
| "start": 743, |
| "end": 769, |
| "text": "Tables 2, 3 and 4. Table 2", |
| "ref_id": "TABREF2" |
| }, |
| { |
| "start": 782, |
| "end": 796, |
| "text": "Tables 3 and 4", |
| "ref_id": "TABREF3" |
| }, |
| { |
| "start": 861, |
| "end": 872, |
| "text": "Tables 2, 3", |
| "ref_id": "TABREF2" |
| }, |
| { |
| "start": 1038, |
| "end": 1046, |
| "text": "Figure 4", |
| "ref_id": null |
| }, |
| { |
| "start": 1110, |
| "end": 1118, |
| "text": "Figure 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experiments", |
| "sec_num": "4" |
| }, |
| { |
| "text": "However, comparing with the results from the whole corpus, we notice some differences. As Table 2, Table 3 and Table 4 show, some categories that are subdivided early with the whole corpus are subdivided later in case of subcorpora. For example, p-ADJ 4 , which is subdivided second with the whole corpus, is subdivided at t = 22 with the s01 corpus. We conjecture that the category corresponding to adjectives is subdivided later because of the low frequency of the adjectives. A category has a low subdivision score if its children categories do not occur uniformly, although the frequency of the category may be high.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 90, |
| "end": 118, |
| "text": "Table 2, Table 3 and Table 4", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experiments", |
| "sec_num": "4" |
| }, |
| { |
| "text": "Once a certain category is subdivided, the subdivision scores of its children categories tend to be larger than the subdivision scores of its sibling categories. For example, in Table 2 , because the category p-v is subdivided early, its children categories are subdivided in succeeding steps. In the part-of-speech system we use, the category p-v has many descendant categories. The number of leaf categories which are descendant of p-v is 361. It is more than half of 509, which is the total number of leaf categories. As a result, category p-ADJ is subdivided later. In fact, when the category corresponding to verbs were subdivided, the maximum subdivision shot up, as shown in Figure 3 and Table 3 .", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 178, |
| "end": 185, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| }, |
| { |
| "start": 682, |
| "end": 690, |
| "text": "Figure 3", |
| "ref_id": null |
| }, |
| { |
| "start": 695, |
| "end": 702, |
| "text": "Table 3", |
| "ref_id": "TABREF3" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experiments", |
| "sec_num": "4" |
| }, |
| { |
| "text": "It is difficult to estimate a proper value for the threshold of the subdivision score ET, because the maximum subdivision score often shoots up.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Proper Value for the Threshold of Subdivision Score", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "If we set ET = 2, the algorithm terminates at t = 8, (GI = 83 with the whole corpus. But the largest subdivision score appears at t = 16. If we set ET = 1.2 in order to call the largest subdivision score, the algorithm terminates at t = 34, IGI = 249. The category set at t = 34 seems to be too large.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Proper Value for the Threshold of Subdivision Score", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "In order to solve the problem mentioned in Section 5.2, we are considering improving the generating algorithm. We are considering other methods for separating a category from its parent instead of subdividing a category at each step. Furthermore, we are considering using other part-of-speech systems in which the leaf nodes are morphemes and not parts-of-speech.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Future Work", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "In this work, we aimed to generate a proper category set. But this work lacks an objective evaluation of the aptness of a category set. We have to find a measure of aptness and evaluate the algorithm by this measure. We would like to use a linguistic one.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Future Work", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Using this measure, we have to run experiments that determine two parameters of the generating algorithm. Then, we have to compare our method to other existing methods.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Future Work", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "We think we have to perform more experiments with other data. For example, we would like to do an experiment using a larger corpus or an experiment with corpus in other domains or other languages.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Future Work", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "In this paper, we proposed a method to generate a proper categorization of morphemes given a hierarchical part-of-speech system and a corpus tagged using this part-of-speech system. Then, we ran experiments to test the generating algorithm. As a result, we obtained a moderately proper categorization, and found several candidates for improvement.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "6" |
| }, |
| { |
| "text": "N, NO, PN, RG, NT and SYM stand for noun, number, propernoun, region, nation and symbol, respectively.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "1n other words, the rewritten corpus is obtained by replacing each leaf category x by r'G(x).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "ADJ and v stand for adjective and verb, respectively.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "We would like to thank Bipin Indurkhya for his advice on expression in English.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgements", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Tagset reduction without information loss", |
| "authors": [ |
| { |
| "first": "Thorsten", |
| "middle": [], |
| "last": "Brants", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "Meeting of the Association for Computational Linguistics", |
| "volume": "", |
| "issue": "", |
| "pages": "287--289", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Brants, Thorsten. 1995. Tagset reduction without information loss. In Meeting of the Association for Computational Linguistics, pages 287-289.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "How to evaluate and compare tagsets: a proposal", |
| "authors": [ |
| { |
| "first": "Herve", |
| "middle": [], |
| "last": "Dejean", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "LREC", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dejean, Herve. 2000. How to evaluate and compare tagsets: a proposal. In LREC 2000.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "Category set before subdivision of category v \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \u2022 Figure 2: Category set after subdivision of category v 3.2 Subdivision Score", |
| "type_str": "figure", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF1": { |
| "text": "with the subcorpus \"s01\", and", |
| "type_str": "figure", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF2": { |
| "text": "Figure 4: Size of category set", |
| "type_str": "figure", |
| "uris": null, |
| "num": null |
| }, |
| "TABREF1": { |
| "text": "", |
| "html": null, |
| "num": null, |
| "content": "<table><tr><td/><td colspan=\"2\">: Size of the corpora used in the experiment</td></tr><tr><td colspan=\"3\">name Number of Sentences Number of Morphemes</td></tr><tr><td>sOl</td><td>934</td><td>44544</td></tr><tr><td>s02</td><td>867</td><td>42915</td></tr><tr><td>s03</td><td>853</td><td>44398</td></tr><tr><td>s04</td><td>901</td><td>44273</td></tr><tr><td>s05</td><td>891</td><td>45058</td></tr><tr><td>s06</td><td>946</td><td>44586</td></tr><tr><td>s07</td><td>888</td><td>44418</td></tr><tr><td>s08</td><td>1015</td><td>43811</td></tr><tr><td>s09</td><td>874</td><td>44121</td></tr><tr><td>slO</td><td>885</td><td>43896</td></tr><tr><td>whole</td><td>18672</td><td>888000</td></tr></table>", |
| "type_str": "table" |
| }, |
| "TABREF2": { |
| "text": "Subdivided categories with subcorpus \"s01\"", |
| "html": null, |
| "num": null, |
| "content": "<table><tr><td>t</td><td>Et</td><td>jt</td></tr></table>", |
| "type_str": "table" |
| }, |
| "TABREF3": { |
| "text": "Subdivided categories with whole corpus (1)", |
| "html": null, |
| "num": null, |
| "content": "<table/>", |
| "type_str": "table" |
| }, |
| "TABREF4": { |
| "text": "", |
| "html": null, |
| "num": null, |
| "content": "<table><tr><td>: Subdivided categories with whole corpus (2)</td></tr></table>", |
| "type_str": "table" |
| } |
| } |
| } |
| } |
