{ "paper_id": "O04-2004", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:59:52.144727Z" }, "title": "Mencius: A Chinese Named Entity Recognizer Using the Maximum Entropy-based Hybrid Model", "authors": [ { "first": "Tzong-Han", "middle": [], "last": "Tsai", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Taiwan University", "location": { "settlement": "Taipei", "country": "Taiwan, R.O.C" } }, "email": "thtsai@iis.sinica.edu.tw" }, { "first": "Shih-Hung", "middle": [], "last": "Wu", "suffix": "", "affiliation": {}, "email": "shwu@iis.sinica.edu.tw" }, { "first": "Cheng-Wei", "middle": [], "last": "Lee", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Wei", "middle": [], "last": "Shih", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Wen-Lian", "middle": [], "last": "Hsu", "suffix": "", "affiliation": {}, "email": "hsu@iis.sinica.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper presents a Chinese named entity recognizer (NER): Mencius. It aims to address Chinese NER problems by combining the advantages of rule-based and machine learning (ML) based NER systems. Rule-based NER systems can explicitly encode human comprehension and can be tuned conveniently, while ML-based systems are robust, portable and inexpensive to develop. Our hybrid system incorporates a rule-based knowledge representation and template-matching tool, called InfoMap [Wu et al. 2002], into a maximum entropy (ME) framework. Named entities are represented in InfoMap as templates, which serve as ME features in Mencius. These features are edited manually, and their weights are estimated by the ME framework according to the training data. To understand how word segmentation might influence Chinese NER and the differences between a pure template-based method and our hybrid method, we configure Mencius using four distinct settings. The F-Measures of person names (PER), location names (LOC) and organization names (ORG) of the best configuration in our experiment were respectively 94.3%, 77.8% and 75.3%. From comparing the experiment results obtained using these configurations reveals that hybrid NER Systems always perform better performance in identifying person names. On the other hand, they have a little difficulty identifying location and organization names. Furthermore, using a word segmentation module improves the performance of pure Template-based NER Systems, but, it has little effect on hybrid NER systems.", "pdf_parse": { "paper_id": "O04-2004", "_pdf_hash": "", "abstract": [ { "text": "This paper presents a Chinese named entity recognizer (NER): Mencius. It aims to address Chinese NER problems by combining the advantages of rule-based and machine learning (ML) based NER systems. Rule-based NER systems can explicitly encode human comprehension and can be tuned conveniently, while ML-based systems are robust, portable and inexpensive to develop. Our hybrid system incorporates a rule-based knowledge representation and template-matching tool, called InfoMap [Wu et al. 2002], into a maximum entropy (ME) framework. Named entities are represented in InfoMap as templates, which serve as ME features in Mencius. These features are edited manually, and their weights are estimated by the ME framework according to the training data. To understand how word segmentation might influence Chinese NER and the differences between a pure template-based method and our hybrid method, we configure Mencius using four distinct settings. The F-Measures of person names (PER), location names (LOC) and organization names (ORG) of the best configuration in our experiment were respectively 94.3%, 77.8% and 75.3%. From comparing the experiment results obtained using these configurations reveals that hybrid NER Systems always perform better performance in identifying person names. On the other hand, they have a little difficulty identifying location and organization names. Furthermore, using a word segmentation module improves the performance of pure Template-based NER Systems, but, it has little effect on hybrid NER systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Information Extraction (IE) is the task of extracting information of interest from unconstrained text. IE involves two main tasks: the recognition of named entities, and the recognition of the relationships among these named entities. Named Entity Recognition (NER) involves the identification of proper names in text and classification of them into different types of named entities (e.g., persons, organizations, locations). NER is important not only in IE [Grishman 2002 ] but also in lexical acquisition for the development of robust NLP systems [Coates-Stephens 1992] . Moreover, NER has proven useful for tasks such as document indexing and the maintenance of databases containing identified named entities.", "cite_spans": [ { "start": 459, "end": 473, "text": "[Grishman 2002", "ref_id": null }, { "start": 550, "end": 572, "text": "[Coates-Stephens 1992]", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "During the last decade, NER has drawn much attention at Message Understanding Conferences (MUC) [Chinchor 1995a ] [Chinchor 1998a ]. Both rule-based and machine learning NER systems have had some success. Traditional rule-based approaches have used manually constructed finite state patterns, which match text against a sequence of words. Such systems (like the University of Edinburgh's LTG [Mikheev et al. 1998 ]) do not need very much training data and can encode expert human knowledge. However, rule-based approaches lack robustness and portability. Each new source of text requires significant tweaking of the rules to maintain optimal performance, and the maintenance costs can be quite steep.", "cite_spans": [ { "start": 96, "end": 111, "text": "[Chinchor 1995a", "ref_id": null }, { "start": 114, "end": 129, "text": "[Chinchor 1998a", "ref_id": null }, { "start": 392, "end": 412, "text": "[Mikheev et al. 1998", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Another popular approach in NER is machine-learning (ML). ML is attractive in that it is more portable and less expensive to maintain. Representative ML approaches used in NER are HMM (BBN's IdentiFinder in [Miller et al. 1998 ] [Bikel et al. 1999] and Maximum Entropy (ME) (New York University's MEME in [Borthwick et al. 1998 ] [Borthwick 1999] ). However, ML systems are relatively inexpensive to develop, and the outputs of these systems are difficult to interpret. In addition, it is difficult to improve the system performance through error analysis. The performance of an ML system can be very poor when the amount of training data is insufficient. Furthermore, the performance of ML systems is worse than that of rule-based ones by about 2%, as revealed at MUC-6 [Chinchor 1995b ] and MUC-7 [Chinchor 1998b ]. This might be due to the fact that current ML approaches can not capture non-parametric factors as effectively as human experts who handcraft the rules. Nonetheless, ML approaches do provide important statistical information that is unattainable by human experts. Currently, the F-measures of English rule-based and ML NER systems are in the range of 85% ~ 94%, based on MUC-7 data [Chinchor 1998c ]. This is higher than the average performance of Chinese NER systems, which ranges from 79% to 86% ].", "cite_spans": [ { "start": 207, "end": 226, "text": "[Miller et al. 1998", "ref_id": "BIBREF15" }, { "start": 229, "end": 248, "text": "[Bikel et al. 1999]", "ref_id": "BIBREF1" }, { "start": 305, "end": 327, "text": "[Borthwick et al. 1998", "ref_id": "BIBREF2" }, { "start": 330, "end": 346, "text": "[Borthwick 1999]", "ref_id": "BIBREF3" }, { "start": 771, "end": 786, "text": "[Chinchor 1995b", "ref_id": null }, { "start": 799, "end": 814, "text": "[Chinchor 1998b", "ref_id": null }, { "start": 1200, "end": 1215, "text": "[Chinchor 1998c", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In this paper, we address the problem of Chinese NER. In Chinese sentences, there are no spaces between words, no capital letters to denote proper names, no sentence breaks, and, worst of all, no standard definition of \"words.\" As a result, word boundaries cannot, at times, be discerned without a context. In addition, the length of a named entity is longer on average than Maximum Entropy-based Hybrid Model that of an English one; thus, the complexity of a Chinese NER system is greater.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Previous works [Chen et al. 1998 ] [Yu et al. 1998 ] [Sun et al., 2002] on Chinese NER have relied on the word segmentation module. However, an error in the word segmentation step might lead to errors in NER results. Therefore, we want to compare the results of NER with/without performing word segmentation. Without word segmentation, a character-based tagger is used, which treats each character as a token and combines the tagged outcomes of contiguous characters to form an NER output. With word segmentation, we treat each word or character as a token, and combine the tagged outcomes of contiguous tokens to form an NER output. Borthwick [1999] used an ME framework to integrate many NLP resources, including previous systems such as Proteus, a POS tagger. Mencius, the Chinese named entity recognizer presented here, incorporates a rule-based knowledge representation and a template-matching tool, called InfoMap [Wu et al. 2002] , into a maximum entropy (ME) framework. Named entities are represented in InfoMap as templates, which serve as ME features in Mencius. These features are edited manually, and their weights are estimated by means of the ME framework according to the training data. This paper is organized as follows. Section 2 provides the ME-based framework for NER. Section 3 describes features and how they are represented in our knowledge representation system, InfoMap. The data set and experimental results are discussed in section 4. Section 5 gives our conclusions and possible extensions of the current work.", "cite_spans": [ { "start": 15, "end": 32, "text": "[Chen et al. 1998", "ref_id": "BIBREF9" }, { "start": 35, "end": 50, "text": "[Yu et al. 1998", "ref_id": "BIBREF19" }, { "start": 53, "end": 71, "text": "[Sun et al., 2002]", "ref_id": "BIBREF16" }, { "start": 634, "end": 650, "text": "Borthwick [1999]", "ref_id": "BIBREF3" }, { "start": 920, "end": 936, "text": "[Wu et al. 2002]", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "For our purpose, we regard each character as a token. Consider a test corpus and a set of n named entity categories. Since a named entity can have more than one token, we associate the following two tags with each category x: x_begin and x_continue. In addition, we use the tag unknown to indicate that a token is not part of a named entity. The NER problem can then be rephrased as the problem of assigning one of 2n + 1 tags to each token. In Mencius, there are 3 named entity categories and 7 tags: person_begin, person_continue, location_begin, location_continue, organization_begin, organization_continue and unknown. For example, the phrase [\uf9e1 \u9060 \u54f2 \u5728 \u9ad8 \u96c4 \u5e02] (Lee, Yuan Tseh in Kaohsiung City) could be tagged as _begin, [person person_continue, person_continue, unknown, location_begin, location_continue, location_continue] .", "cite_spans": [ { "start": 725, "end": 829, "text": "[person person_continue, person_continue, unknown, location_begin, location_continue, location_continue]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy-Based NER Framework", "sec_num": "2." }, { "text": "ME is a flexible statistical model which assigns an outcome for each token based on its history and features. Outcome space is comprised of the seven Mencius tags for an ME formulation of NER. ME computes the probability p(o|h) for any o from the space of all possible outcomes O, and for every h from the space of all possible histories H. A history is composed of all the conditioning data that enable one to assign probabilities to the space of outcomes. In NER, history can be viewed as consisting of the all information derivable from the test corpus relavant to the current token.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "The computation of p(o|h) in ME depends on a set of binary-valued features, which are helpful in making a prediction about the outcome. For instance, one of our features is as follows: when the current character is a known surname, it is likely to be the leading character of a person name. More formally, we can represent this feature as Here, Current-Char-Surname(h) is a binary function that returns the value true if the current character of the history h is in the surname list.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "Given a set of features and a training corpus, the ME estimation process produces a model in which every feature f i has a weight \u03b1 i . This allows us to compute the conditional probability as follows [Berger et al. 1996] :", "cite_spans": [ { "start": 201, "end": 221, "text": "[Berger et al. 1996]", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "\u220f = i o h f i i h Z h o p ) , ( ) ( 1 ) | ( \u03b1 . (2)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "Intuitively, the probability is the multiplication of the weights of active features (i.e., those f i (h,o) = 1). The weight \u03b1 i is estimated by means of a procedure called Generalized Iterative Scaling (GIS) [Darroch et al. 1972] . This is an iterative method that improves estimation of the weights at each iteration. The ME estimation technique guarantees that for every feature f i , the expected value of\u03b1 i equals the empirical expectation of\u03b1 i in the training corpus.", "cite_spans": [ { "start": 209, "end": 230, "text": "[Darroch et al. 1972]", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "As Borthwick [1999] remarked, ME allows the modeler to concentrate on finding the features that characterize the problem while letting the ME estimation routine deal with assigning relative weights to the features.", "cite_spans": [ { "start": 3, "end": 19, "text": "Borthwick [1999]", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Maximum Entropy", "sec_num": "2.1" }, { "text": "After an ME model has been trained and the proper weight \u03b1 i has been assigned to each feature f i , decoding (i.e., marking up) a new piece of text becomes a simple task. First, Mencius tokenizes the text and preprocesses the testing sentence. Then for each token, it checks which Maximum Entropy-based Hybrid Model features are active and combines the\u03b1 i of the active features according to equation 2. Finally, a Viterbi search is run to find the highest probability path through the lattice of conditional probabilities that does not produce any invalid tag sequences (for instance, the sequence [person_begin, location_continue] is invalid). Further details on Viterbi search can be found in [Viterbi 1967 ].", "cite_spans": [ { "start": 697, "end": 710, "text": "[Viterbi 1967", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "2.2" }, { "text": "We divide features that can be used to recognize named entities into four categories according to whether they are external or not and whether they are category dependent or not. McDonald defined internal and external features in [McDonald 1996] . Internal evidence is found within the entity, while external evidence is gathered from its context. We use category-independent features to distinguish named entities from non-named entities (e.g., first-character-of--a-sentence, capital-letter, out-of-vocabulary), and use category-dependent features to distinguish between different named entity categories (for example, surname and given name lists are used to recognize person names). However, to simplify our design, we only use internal features that are category-dependent in this paper.", "cite_spans": [ { "start": 230, "end": 245, "text": "[McDonald 1996]", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3." }, { "text": "To the calculate values of location features and organization features, Mencius uses InfoMap. InfoMap is our knowledge representation and template matching tool, which represents location or organization names as templates. An input string (sentence) is first matched to one or more location or organization templates by InfoMap and then passed to Mencius; there, it is assigned feature values which further distinguish which named entity category it falls into.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "InfoMap -Our Knowledge Representation System", "sec_num": "3.1" }, { "text": "InfoMap is a hierarchical knowledge representation scheme, consisting of several domains, each with a tree-like taxonomy. The basic units of information in InfoMap are called generic nodes, which represent concepts, and function nodes, which represent the relationships among the generic nodes of one specific domain. In addition, generic nodes can also contain cross references to other nodes to avoid needless repetition.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Knowledge Representation Scheme in InfoMap", "sec_num": "3.1.1" }, { "text": "In Mencius, we apply the geographical taxonomy of InfoMap called GeoMap. Our location and organization templates refer to generic nodes in Geomap. As shown in Figure 1 , GeoMap has three sub-domains: World, Mainland China, and Taiwan. Under the sub-domain Taiwan, there are four attributes: Cities, Parks, Counties and City Districts. Moreover, these attributes can be further divided; for example, Counties can be divided into individual counties:", "cite_spans": [], "ref_spans": [ { "start": 159, "end": 167, "text": "Figure 1", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Knowledge Representation Scheme in InfoMap", "sec_num": "3.1.1" }, { "text": "Taipei County, Taoyuan County, etc. In InfoMap, we refer to generic nodes (or concept node) by means of paths. A path of generic nodes consists of all the node names from the root of the domain to the specific generic node, where function nodes are omitted. The node names are separated by periods. For example, the path for the \"Taipei County\" node is \"GeoMap.Counties.Taipei County.\" ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Knowledge Representation Scheme in InfoMap", "sec_num": "3.1.1" }, { "text": "In InfoMap, text templates are stored in generic nodes. Templates can consist of character strings, wildcards (see $$ in Table 1) , and references to other generic nodes in InfoMap. For example, the template [ \u901a \u7528 \u5730 \uf9e4 . \u53f0 \u7063 . \u7e23 ]:$$(2..4): \u5c40 ( [GeoMap. Taiwan. Counties]: $$(2..4):Department ) can be used to recognize county level governmental departments in Taiwan. The syntax used in InfoMap templates are shown in Table 1 . The first part of our sample template shown above (enclosed by \"[]\") is a path that refers to the generic node \"Counties.\" The second element is a wildcard, ($$) which must be 2 to 4 characters in length. The third element is a specified character \"\u5c40\" (Department). ", "cite_spans": [], "ref_spans": [ { "start": 121, "end": 129, "text": "Table 1)", "ref_id": "TABREF0" }, { "start": 418, "end": 425, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "InfoMap Templates", "sec_num": "3.1.2" }, { "text": "Recall that category-dependent features are used to distinguish among different named entity categories.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Category-Dependent Internal Features", "sec_num": "3.2" }, { "text": "Mencius only deals with a surname plus a first name (usually composed of two characters), for example, \u9673\u6c34\u6241 (Chen Shui-bian). There are various other ways to identify a person in a sentence, such as \u9673\u5148\u751f (Mr. Chen) and\uf934\u9673 (Old Chen), which have not been incorporated into the current system. Furthermore, we do not target transliterated names, such as \u5e03\u5e0c (Bush), since they do not follow Chinese name composition rules. We use a table of frequently occurring names to process our candidate test data. If a character and its context (history) correspond to a feature condition, the value of the current character for that feature will be set to 1. Feature conditions, examples and explanations for each feature are shown in Table 2 . In the feature condition column, c -1 , c 0 , and c 1 represent the preceding character, the current character, and the following character, respectively.", "cite_spans": [], "ref_spans": [ { "start": 720, "end": 727, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Features for Recognizing Person Names", "sec_num": "3.2.1" }, { "text": "Current-Char-Person-Surname: This feature is set to 1 if c 0 c 1 c 2 or c 0 c 1 is in the person name database. For example, in the case of c 0 c 1 c 2 = \u9673\u6c34\u6241, the feature Current-Char-Person-Surname for \u9673 is active since c 0 and its following characters c 1 c 2 satisfy the feature condition.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Person Names", "sec_num": "3.2.1" }, { "text": "Current-Char-Person-Given-Name: This feature is set to 1 if c -2 c -1 c 0 , c -1 c 0 , or c -1 c 0 c 1 is in the person name database.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Person Names", "sec_num": "3.2.1" }, { "text": "Current-Char-Surname: This feature is set to 1 if c 0 is in the top 300 popular surname list. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Person Names", "sec_num": "3.2.1" }, { "text": "In general, locations are divided into four types: administrative division, public area (park, airport, or port), landmark (road, road section, cross section or address), and landform (mountain, river, sea, or ocean Two features are associated with each InfoMap template category x (e.g., location and organization). The first is Current-Char-InfoMap-x-Begin, which is set to 1 for the first character of a matched string and set to 0 for the remaining characters. The other is Current-Char-InfoMap-x-Continue, which is set to 1 for all the characters of matched string except for the first character and set to 0 for the first character. The intuition behind this is as follows: InfoMap can be used to help ME detect which character in a sentence is the first character of the location name and which characters are the remaining characters of a location name. That is, Current-Char-InfoMap-x-Begin is helpful for determining which character should be tagged as x_begin, while Current-Char-InfoMap-x-Continue is helpful for determining which character should be tagged as x_continue if we build an InfoMap template for that category x. The two features associated with x category are shown below: ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Location Names", "sec_num": "3.2.2" }, { "text": "\u23a9 \u23a8 \u23a7 = = =", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Location Names", "sec_num": "3.2.2" }, { "text": "When recognizing a location name in a sentence, we test if any location templates match the sentence. If several matched templates overlap, we select the longest matched one. As mentioned above, the feature Current-Character-InfoMap-Location-Begin of the first character of the matched string is set to 1 while the feature Current-Character-InfoMap-Location--Continue of the remaining characters of the matched string is set to 1. Table 3 shows the necessary conditions for each organization feature and gives examples of matched data. ", "cite_spans": [], "ref_spans": [ { "start": 431, "end": 438, "text": "Table 3", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Features for Recognizing Location Names", "sec_num": "3.2.2" }, { "text": "Organizations include named corporate, governmental, or other organizational entities. The difficulty in recognizing an organization name is that it usually begins with a location name, such as \u53f0\uf963\u5e02\u5730\u6aa2\u7f72 (Taipei District Public Prosecutors Office). Therefore, traditional machine learning NER systems can only identify the location part rather than the full organization name. For example, the system only extracts \u53f0\uf963\u5e02 (Taipei City) from \u53f0\uf963\u5e02 SOGO \u767e\u8ca8\u9031\uf98e\u6176 (Taipei SOGO Department Store Anniversary) rather than \u53f0\uf963\u5e02 SOGO \u767e\u8ca8 (Taipei SOGO Department Store). According to our analysis of the structure of Chinese organization names, they mostly end with a specific keyword or begin with a location name. Therefore, we use those keywords and location names as the boundary markers of organization names. Based on our observation, we categorize organization names into four types according to their boundary markers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features for Recognizing Organization Names", "sec_num": "3.2.3" }, { "text": "The organization names in this category begin with by one or more geographical names and ended by an organization keyword. For example, \u53f0\uf963\u5e02 (Taipei City) is the left boundary marker of \u53f0\uf963\u5e02\u6377\u904b\u516c\u53f8 (Taipei City Rapid Transit Corporation), while an organization keyword, \u516c\u53f8 (Corporation), is the right boundary marker.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Type I: With left and right boundary markers", "sec_num": null }, { "text": "The organization names in this category begin with by one or more than one geographical names, but the organization keyword (e.g., \u516c\u53f8 (Corporation)) is omitted. For example, \u53f0\u7063\u6377\u5b89\u7279 (Giant Taiwan) only contains the left boundary \u53f0\u7063 (Taiwan).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Type II: With a left boundary marker", "sec_num": null }, { "text": "The organization names in this category end with an organization keyword. For example, \u6377\u5b89 \u7279\u516c\u53f8 (Giant Corporation) only contains the right boundary \u516c\u53f8 (Corporation). $$(2..4):Department) to recognize county level government departments in Taiwan. However, in Types II, III, and IV, organization names cannot be recognized by templates. Therefore, the maximum entropy model uses features of characters (from c -2 to c 2) , tags (from t -2 to t 2 ), and organization keywords, e.g., \u516c\u53f8 (Corporation), to find the most likely tag sequences and recognize them.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Type III: With a right boundary marker", "sec_num": null }, { "text": "When a string matches an organization template, the feature Current-Character-InfoMap--Organization-Start of the first character is set to 1. In addition, the feature Current--Character-InfoMap-Organization-Continue of the remaining characters is set to 1. The necessary conditions for each organization feature and examples of matched data are shown in Table 4 . These features are helpful for recognizing organization names.", "cite_spans": [], "ref_spans": [ { "start": 354, "end": 361, "text": "Table 4", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Type III: With a right boundary marker", "sec_num": null }, { "text": "For Chinese NER, the most famous corpus is MET-2 [6]. There are two main differences between our corpus and MET-2: the number of domains and the amount of data. First, MET-2 contains only one domain (Accident), while our corpus, which was collected from the online United Daily News in December 2002 (http://www.udn.com.tw), contains six domains: Local News, Social Affairs, Investment, Politics, Headline News and Business, which provide a greater variety of organization names than a single domain corpus can. The full location names and organization names are comparatively longer, and our corpus contains more location names and addresses at the county level. Therefore, the patterns of location names and organization names are more complex in our corpus.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Sets", "sec_num": "4.1" }, { "text": "Secondly, our corpus is much larger than MET2, which contains 174 Chinese PER, 750 LOC, and 377 ORG. Our corpus contains 1,242 Chinese PER, 954 LOC, and 1,147 ORG in 10,000 sentences (about 126,872 Chinese characters). The statistics of our data are shown in Table 5 . ", "cite_spans": [], "ref_spans": [ { "start": 259, "end": 266, "text": "Table 5", "ref_id": "TABREF6" } ], "eq_spans": [], "section": "Data Sets", "sec_num": "4.1" }, { "text": "To understand how word segmentation might influence Chinese NER and the differences between a pure template-based method and our hybrid method, we configure Mencius using the following four settings: (1) Template-based with Char-based Tokenization (TC), 2Template-based with Word-based Tokenization (TW), (3) Hybrid with Char-based Tokenization (HC), and (4) Hybrid with Word-based Tokenization (HW). Following the standard 10-fold cross-validation method, we tested Mencius with each configuration using the data set mentioned in section 4.1. The following subsections provide details about each configuration and the results obtained.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Results", "sec_num": "4.2" }, { "text": "In this experiment, we regarded each character as a token, and used a person name list and InfoMap templates to recognize all named entities. The number of lexicons in the person name lists and gazetteers was 32000. As shown in Table 6 , the obtained F-Measures of PER, LOC and ORG were 76.2%, 75.4% and 75.1%, respectively. ", "cite_spans": [], "ref_spans": [ { "start": 228, "end": 235, "text": "Table 6", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Template-based with Char-based Tokenization (TC)", "sec_num": "4.2.1" }, { "text": "In this experiment, we used a word segmentation module based on the 100,000-word CKIP Traditional Chinese dictionary to split sentences into tokens. This module combines forward and backward longest matching algorithms in the following way: if the segmentation results of the two algorithms agree in certain substrings, this module outputs tokens in those substrings. While in the part which the segmentation results of the two algorithms differ, this module skips word tokens and only outputs character tokens. In the previous test, 98% of the word tokens were valid words. Then, we used person name lists and InfoMap templates to recognize all the named entities. The number of lexicons in the person name lists and gazetteers was 32,000. As shown in Table 6 , the obtained F-Measures of PER, LOC and ORG were 89.0%, 74.1% and 71.6%, respectively. ", "cite_spans": [], "ref_spans": [ { "start": 753, "end": 760, "text": "Table 6", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Template-based with Word-based Tokenization (TW)", "sec_num": "4.2.2" }, { "text": "In this experiment, we regarded each character as a token without performing any word segmentation. We then integrated person name lists, location templates, and organization templates into a Maximum-Entropy-Based framework. As shown in Table 8 , the obtained F-Measures of PER, LOC and ORG were 94.3%, 77.8% and 75.3%, respectively. ", "cite_spans": [], "ref_spans": [ { "start": 237, "end": 244, "text": "Table 8", "ref_id": "TABREF9" } ], "eq_spans": [], "section": "Hybrid with Char-based Tokenization (HC)", "sec_num": "4.2.3" }, { "text": "In this experiment, we used the same word segmentation module described in section 4.2.2 to split sentences into tokens. Then, we integrated person name lists, location templates, and organization templates into a Maximum-Entropy-Based framework. As shown in Table 9 , the obtained F-Measures of PER, LOC and ORG were 95.9%, 73.4% and 76.1%, respectively. ", "cite_spans": [], "ref_spans": [ { "start": 259, "end": 266, "text": "Table 9", "ref_id": "TABREF10" } ], "eq_spans": [], "section": "Hybrid System with Word-based Tokenization (HW)", "sec_num": "4.2.4" }, { "text": "We observed that TW achieved much higher precision than TC in PER. When word segmentation is not performed, some trigrams and quadgrams may falsely appear to be person names. Take the sentence \"\u65b0\u53e4\u5178\u4e3b\u7fa9\" for example. TC would extract \"\u53e4\u5178\u4e3b\" as a person", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TC versus TW", "sec_num": null }, { "text": "Maximum Entropy-based Hybrid Model name since \"\u53e4\u5178\u4e3b\" matches our family-name trigram template. However, in TW, thanks to word segmentation, \"\u53e4\u5178\" and \"\u4e3b\u7fa9\" would be marked as tokens first and would not match the family-name trigram template.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TC versus TW", "sec_num": null }, { "text": "We observed that HW achieved similar precision to that of HC in all three NE categories. HW also achieved recall rates similar to those achieved by HC with PER and ORG NEs. In the case of PER NEs, this is because the length of person names is 2 to 4 characters. Therefore, a five-character long window (-2 to +2) is sufficient to recognize a person name. As far as recognizing LOC NEs is concerned, HW's recall rate was worse than HC's. This is because the word segmentation module marks occupational titles as tokens, for example: \"\u53f0\uf963\u5e02\u9577\". HW cannot extract the LOC NE \"\u53f0\uf963\u5e02\" from \"\u53f0\uf963\u5e02\u9577\" because it has already been defined as a token. To recognize LOC and ORG NEs, we need higher-level features and more external features. Since Mencius lacks these kinds of features, HW doesn't achieve significantly better performance than HC.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "HC versus HW", "sec_num": null }, { "text": "We observed that in PER, HC achieved much higher precision than TC, while in LOC and ORG, HC performed slightly better than TC. This is because most of the key features for identifying a person name are close to the person name, or inside the personal name. Take the sentence \"\uf9f7 \u5373\uf99a\u7d61\u6d77\u9dd7\u76f4\u5347\u6a5f\" as an example; when we wish to determine whether \"\uf99a\u7d61\u6d77\" is a person name, we can see that \"\uf9f7\u5373\" seldom appears before a person name, and that \"\u9dd7\" seldom appears after a person name. In HC, ME can use this information to determine that \"\uf99a\u7d61\u6d77\" is not a person name, but to recognize a location name and an organization name, we need wider context and features, such as sentence analysis or shallow parsing. Take \"\u5982\u99ac\u516c\u3001\u4e03\u7f8e\u3001\u671b\u5b89\u3001 \uf91f\u5dbc\u3001\uf93d\u5cf6\u3001\u99ac\u7956\u548c\uf90a\u9580\u7b49\uf9ea\u5cf6\u70ba\u7ba1\u5236\u822a\u7dda\" as an example; the two preceding characters are \"\u7f8e\"and \"\u3001\", and the two following characters are \"\u3001\"and \"\uf91f\". ME cannot use this information to identify a location name.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TC versus HC", "sec_num": null }, { "text": "We observed that HW achieved better precision than TW in identifying personal names. This is because in HW, ME can use context information to filter some trigrams and 4 grams, which are not personal names. Take \" \u738b \uf90a \u5e73 \u548c \u5176 \u4ed6 \u59d4 \u54e1 \" as an example; it matches the double-family-name quadgram template because \" \u738b \" and \" \uf90a \" are both family names. However, \"\u738b\uf90a\u5e73\" is the correct person name. In HW, ME can use the information that \"\u738b\uf90a \u5e73\" has appeared in the training corpus and been tagged as a PER NE to identify the person name \"\u738b\uf90a\u5e73\" in a sentence. We also observed that HW achieved better recall than TW in identifying person names. This is because in HW, ME can use the information that bigram personal names are tagged as PER NEs from the training data, but TW cannot because we don't have bigram-person-name templates. In addition, some person names are in the dictionary, so some tokens are person names. Take \"\u9673\u5efa\u4ec1 \u7684 \u4f5c\u70ba\" as an example. Although the token \"\u9673 \u5efa\u4ec1\" cannot match any person name template, in HW, ME can use context information and training data to recognize \"\u9673\u5efa\u4ec1\". To identify location names, ME needs a wider context to detect location names, so HW's recall is worse than TW's. However, ME can filter out some unreasonable trigrams, such as \"\u9ec3\u69ae\u6751\", because it matches a location name template $$(2..3): \u6751 , which represents a village in Taiwan. Therefore, ME achieves bigger precision in identifying location names.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TW versus HW", "sec_num": null }, { "text": "In this paper, we have presented a Chinese NER system, called Mencius. We configured Mencius according to the following settings of to analyze the effects of using a Maximum Entropy-based Framework and a word segmentation module: (1) Template-based with Char-based Tokenization (TC), (2) Template-based with Word-based Tokenization (TW), (3) Hybrid with Char-based Tokenization (HC), and (4) Hybrid with Word-based Tokenization (HW). The experimental results showed that whether a character or a word was taken as a token, the hybrid NER System always performed better in identifying person names. However,this had little effect on the identification of location and organization names. This is because the context information around a location name or an organization name is more complex than that around a person name. In addition, using a word segmentation module improved the performance of the pure Template-based NER System. However, it had little effect with the hybrid NER systems. The current version of Mencius lacks sentence parsing templates and shallow parsing tools to handle such complex information. We will add these functions in the future.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "5." } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "A maximum entropy approach to natural language processing", "authors": [ { "first": "A", "middle": [], "last": "Berger", "suffix": "" }, { "first": "S", "middle": [ "A" ], "last": "Della Pietra", "suffix": "" }, { "first": "Della", "middle": [], "last": "Pietra", "suffix": "" }, { "first": "V", "middle": [ "J" ], "last": "", "suffix": "" } ], "year": 1996, "venue": "Computational Linguistics", "volume": "22", "issue": "", "pages": "39--71", "other_ids": {}, "num": null, "urls": [], "raw_text": "Berger, A., Della Pietra, S. A., and Della Pietra, V. J., \"A maximum entropy approach to natural language processing,\" Computational Linguistics, vol. 22, pp. 39-71, 1996.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "An Algorithm that Learns What's in a Name", "authors": [ { "first": "D", "middle": [], "last": "Bikel", "suffix": "" }, { "first": "R", "middle": [], "last": "Schwartz", "suffix": "" }, { "first": "R", "middle": [], "last": "Weischedel", "suffix": "" } ], "year": 1999, "venue": "Machine Learning", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bikel, D., Schwartz, R., and Weischedel, R., \"An Algorithm that Learns What's in a Name,\" Machine Learning, 1999.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "NYU: Description of the MENE Named Entity System as Used in MUC-7", "authors": [ { "first": "A", "middle": [], "last": "Borthwick", "suffix": "" }, { "first": "J", "middle": [], "last": "Sterling", "suffix": "" }, { "first": "E", "middle": [], "last": "Agichtein", "suffix": "" }, { "first": "R", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Borthwick, A., Sterling J., Agichtein, E., and Grishman, R., \"NYU: Description of the MENE Named Entity System as Used in MUC-7,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "A Maximum Entropy Approach to Named Entity Recognition", "authors": [ { "first": "A", "middle": [], "last": "Borthwick", "suffix": "" } ], "year": 1999, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Borthwick, A., \"A Maximum Entropy Approach to Named Entity Recognition,\" New York University, 1999.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "MUC-6 Named Entity Task Definition (Version 2.1),\" presented at the 6th Message Understanding Conference", "authors": [ { "first": "N", "middle": [], "last": "Chinchor", "suffix": "" } ], "year": 1995, "venue": "Maximum Entropy-based Hybrid Model", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinchor, N., \"MUC-6 Named Entity Task Definition (Version 2.1),\" presented at the 6th Message Understanding Conference, Columbia, Maryland, 1995. Maximum Entropy-based Hybrid Model", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Statistical Significance of MUC-6 Results", "authors": [ { "first": "N", "middle": [], "last": "Chinchor", "suffix": "" } ], "year": 1995, "venue": "presented at the 6th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinchor, N., \"Statistical Significance of MUC-6 Results,\" presented at the 6th Message Understanding Conference, Columbia, Maryland, 1995.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "MUC-7 Named Entity Task Definition (Version 3.5),\" presented at the 7th Message Understanding Conference", "authors": [ { "first": "N", "middle": [], "last": "Chinchor", "suffix": "" } ], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinchor, N., \"MUC-7 Named Entity Task Definition (Version 3.5),\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Statistical Significance of MUC-7 Results", "authors": [ { "first": "N", "middle": [], "last": "Chinchor", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinchor, N., \"Statistical Significance of MUC-7 Results,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "MUC-7 Test Score Reports for all Participants and all Tasks", "authors": [ { "first": "N", "middle": [], "last": "Chinchor", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chinchor, N., \"MUC-7 Test Score Reports for all Participants and all Tasks,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Description of the NTU System Used for MET2", "authors": [ { "first": "H", "middle": [ "H" ], "last": "Chen", "suffix": "" }, { "first": "Y", "middle": [ "W" ], "last": "Ding", "suffix": "" }, { "first": "S", "middle": [ "C" ], "last": "Tsai", "suffix": "" }, { "first": "G", "middle": [ "W" ], "last": "Bian", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, H. H., Ding, Y. W., Tsai, S. C., and Bian, G. W., \"Description of the NTU System Used for MET2,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "The Analysis and Acquisition of Proper Names for Robust Text Understanding", "authors": [ { "first": "S", "middle": [], "last": "Coates-Stephens", "suffix": "" } ], "year": 1992, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Coates-Stephens, S., \"The Analysis and Acquisition of Proper Names for Robust Text Understanding,\" in Dept. of Computer Science. London: City University, 1992.", "links": null }, "BIBREF11": { "ref_id": "b11", "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 Mathematicl Statistics", "volume": "43", "issue": "", "pages": "1470--1480", "other_ids": {}, "num": null, "urls": [], "raw_text": "Darroch, J. N. and Ratcliff, D., \"Generalized iterative scaling for log-linear models,\" Annals of Mathematicl Statistics, vol. 43, pp. 1470-1480, 1972.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "in Information Extraction: A Multidisciplinary Approach to an Emerging Information Technology", "authors": [ { "first": "R", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "10--26", "other_ids": {}, "num": null, "urls": [], "raw_text": "Grishman, R., \"Information Extraction: Techniques and Challenges,\" in Information Extraction: A Multidisciplinary Approach to an Emerging Information Technology, J. G. Carbonell, Ed. Frascati, Italy: Springer, 1997, pp. 10-26.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Internal and External Evidence in the Identification and Semantic Categorization of Proper Names", "authors": [ { "first": "D", "middle": [], "last": "Mcdonald", "suffix": "" } ], "year": 1996, "venue": "Corpus Processing for Lexical Acquisition", "volume": "", "issue": "", "pages": "21--39", "other_ids": {}, "num": null, "urls": [], "raw_text": "McDonald, D., \"Internal and External Evidence in the Identification and Semantic Categorization of Proper Names,\" in Corpus Processing for Lexical Acquisition, J. Pustejovsky, Ed. Cambridge, MA: MIT Press, 1996, pp. 21-39.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Description of the LTG System Used for MUC-7", "authors": [ { "first": "A", "middle": [], "last": "Mikheev", "suffix": "" }, { "first": "C", "middle": [], "last": "Grover", "suffix": "" }, { "first": "M", "middle": [], "last": "Moensk", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mikheev, A., Grover, C., and Moensk, M., \"Description of the LTG System Used for MUC-7,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "BBN: Description of the SIFT System as Used for MUC-7", "authors": [ { "first": "S", "middle": [], "last": "Miller", "suffix": "" }, { "first": "M", "middle": [], "last": "Crystal", "suffix": "" }, { "first": "H", "middle": [], "last": "Fox", "suffix": "" }, { "first": "L", "middle": [], "last": "Ramshaw", "suffix": "" }, { "first": "R", "middle": [], "last": "Schwartz", "suffix": "" }, { "first": "R", "middle": [], "last": "Stone", "suffix": "" }, { "first": "R", "middle": [], "last": "Weischedel", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Miller, S., Crystal, M., Fox, H., Ramshaw, L., Schwartz, R., Stone, R., and Weischedel, R., \"BBN: Description of the SIFT System as Used for MUC-7,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Chinese Named Entity Identification Using Class-based Language Model", "authors": [ { "first": "J", "middle": [], "last": "Sun", "suffix": "" }, { "first": "J", "middle": [ "F" ], "last": "Gao", "suffix": "" }, { "first": "L", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "M", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "C", "middle": [ "N" ], "last": "Huang", "suffix": "" } ], "year": 2002, "venue": "presented at the 19th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sun, J., Gao, J. F., Zhang, L., Zhou, M., and Huang, C. N., \"Chinese Named Entity Identification Using Class-based Language Model,\" presented at the 19th International Conference on Computational Linguistics,, 2002.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding Algorithm", "authors": [ { "first": "A", "middle": [ "J" ], "last": "Viterbi", "suffix": "" } ], "year": 1967, "venue": "IEEE Transactions on Information Theory", "volume": "IT", "issue": "", "pages": "260--269", "other_ids": {}, "num": null, "urls": [], "raw_text": "Viterbi, A. J., \"Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding Algorithm,\" IEEE Transactions on Information Theory, vol. IT, pp. 260-269, 1967.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "FAQ-centered Organizational Memory", "authors": [ { "first": "S", "middle": [ "H" ], "last": "Wu", "suffix": "" }, { "first": "M", "middle": [ "Y" ], "last": "Day", "suffix": "" }, { "first": "T", "middle": [ "H" ], "last": "Tsai", "suffix": "" }, { "first": "W", "middle": [ "L" ], "last": "Hsu", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wu, S. H., Day, M. Y., Tsai, T. H., and Hsu, W. L., \"FAQ-centered Organizational Memory,\" in Knowledge Management and Organizational Memories, R. Dieng-Kuntz, Ed. Boston: Kluwer Academic Publishers, 2002.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Description of the Kent Ridge Digital Labs System Used for MUC-7", "authors": [ { "first": "S", "middle": [ "H" ], "last": "Yu", "suffix": "" }, { "first": "S", "middle": [ "H" ], "last": "Bai", "suffix": "" }, { "first": "P", "middle": [], "last": "Wu", "suffix": "" } ], "year": 1998, "venue": "presented at the 7th Message Understanding Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yu, S. H., Bai, S. H., and Wu, P., \"Description of the Kent Ridge Digital Labs System Used for MUC-7,\" presented at the 7th Message Understanding Conference, Fairfax, Virginia, 1998.", "links": null } }, "ref_entries": { "FIGREF1": { "type_str": "figure", "num": null, "uris": null, "text": "A partial view of GeoMap." }, "FIGREF2": { "type_str": "figure", "num": null, "uris": null, "text": "This feature is set to 1 if c 0 c 1 or c -1 c 0 is in the given name database.Current-Char-Freq-Given-Name-Character: (c 0 and c 1 ) or (c -1 and c 0 ) is in the frequently given name character list Current-Char-Speaking-Verb: c 0 or c 0 c 1 or c -1 c 0 is in the speaking verb list. This feature distinguishes a trigram containing a speaking verb, such as \u9673\u6c96\uf96f (Chen Chong said), from a real person name.Current-Char-Title: c 0 or c 0 c 1 or c -1 c 0 is in the title list. This feature distinguishes a trigram containing a title, such as \u9673\u5148\u751f (Mr. Chen), from a real person name." }, "TABREF0": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
Symbol SemanticsExample TemplateSample
Matching String
:Concatenate two stringsA:BAB
$$(m..n) Wildcards (the number ofA:$$(1..2):BACB, ADDB,
characters can be from m to n;ACDB
both m and n have to be
non-negative integers)
[p]A path to a generic node[GeoMap.Taiwan.Counties]Taipei County,
Taoyuan County,
Hsinchu County,
etc.
" }, "TABREF1": { "num": null, "html": null, "text": "", "type_str": "table", "content": "" }, "TABREF2": { "num": null, "html": null, "text": "). An administrative division name usually contains one or more Maximum Entropy-based Hybrid Model location names in a hierarchical order, such as \u5b89\u5927\uf976\uf96d\u591a\uf9d4\u591a\u5e02 (Toronto, Ontario). A public area name is composed of a Region-Name and a Place-Name. However, the Region-Name is usually omitted from news content if it was previously mentioned. For example, \uf9d4\u6566\u6d77\u5fb7\u516c\u5712 (Hyde Park, London) contains the Region-Name \uf9d4\u6566 (London) and the Place-Name \u6d77\u5fb7\u516c\u5712 (Hyde Park). But \"Hyde Park, London\" is usually abbreviated as \"Hyde Park\" within a report. The same rule can be applied to landmark names. A landmark name includes a Region-Name and a Position-Name. In a news article, the Region-Name can be omitted if the Place-Name has been mentioned previously. For example, \u6eab \u54e5 \u83ef \u5e02 \uf90f \u4f2f \u905c \u8857 \u4e94 \u865f (No. 5, Robson St., Vancouver City) will be stated as \uf90f\u4f2f\u905c\u8857\u4e94\u865f (No. 5, Robson St.) later in the report.In Mencius, we build templates to recognize three types of location names. Our administrative division templates contain more than one set of location names in a hierarchical order.For example, the template, [ \u901a\u7528\u5730\uf9e4 . \u53f0\u7063. \u5e02 ]:[ \u901a\u7528\u5730\uf9e4. \u53f0\u7063. \u5404\u5e02\ufa08 \u653f\u5340 ] ([GeoMap.Taiwan.Cities]:[GeoMap.Taiwan.City Districts]) can be used to recognize all city districts in Taiwan. In addition, public area templates contain one set of location names and a set of Place-Name. For example, [ \u901a \u7528 \u5730 \uf9e4 . \u53f0 \u7063 . \u5e02 ]:[ \u901a \u7528 \u5730 \uf9e4 . \u53f0 \u7063 . \u516c \u5712 ] ([GeoMap.Taiwan.Cities]:[GeoMap.Taiwan.Parks]) can be used to recognize all city parks in Taiwan. Landmark templates are built in the same way.For example, [ \u901a \u7528 \u5730 \uf9e4 . \u53f0 \u7063 .", "type_str": "table", "content": "
\u5e02]:$$(2..4):\uf937 ([GeoMap.Taiwan.Cities]:$$(2..4):Road) can be used to recognize roads in
Taiwan.
" }, "TABREF4": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
FeatureFeatureExampleExplanations
Conditions
Current-Char-InfoMap-Location-Beginc 0~cn-1 matches\"\u53f0\"\uf963\u7e23\u677fProbably the
an InfoMap\u6a4b\u5e02leading
locationcharacter of a
template, wherelocation name.
the character
length of the
template is n
Current-Char-InfoMap-Location-Continue c a \u2026c 0 \u2026 . c b\u53f0\"\uf963\"\u7e23\u677fProbably a
matches an\u6a4b\u5e02continuing
InfoMap locationcharacter of a
template, where alocation name.
is a negative
integer and b is a
non-negative
integer
" }, "TABREF5": { "num": null, "html": null, "text": "In this category, both left and right boundaries as above mentioned are omitted, for example, \u6377 \u5b89\u7279 (Giant). The organization names in this category are usually in abbreviated form.In Mencius, we build templates for recognizing Type I organization names. Each organization template begins with a location name in GeoMap and ends with an organization keyword. For example, we can build [\u901a\u7528\u5730\uf9e4.\u53f0\u7063.\u5e02]:$$(2..4):\u5c40([GeoMap.Taiwan.Cities]:", "type_str": "table", "content": "
FeatureFeatureExampleExplanations
Conditions
Current-Char-InfoMap-Organization-Beginc 0~cn-1 matches\"\u53f0\"\uf963\u5e02Probably the
an InfoMap\u6377\u904b\u516c\u53f8leading
organizationcharacter of an
template, whereorganization
the charactername
length of the
template is n
Current-Char-InfoMap-Organization-Continue c a \u2026c 0 \u2026 . c b\u53f0\"\uf963\"\u5e02Probably the
matches an\u6377\u904b\u516c\u53f8leading
InfoMapcharacter of an
organizationorganization
template, wherename
a is a negative
integer and b is
a non-negative
integer
Current-Char-Organization-Keywordc 0 or c 0 c 1 or c -1 c 0\"\u516c\"\u53f8,Probably part
is in the\u516c \"\u53f8\"of an
organizationorganization
keyword listkeyword
Type IV: No boundary marker
" }, "TABREF6": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
DomainNumber of Named EntitiesSize (in characters)
PERLOCORG
Local News841399711835
Social Affairs31028735437719
Investment20633314397
Politics41920923317168
Headline News2677024319938
Business14218618725815
Total12429541147126872
" }, "TABREF7": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
NEP(%)R(%)F(%)
PER64.7792.5976.22
LOC76.4174.4275.40
ORG85.6066.9375.12
Total72.9578.6275.67
" }, "TABREF8": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
NEP(%)R(%)F(%)
PER88.6989.3289.00
LOC76.9271.4474.08
ORG85.6661.4471.55
Total84.1474.7079.14
" }, "TABREF9": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
NEP(%)R(%)F(%)
PER96.9791.7194.27
LOC80.9674.8177.76
ORG87.1666.2275.26
Total89.0578.1883.26
" }, "TABREF10": { "num": null, "html": null, "text": "", "type_str": "table", "content": "
NEP(%)R(%)F(%)
PER98.7493.3195.94
LOC81.4666.7373.36
ORG87.5467.2976.09
Total90.3376.6682.93
" } } } }