| { | |
| "paper_id": "Y07-1036", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T13:46:49.313753Z" | |
| }, | |
| "title": "Acquisition of Named-Entity-Related Relations for Searching *", | |
| "authors": [ | |
| { | |
| "first": "Tri-Thanh", | |
| "middle": [], | |
| "last": "Nguyen", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Japan Advanced Institute of Science and Technology", | |
| "location": { | |
| "addrLine": "1-1", | |
| "postCode": "923-1292", | |
| "settlement": "Asahidai, Ishikawa", | |
| "region": "Nomi", | |
| "country": "Japan" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Akira", | |
| "middle": [], | |
| "last": "Shimazu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Japan Advanced Institute of Science and Technology", | |
| "location": { | |
| "addrLine": "1-1", | |
| "postCode": "923-1292", | |
| "settlement": "Asahidai, Ishikawa", | |
| "region": "Nomi", | |
| "country": "Japan" | |
| } | |
| }, | |
| "email": "shimazu@jaist.ac.jp" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Named entities (NEs) are important in many Natural Language Processing (NLP) applications, and discovering NE-related relations in texts may be beneficial for these applications. This paper proposes a method to extract the ISA relation between a \"named entity\" and its category, and an IS-RELATED-TO relation between the category and its related object. Based on the pattern extraction algorithm \"Person Category Extraction\" (PCE), we extend it for solving our problem. Our experiments on Wall Street Journal (WSJ) corpus show promising results. We also demonstrate a possible application of these relations by utilizing them for semantic search.", | |
| "pdf_parse": { | |
| "paper_id": "Y07-1036", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Named entities (NEs) are important in many Natural Language Processing (NLP) applications, and discovering NE-related relations in texts may be beneficial for these applications. This paper proposes a method to extract the ISA relation between a \"named entity\" and its category, and an IS-RELATED-TO relation between the category and its related object. Based on the pattern extraction algorithm \"Person Category Extraction\" (PCE), we extend it for solving our problem. Our experiments on Wall Street Journal (WSJ) corpus show promising results. We also demonstrate a possible application of these relations by utilizing them for semantic search.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Text documents often contain valuable relations among entities. For example, in the sentence taken from the WSJ corpus:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "There's a generally more positive attitude toward the economy, said Bette Raptapoulos, analyst for Prudential-Bache Securities Inc., \u2026 (1) there are relations: \"Bette Raptapoulos\" is-a analyst, and analyst for \"Prudential-Bache Securities Inc.\" Such relations may be beneficial in many NLP applications, such as for answering Who and List questions, e.g., \"Who is Bette Raptapoulos?'' or \"Give me the list of analyst for Prudential-Bache Securities Inc.\" Relations in text documents can be extracted by pattern extraction as in (Brin 1998) . (Brin 1998) presented Dual Iterative Pattern Relation Extraction (DIPRE), and used DIPRE to extract \u2329author, title\u232a tuples describing the relation: the author of the book title is author. Based on Brin's model, (Agichtein and Gravano 2000) presented the Snowball system to extract \u2329organization, location\u232a tuples indicating that the headquarters of organization is in location. (Nguyen and Shimazu 2007) developed the PCE system for extracting \u2329person, category\u232a tuples describing that the person is-a category.", | |
| "cite_spans": [ | |
| { | |
| "start": 421, | |
| "end": 454, | |
| "text": "Prudential-Bache Securities Inc.\"", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 528, | |
| "end": 539, | |
| "text": "(Brin 1998)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 542, | |
| "end": 553, | |
| "text": "(Brin 1998)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 753, | |
| "end": 781, | |
| "text": "(Agichtein and Gravano 2000)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 920, | |
| "end": 945, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "This study proposes to automatically extract quadruples \u2329ne, category, related-to, object\u232a describing that the named entity ne ISA category, and the category IS-RELATED-TO object. We call such relations \"named-entity is-a category\" relation, and \"category related-to object\" relation. The related-to and object of a quadruple may be null. We extend PCE algorithm to extract quadruples, and build a semantic search system to utilize extracted quadruples for answering some types of questions. The remainder of this paper is organized as follows: Section 2 summarises some related work. Section 3 describes the original PCE algorithm and our extraction model. Section 4 gives a possible application of the extracted quadruples; Section 5 presents experiments and evaluation; Conclusions are given in the last section. (Brin 1998) presented the DIPRE algorithm for extracting relations, and used DIPRE to extract (author, title) tuples having the relation: the author of the book title is author. Starting with a small number of (author, title) seed tuples, DIPRE finds the occurrences of tuples in order to generate new patterns. New patterns are, again, used to extract further (author, title) tuples. The DIPRE algorithm is graphically depicted in Figure 1 . Based on DIPRE, (Agichtein and Gravano 2000) introduced another method of generating new patterns, and developed the Snowball system for extracting (organization, location) tuples expressing the relation: the headquarters of organization is in location.", | |
| "cite_spans": [ | |
| { | |
| "start": 816, | |
| "end": 827, | |
| "text": "(Brin 1998)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 1275, | |
| "end": 1303, | |
| "text": "(Agichtein and Gravano 2000)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1248, | |
| "end": 1256, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Current Named Entity Recognition (NER) systems often operate based on a predefined set of named entity classes, and assign a unique class to a discovered named entity (Chieu and Tou 2003) . This is not natural, since the potential classes of named entity is large, and a named entity may belong to more than one class. For example, a person named entity may be both \"executive vice president\" and \"chief financial officer\" as expressed in the sentence: \"Daniel Akerson, executive vice president and chief financial officer, said MCI's growth is being fueled by \u2026\" With the purpose of extending the number of named entity classes, (Nguyen and Shimazu 2007) proposed the \"Person Category Extraction\" (PCE) algorithm to automatically extract fine-grained categories of person Named Entities from text corpora. Based on DPIRE, (Nguyen and Shimazu 2007) introduced new types of patterns based on part-of-speech (POS) and chunk tags. One more proposal of their study which improved the perfomance of PCE a lot was the use of a validation function in the extraction procedure. Details of the PCE algorithm are provided in Section 3.1.", | |
| "cite_spans": [ | |
| { | |
| "start": 167, | |
| "end": 187, | |
| "text": "(Chieu and Tou 2003)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 630, | |
| "end": 655, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 823, | |
| "end": 848, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "In this section, we describe the original PCE algorithm, and our extension for extracting \u2329ne, category, related-to, object\u232a quadruples.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Extraction system", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The PCE algorithm is depicted in Figure 2 , and the description is given in Figure 3 . Starting with two seed patterns, PCE extracted \u2329person, category\u232a tuples. The extracted tuples were used to extract occurrences of \u2329person, category\u232a tuples in texts for generating new patterns. Again, new patterns were used to extract new tuples. The process terminated when no more patterns were produced. A pattern is defined as a 4-tuple: where order (a Boolean value) indicates the occurrence order of person and category in a sentence; person_slot is a slot which will be replaced with a person named entity; middle is the string surrounded by person and category; category_pattern is defined as: category_pattern:=noun_phrase 1 (and noun_phrase 2 )? 1", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 33, | |
| "end": 41, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF3" | |
| }, | |
| { | |
| "start": 76, | |
| "end": 84, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "where noun_phrase i is a regular expression that matches a noun phrase with added POS tags.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "In order to extract new \u2329person, category\u232a tuples, for every sentence s, from a pattern (whose order is true), and for each person NE named_entity in s, we construct a regular expression: * named_entity middle category_pattern * If s matches the above regular expression, the \u2329person, category\u232a tuple is extracted according to the algorithm in Figure 4 , where is_valid(category) is a function that returns true if category is a sort of 'person', and false otherwise. The purpose of this function is to ignore unexpected matches, i.e., matches that give incorrect tuples. The is_valid function operates based on the fact that if a person is-a category, then the category must be a sort (or subtype) of person. Since a category is a sort of person is equivalent to the category is a hyponym of person (or person is a hypernym of the category), this constraint is checked by using WordNet (Fellbaum 1998) which contains hyponymic and hypernymic relations among concepts. When order is false, named_entity and category_pattern are switched. PCE can extract two tuples from a match, if there are two.", | |
| "cite_spans": [ | |
| { | |
| "start": 887, | |
| "end": 902, | |
| "text": "(Fellbaum 1998)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 344, | |
| "end": 352, | |
| "text": "Figure 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Occurrences: An occurrence of a \u2329person, category\u232a tuple is defined as a 4-tuple:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "(order, person, middle, category),", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "where middle is a string surrounded by person and category. An occurrence of a (person, category) tuple is extracted if a sentence s matches the regular expression: * person middle category * or * category middle person * After extracting occurrences from the text corpus, they are used to generate new patterns. However, a middle of an occurrence is not necessarily reliable, (Nguyen and Shimazu) proposed a method to retain reliable ones based on two criteria: repetition and diversity as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Repetition of a middle (repetition(middle)) is the number of times the middle appears between the person and category of (person, category) tuples of same person.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Diversity of a middle (diversity(middle)) is the number of times the middle appears between the person and category of (person, category) tuples of different persons.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "A middle that has repetition(middle)>threshold R seems reliable and is kept. A pattern seems specific if it is generated based on tuples of a person, so only middles that have diversity(middle)>threshold D are kept to make the generated patterns general (Condition 1). If a middle contains a verb phrase, the verb phrase should express the relation person is-a category (Condition 2).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "These two conditions are used in the pattern generation procedure as described in Figure 5 . In the experiments, for the simplity, threshold R and threshold D are set to the same value which was called threshold for short.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 82, | |
| "end": 90, | |
| "text": "Figure 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Pattern types: Patterns whose middles are generated directly from the middles of occurrences are called exact patterns. Exact patterns are relately reliable, howerver, they have low coverage. In order to increase the coverage, (Nguyen and Shimazu 2007) introduced two more types of patterns, i.e., sketch and extended sketch patterns. An example of middle of an exact pattern with order true is:", | |
| "cite_spans": [ | |
| { | |
| "start": 227, | |
| "end": 252, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\" ] ,/, [NP ABC/NNP ] [NP 's/POS \"", | |
| "eq_num": "(2)" | |
| } | |
| ], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The exact pattern with middle (2) can match the sentence:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "[NP Harvey/NNP Dzodin/NNP ] ,/, [NP ABC/NNP ] [NP 's/POS vice/NN president/NN ] \u2026", | |
| "eq_num": "(3)" | |
| } | |
| ], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "However, this pattern can not match a similar sentence that describes the \"director\" of another company, e.g., IBM, in the same syntax as (3):", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "Input: Text corpus D; Seed pattern set P S Output: The list L of quadruples Preprocessing: Find NEs in every sentence by an NER; Remove sentences that contain no person NE. Add part-of-speech (POS) and chunk tags for every sentence by a shallow parser; 1. P\u2190 P ; L \u2190 \u2205; If P is empty then return; else go to Step 2; If middle (2) is modified so that its pattern can match (4), then expected relations in both (3) and (4) can be extracted. In order to do this, (2) is converted into a template that can match other sequences having similar structure. Concretely, nouns, adjectives, cardinals and articles in a middle are replaced with a variable $word that matches a word. Below is the template constructed from the middle 2:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "\" ] ,/, [NP $word/NNP ] [NP 's/POS \"", | |
| "eq_num": "(5)" | |
| } | |
| ], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "This template was called the sketch of a middle. A new pattern type whose the middle is replaced with a sketch was called sketch pattern. Details about the extended sketch patterns and other information can be seen in (Nguyen and Shimazu 2007) .", | |
| "cite_spans": [ | |
| { | |
| "start": 218, | |
| "end": 243, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "PCE algorithm", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The purpose of PCE is to extract \u2329person, category\u232a tuples, in which category can be used as the fine-grained type of person, so the set of NE types can be expanded by automatically extracting from texts. When we extract the tuple \u2329\"Bette Raptapoulos\", 'analyst'\u232a from (1), we only have information: \"Bette Raptapoulos\" is-a 'analyst'. If we can extract the relation: 'analyst' for \"Prudential-Bache Securities Inc.\", we will have complete information about \"Bette Raptapoulos\". Since PCE can be used to extract tuples of other NE types, such as organization and location, we propose to extend the PCE to extract \u2329ne, category, related-to, object\u232a quadruples describing the relations: ne is-a category, and category related-to object (or relatedto relations for short). From our observations, the related-to relations can be expressed in the following ways:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "a) The category and object are linked by a preposition: \"category preposition object\", e.g., \"analyst for Prudential-Bache Securities Inc.\"", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "b) The category and object are connected by a possessive apostrophe: \"object's category\", e.g., \"Semi-Tech's chief executive officer\". This can be interpreted as \"category of object\", e.g., \"chief executive officer of Semi-Tech\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "c) The object and ne are linked by a preposition: \"category ne preposition object\", e.g., \"\u2026 said economist David Littmann of Manufacturers National Bank\u2026\", from which an expected quadruple is \u2329\"David Littmann\", 'economist', 'of', \"Manufacturers National Bank\"\u232a.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "d) The object is embedded in category, e.g., \"IBM president\". This can also be interpreted as \"category of object\", e.g., \"president of IBM\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "e) The related-to relation is implicitly expressed, e.g., \"Mr. Baird, who heads the Manhattan U.S. attorney's securities-fraud unit, denied the quote \u2026\", from which an expected quadruple is \u2329'Baird', 'header', 'of', \"securities-fraud unit\"\u232a.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "Since case e) does not have fixed expressions, we do not treat such cases. In case d), because object is already embedded in category, we do not need to extract the object. For cases a), b) and c), we build regular expressions to extract the object. We modify the procedure in Figure 4 to extract \u2329ne, category, related-to, object\u232a quadruples instead of \u2329ne, category\u232a tuples. Let category_str be the string containing the category, the regular expressions corresponding to each case are (we omit POS and chunk tags for readability): Figure 4 ), if the current processing sentence matches one of the above regular expressions, the object is produced by removing POS tags in noun_phrase; related-to is the preposition after removing POS tags in cases a) and c); related-to is 'of' in case b), then, \u2329ne, category, related-to, object\u232a quadruples are returned instead of tuples. If no regular expressions match the current processing sentence, the object and related-to are null. We call our new algorithm NECOE which stands for \"Named Entity Category Object Extraction\".", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 277, | |
| "end": 285, | |
| "text": "Figure 4", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 534, | |
| "end": 542, | |
| "text": "Figure 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Named entity category object extraction", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "The extracted \u2329ne, category, related-to, object\u232a quadruples are valuable for NLP applications. In this section, we use them for answering some types of questions. If ne in a quadruple is a person, the quadruple helps answer the query: \"Who is ne?\". If ne is of another type, such as organization or location, the quadruple helps answer the query: \"What is ne?\" For answering the question, we just search for a quadruple having the same ne as that of the question. If a quadruple is found, then the answer is:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Utilizing quadruples for semantic search", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "ne is a category related-to object.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Utilizing quadruples for semantic search", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "The extracted quadruples also help answer list questions, e.g., \"Give me the list of analyst for Prudential-Bache Securities Inc.\" The general form of this question type is \"Give me the list of category [related-to object]\", where the part in square brackets is optional. For answering this question type, we search for the list L of quadruples having the same category (related-to and object) as those of the question. The answer is the list of nes of quadruples in L.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Utilizing quadruples for semantic search", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "If related-to of a question is 'of', we also search for quadruples whose object is embedded in category (case (d) as discussed in Section 3.2). For example, if the question is \"Give me the list of president of IBM\", we also search for quadruples whose category is \"IBM president\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Utilizing quadruples for semantic search", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "Since person related questions takes a large portion among NE-related questions, as seen in Text Retrieval Conference (TREC) 9 question-answering track 2 , our experiments concentrated on extracting person named-entity-related relations.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments and evaluation", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "We used the same corpus as that used in (Nguyen and Shimazu 2007) , i.e., the Wall Street Journal (WSJ) corpus which consists of 595 files. After extracting the body part and removing other parts, e.g., the headers, a plain text collection of nearly 3 million sentences with the size of 308 MB was produced. From this text collection, a test set of 1,000 sentences was randomly selected. From the test set, 385 \"ne is-a category\" relations (called is-a relation for short) were manually extracted. Among 385 is-a relations, 199 relations had additional related-to relations. The distribution of related-to relations according to cases discussed in Section 3.2 is given in Table 1. In the preprocessing step of the algorithm in Figure 4 , all NEs in this plain text collection were tagged by LingPipe 3 . After removing sentences that contain no person NE, a collection of 667,981 sentences was produced (we call this big-dataset). Next, OpenNLP 4 was used to add POS and chunk tags for the big-dataset. ", | |
| "cite_spans": [ | |
| { | |
| "start": 40, | |
| "end": 65, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 727, | |
| "end": 735, | |
| "text": "Figure 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Dataset", | |
| "sec_num": "6.1." | |
| }, | |
| { | |
| "text": "Besides NECOE, we also extended the baseline program in (Nguyen and Shimazu 2007) to extract quadruples, and called this program NECOE-NoValidation, since it had no category validation function. We ran NECOE on the big-dataset to get patterns. These patterns are used to extract quadruples on the test set for evaluation. Since a related-to relation was extracted after an is-a relation was extracted, we evaluated the results of the two relations in quadruples separately. PCE used a threshold in the pattern generation procedure, and the proper value of threshold selected from experiments was 3. In our experiments, we also set the value of this threshold to 3. The results of our experiments are shown in Table 2 . Since extended sketch patterns did not increase the coverage much, we do not show their results. Figure 6 shows the growth of the number of extracted quadruples and distinct categories from the big-dataset. The figure shows that the number of actual categories (40761) is relatively large.", | |
| "cite_spans": [ | |
| { | |
| "start": 56, | |
| "end": 81, | |
| "text": "(Nguyen and Shimazu 2007)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 709, | |
| "end": 716, | |
| "text": "Table 2", | |
| "ref_id": "TABREF2" | |
| }, | |
| { | |
| "start": 816, | |
| "end": 824, | |
| "text": "Figure 6", | |
| "ref_id": "FIGREF8" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments and evaluation", | |
| "sec_num": "6.2." | |
| }, | |
| { | |
| "text": "Though sketch patterns help to extract only 5.5% of the total extracted quadruples, their discovered categories comprise 24% of total distinct categories. 3479Top ranked part-time CIA employee (1), partnership analyst (1), parliament deputy (1), parts marketing administrator (1), patent specialist (1), personal translator (1), freight carrier 1Bottom ranked Related-to relation managing director of investment bank, vice president for economic research, president of Trans World International Inc., deputy of Japanese equities, manager of sales Figure 7 : The prototype of a semantic search system. Table 3 lists some top, bottom ranked categories along with their frequencies, as well as some related-to relations extracted by NECOE. From our observations, the reason that decreases the precision of related-to relations is derived from the extraction of incorrect is-a relations. The precision of related-to relations is 100%, if it is calculated on correctly extracted is-a relations. Since the related-to relations has a relatively large portion in is-a relations that can not be extracted by PCE, this is the reason that decreases the recall of related-to relations. Figure 7 introduces the prototype of a semantic search system that utilizes the relations in quadruples extracted from the WSJ corpus.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 547, | |
| "end": 555, | |
| "text": "Figure 7", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 601, | |
| "end": 608, | |
| "text": "Table 3", | |
| "ref_id": "TABREF3" | |
| }, | |
| { | |
| "start": 1174, | |
| "end": 1182, | |
| "text": "Figure 7", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments and evaluation", | |
| "sec_num": "6.2." | |
| }, | |
| { | |
| "text": "In this paper, we proposed a method for automatically extracting from text documents \u2329ne, category, related-to, object\u232a quadruples describing that \"ne ISA category\", and \"category IS-RELATED-TO object\". We extended PCE algorithm to extract these quadruples. Our experiments on the Wall Street Journal corpus obtained relatively good results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "7." | |
| }, | |
| { | |
| "text": "'?' stands for there is zero or one. Since PCE works on sentences which has been parsed by a shallow parser, each sentence contains POS and chunk tags. In this paper, we ignore POS and chunk tags in all regular expressions for readability.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://tangra.si.umich.edu/clair/NSIR/cgi-bin/trec-question.cgi?collection=9&script=html/nsir.cgi 3 http://www.alias-i.com/lingpipe/index.html 4 http://opennlp.sourceforge.net", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "We also utilize the extracted quadruples in a semantic search system for answering some types of questions.Our algorithm can be applied to extract quadruples of other NE types, such as organization and location.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "acknowledgement", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Snowball: Extracting Relations from Large Plaintext Collections", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Agichtein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Gravano", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of the 5th ACM International Conference on Digital Libraries", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "85--94", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Agichtein, E. and L. Gravano, 2000. Snowball: Extracting Relations from Large Plaintext Collections. Proceedings of the 5th ACM International Conference on Digital Libraries, pp. 85-94.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Extracting Patterns and Relations from the World Wide Web", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Brin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "Proceedings of the 6th International Conference on Extending Database Technology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "172--183", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Brin, S., 1998. Extracting Patterns and Relations from the World Wide Web. Proceedings of the 6th International Conference on Extending Database Technology, pp. 172-183.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Named Entity Recognition with a Maximum Entropy Approach", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Chieu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Tou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of Conference on Computational Natural Language Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "160--163", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chieu, H. and N. Tou, 2003. Named Entity Recognition with a Maximum Entropy Approach, Proceedings of Conference on Computational Natural Language Learning 2003 (CoNLL-2003), pp. 160-163.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "WordNet: An Electronic Lexical Database and Some of Its Applications", | |
| "authors": [ | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Fellbaum", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Fellbaum, C., editor, 1998. WordNet: An Electronic Lexical Database and Some of Its Applications, MIT Press.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Automatic Extraction of the Fine Category of Person Named Entities from Text Corpora", | |
| "authors": [ | |
| { | |
| "first": "T", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Nguyen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Shimazu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "IEICE Transactions on Information and Systems, Special section on Knowledge, Information and Creativity Support System", | |
| "volume": "", | |
| "issue": "10", | |
| "pages": "1542--1549", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nguyen, T. T. and A. Shimazu, 2007. Automatic Extraction of the Fine Category of Person Named Entities from Text Corpora. IEICE Transactions on Information and Systems, Special section on Knowledge, Information and Creativity Support System, Vol. E90-D, No. 10, 1542-1549.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF1": { | |
| "text": "The DIPRE model.", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF3": { | |
| "text": "The PCE model.PatternsText corpus", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF4": { | |
| "text": "Extract the list L' of quadruples from sentences that match a pattern in P; L \u2190 L + L'; Let D' be list of sentences from which quadruples in L' were extracted, D \u2190 D -D'; If D is empty then return; 3. Extract the list of occurrences O of quadruples in D; 4. Generate new patterns set P' from O; P \u2190 P';", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF5": { | |
| "text": "PCE algorithm.", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF6": { | |
| "text": "\u2329person, category\u232a extraction from a match. 1. Group all occurrences in the list O by order and middle; Let the resulting groups be O 1 , O 2 , \u2026, O ; N 2. For each group O i , if the middle satisfies the two conditions, then generate a new pattern: (order, person_slot, middle, category_pattern) Pattern generation procedure. [NP Alan/NNP Baratz/NNP ] ,/, [NP IBM/NNP ] [NP 's/POS director/NN ] \u2026 (4)", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF7": { | |
| "text": "a) * category_str preposition noun_phrase * b) * noun_phrase's category_str * c) * named_entity preposition noun_phrase * After extracting a valid category and an ne (Step 2 of", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF8": { | |
| "text": "The growth of extracted quadruples (a) and distinct categories (b).", | |
| "uris": null, | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "TABREF1": { | |
| "text": "The distribution of related-to relations.", | |
| "html": null, | |
| "num": null, | |
| "content": "<table><tr><td>Case</td><td>a)</td><td>b)</td><td>c)</td><td>d)</td><td>e)</td></tr><tr><td>%</td><td>74.87</td><td>17.74</td><td>2.05</td><td>3.59</td><td>2.05</td></tr></table>", | |
| "type_str": "table" | |
| }, | |
| "TABREF2": { | |
| "text": "Results of experiments.", | |
| "html": null, | |
| "num": null, | |
| "content": "<table><tr><td/><td/><td/><td>Is-a relations</td><td/><td/><td/></tr><tr><td>Pattern</td><td colspan=\"3\">NECOE-NoValidation P(%) R(%) F(%)</td><td>P(%)</td><td>NECOE R(%)</td><td>F(%)</td></tr><tr><td>Seed</td><td>89.03</td><td>35.84</td><td>51.11</td><td>99.26</td><td>35.06</td><td>51.82</td></tr><tr><td>Exact</td><td>63.41</td><td>72.47</td><td>67.64</td><td>94.48</td><td>75.58</td><td>83.98</td></tr><tr><td>Sketch</td><td>62.88</td><td>74.81</td><td>68.33</td><td>94.50</td><td>80.26</td><td>86.80</td></tr><tr><td/><td/><td colspan=\"3\">Related-to relations</td><td/><td/></tr><tr><td/><td>P(%)</td><td>R(%)</td><td>F(%)</td><td>P(%)</td><td>R(%)</td><td>F(%)</td></tr><tr><td>Seed</td><td>92.13</td><td>41.21</td><td>56.94</td><td>97.53</td><td>39.70</td><td>56.43</td></tr><tr><td>Exact</td><td>79.34</td><td>48.24</td><td>60.00</td><td>97.03</td><td>49.25</td><td>63.97</td></tr><tr><td>Sketch</td><td>76.64</td><td>52.76</td><td>62.50</td><td>96.64</td><td>57.64</td><td>72.33</td></tr></table>", | |
| "type_str": "table" | |
| }, | |
| "TABREF3": { | |
| "text": "Some top, bottom ranked categories with their frequencies and some related-to relations. President (22679), Chairman (12835), Analyst (6729), Vice President (6011), Director (5821), Chief Executive Officer (5326), Judge (5050), Dr. (4931), Rep.", | |
| "html": null, | |
| "num": null, | |
| "content": "<table/>", | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |