| { | |
| "paper_id": "M98-1004", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T03:16:17.198226Z" | |
| }, | |
| "title": "Oki Electric Industry : Description of the Oki System as Used for MUC-7", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Fukumoto", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Kansai Lab", | |
| "institution": "R&D group Oki Electric Industry Co., Ltd. Crystal Tower", | |
| "location": { | |
| "addrLine": "1-2-27 Shiromi, Chuo-ku", | |
| "postCode": "540-6025", | |
| "settlement": "Osaka", | |
| "country": "JAPAN" | |
| } | |
| }, | |
| "email": "ffukumoto@kansai.oki.co.jp" | |
| }, | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Masui", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Kansai Lab", | |
| "institution": "R&D group Oki Electric Industry Co., Ltd. Crystal Tower", | |
| "location": { | |
| "addrLine": "1-2-27 Shiromi, Chuo-ku", | |
| "postCode": "540-6025", | |
| "settlement": "Osaka", | |
| "country": "JAPAN" | |
| } | |
| }, | |
| "email": "masui@kansai.oki.co.jp" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Shimohata", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Kansai Lab", | |
| "institution": "R&D group Oki Electric Industry Co., Ltd. Crystal Tower", | |
| "location": { | |
| "addrLine": "1-2-27 Shiromi, Chuo-ku", | |
| "postCode": "540-6025", | |
| "settlement": "Osaka", | |
| "country": "JAPAN" | |
| } | |
| }, | |
| "email": "simohata@kansai.oki.co.jp" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Sasaki", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Kansai Lab", | |
| "institution": "R&D group Oki Electric Industry Co., Ltd. Crystal Tower", | |
| "location": { | |
| "addrLine": "1-2-27 Shiromi, Chuo-ku", | |
| "postCode": "540-6025", | |
| "settlement": "Osaka", | |
| "country": "JAPAN" | |
| } | |
| }, | |
| "email": "sasakig@kansai.oki.co.jp" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "M98-1004", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "This paper describes the Oki Information Extraction system as used for MUC-7 evaluation [1] [2] . The tasks we have conducted are Named Entity, Co-reference, Template Element and Template Relation. Each module is implemented using MT system modules and pattern recognition modules. Our purposes to participate MUC-7 evaluation are to evaluate how MT system modules are eective for other application such as IE system and to develop our information extraction technology based on pattern recognition.", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 91, | |
| "text": "[1]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 92, | |
| "end": 95, | |
| "text": "[2]", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "INTRODUCTION", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Oki's MT system, PENS EE [3] [4] , is a commercial system and is one of major MT systems in Japan. Translation mechanism of PENS EE system basically applies transfer method. There are English-to-Japanese (EJ) and Japanese-to-English (JE) systems and both are based on the same software systems, that is, both MT systems are using the same Grammar Description Language (GDL), and GDL rule translator and interpreter (GDL system). Oki has already spent more than ten years for development and improvement of the MT system. The Oki IE system used for MUC-7 is composed of a surface pattern recognition module and a structural pattern recognition module. The surface pattern recognition module traces a text at surface linguistic level and detects NE elements and co-referred elements without any language analysis system such as lexical analysis and syntax analysis. The structural pattern recognition module traces parse trees of a text, which is generated by parser of MT system. Syntactic and semantic information embedded in the parse tree are used to detect NE elements, co-referred elements and so on. The structural pattern recognition rules are described in GDL and are executed on the GDL system which is based on pattern matching mechanism of tree data. Detected elements of tree data are marked and are extracted after execution of rules.", | |
| "cite_spans": [ | |
| { | |
| "start": 25, | |
| "end": 28, | |
| "text": "[3]", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 29, | |
| "end": 32, | |
| "text": "[4]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "INTRODUCTION", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Oki has submitted two systems: the English system for MUC-7 (NE, CO, TE and TR) and the Japanese system for MET-2 (NE). This is our rst participation in MUC and MET evaluation. In order to develop the systems in a short period, we utilized parsing module of the MT system for a sentence analyzer. The English system is developed using the English sentence analyzer of the PENS EE-EJ and the Japanese system is developed using the Japanese sentence analyzer of PENS EE-JE. We also utilized the GDL system for description of extraction rules (only for description of structural patterns). The GDL system has powerful rule debugging environment, which is very helpful to develop an extraction system in a short period.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "BACKGROUND", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In the MT system, PENS EE, translation rules are executed on the GDL system. A GDL rule consists of pattern matching part and action part. Pattern matching part describes conditions to specify a part of tree data and action part states changes of the specied tree structure and/or modication of node information of the tree. Sample transformation of tree data is shown in Figure 1 . Pattern matching species a sequence of article (art), adjective (adj) and noun and both sides of the sequence are arbitrary number of nodes (\\*1\" and \\*2\"). After transformation, node information of article is embedded into the noun node, and the noun and adjective nodes are moved under the noun phrase node (NP). In our current implementation, a lower node modied its upper node.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 372, | |
| "end": 380, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "OKI MT SYSTEM : PENS EE", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Oki's IE system consists of surface pattern recognition modules, structural pattern recognition modules, ltering programs to convert internal expression of parser to surface expression, and SGML tag processing modules. Architecture of the system is shown in Figure 2 . Figure 2 : Architecture of the Oki MUC system The system rstly recognizes surface level patterns in a text and adds SGML tags for each task. In this analysis, NE elements, CO elements and sub-category of NE elements for the TE task are extracted. In the SGML tag processing module, these tags and original SGML tags are embedded in their adjacent words for morphological and syntax analysis. Each sentence of the tag-processed text is parsed in the morphological and syntax analyzer which is originally used in the MT system. Structural pattern analysis rules for the NE task are implemented in the syntax analysis rules. The recognized patterns are expressed in node attributes in a parse tree. In order to obtain NE results, information of node attributes is extracted and embedded in a text as SGML tags by the NE lter.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 258, | |
| "end": 266, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 269, | |
| "end": 277, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "OVERVIEW OF THE OKI IE SYSTEM", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For the CO task, all parse trees of sentences are converted into one tree structure each child node of which is a parse tree of a sentence. In recognition of the CO task, anaphoric elements are extracted and their antecedents are detected by traversing the tree structure. Information of CO elements is extracted from the tree structure and embedded in the original text as SGML tags by the CO lter. CO-reference numbers are also set by the CO lter. For the TE and TR tasks, a text tree structure is used for identifying TE and TR elements. The TE and TR lters generate TE and TR information in BNF format from the tree structure of a text.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "OVERVIEW OF THE OKI IE SYSTEM", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The NE system consists of a surface pattern recognition module, a structural pattern recognition module and some ltering programs for tag-processing for parsing and post-processing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The NE system", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Surface pattern recognition is processed in the following sub-processes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Surface Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Sequential capitalized words are recognized as NE candidates.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition of capitalized area", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Head word processing A head word of a sentence is removed from NE candidates when it is registered in Non-element word list.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition of capitalized area", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Merging capitalized area Some functional words and prepositions, \\for\" and \\of\", are utilized for recognition of NE candidates. For example, the functional word \\Bank\" and the preposition \\of\" are used for recognition of the NE element, \\Bank of Tokyo\". The functional word \\University\" and the preposition \\of\" are used for recognition of the NE element, \\University o f T okyo\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition of capitalized area", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Type information of NE candidates is recognized by functional words such as \\Mr.\" of a person name, \\Bank\" of a organization name, \\City\" of a location name and so on.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Type recognition of NE elements", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The rests of NE candidates are checked with word list of each NE type. The lists are manually extracted from newspaper articles, index information of a world map and company names from stock market news, etc.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dictionary look up", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In the recognition of NE elements, identied elements are utilized for recognition of their abbreviation which is repeated in a text. For example, when \\Mr. John Doe\" in a text is recognized as a person name, the words \\John\" and \\Doe\" in the same text are also recognized as a person name. Moreover, the abbreviation \\FAA\" in a text is used for checking NE candidates after recognition of \\Federal Aviation Administration\" in the same text.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dictionary look up", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "An original text is SGML-tagged one and NE elements in the text are also SGML-tagged after the surface pattern recognition. In order to parse such a SGML-tagged text, these tags have t o b e concealed. In a parse tree, tag information is expressed in node attributes of a parse tree, therefore, the system can handles information obtained at a surface level pattern recognition during parsing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Tag-processing for Parsing", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "After parsing, several structural pattern rules are applied to a parse tree in order to recognize NE elements. Some of them are as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The subject element of some types of verb such as \\say\", \\die\", \\play\" and so on is recognized as a person name.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The noun phrase in front of relative pronoun \\who\" is recognized as a person name.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The noun phrase, whose appositive phrase is a person name, is recognized as a person name, vice versa.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The noun phrase followed by \\employee\", \\spokesman\" and so on is recognized as an organization name.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The noun phrase with preposition \\in\", \\at\", \\near\" or \\over\", whose appositive phrase is an organization name, is recognized as an organization name.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Pattern Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "After the structural pattern recognition, NE tag information is extracted from a parse tree and added to the pattern tagged text. All the NE tag information is utilized for tagging to non-body part of the text.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Post-processing", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The CO system also consists of a surface pattern recognition module, a structural pattern recognition module and some ltering programs for tag-processing for parsing and post-processing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The CO System", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In the surface pattern recognition of the NE system, abbreviation and repeated one are utilized for recognition of NE elements. This mechanism is also used for surface level recognition of CO elements. For example, when the person name \\Mr. John Doe\" appears in a text, the words \\John\" and \\Doe\" are utilized as abbreviation of the person name.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Surface Level Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Tag-processing for Parsing SGML-tags and CO tags obtained from the surface pattern analysis are concealed for parsing as well as the NE system.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Surface Level Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This is the main module of CO-reference analysis. Firstly, the CO system recognizes co-reference expression of appositions and the expression of \\A is B\". Then, the CO system extracts anaphoric expressions of pronouns and noun phrases with denite articles, and traverses the tree structure from bottom to top and left to right way to detect its antecedent.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Structural Level Recognition", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Information of CO tags is extracted from the tree structure of a text and is added to the pattern tagged text. All the CO tag information is utilized for tagging to non-body part of a text in the same way of the NE system. Moreover, all the co-referred elements have the same co-reference number, therefore, renumbering according to the CO task denition is also done in this module.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Post-processing", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The TE system consists of the following sub-processes.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The TE System", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Entities which are recognized in the NE system are selected. Capitalized elements which are not identied in the NE system will be handled as candidates for TE elements. These elements will be identied as TE elements, if they are related to some descriptor.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition of Entities", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "Noun phrases and prepositional phrases which have functional words of descriptors such as \\Comdr.\", \\President\", \\agency\" and so on are extracted. Semantic information of Descriptors which is used in parsing will be utilized for recognition of type information of TE entities in the next process.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition of Descriptors", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "Relation between entities and descriptors is recognized by structural pattern rules for the TE task. For example, an entity and a descriptor are related on some verb such as \\called\", \\named\" and \\be\", they are recognized as a TE pattern.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Recognition Relation between Entities and Descriptor", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "Some TE patterns are merged using Co-reference information, that is, the same entities (coreferred elements) are recognized as one entity.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Merging TE Patterns", | |
| "sec_num": "4." | |
| }, | |
| { | |
| "text": "The TR system identies relation between TE elements. In our current simple implementation, if a person and an organization are related on some predicate, they are recognized as having \\employeeof\" relation. Moreover, if an organization and a location are related on some predicate, they are recognized as having \\locate-of\" relation. In case of an artifact and an organization, they will have \\product-of\" relation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The TR System", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Table shows results of the workthrough article. As for the TE task, we mainly implemented rules for entity, therefore, the TE score for entity is at an average level according to the results of MUC-6. However, the score for location is very low. It is because the TE system has a few location rules and has no dictionaries to identify the detailed information of a location. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "WORKTHROUGH ARTICLE", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This participation to the MUC-7 and MET-2 tasks was our rst trial and we had to develop our IE systems for the tasks in a short period. Our purpose to participate MUC-7 tasks is to evaluate how MT parsing modules are eective for development of an IE system in a short period and how they work. We found that it is useful to utilize the MT system modules for the other application such as IE system because our MT system has practically achieved robustness and the rule description system called the GDL system is also very helpful due to its pattern matching mechanism. However, parsing modules of the MT system has been originally developed for language transformation in a transfer method, therefore, tree structure of an original language is sometimes converted to the structure of a target language. These kinds of rules caused some diculty of pattern matching in the IE tasks.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CONCLUSION AND FUTURE DIRECTION", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "We have participated the MET-2 task in Japanese which has only the NE task. We are planning to apply our IE technology developed for MUC-7 tasks to a practical Japanese information extraction system. Moreover, it will be useful to participate Japanese CO, TE, TR and ST tasks if they will be dened in the next conference.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "acknowledgement", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Proceedings of 6th Message Understanding Conference (MUC-6)", | |
| "authors": [], | |
| "year": 1995, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Proceedings of 6th Message Understanding Conference (MUC-6), DARPA, (1995).", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "\\Analysis of Lengthy Sentences Using an English Comparative Structure Model", | |
| "authors": [ | |
| { | |
| "first": "F", | |
| "middle": [], | |
| "last": "Masui", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Tsunashima", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Sugio", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Tazoe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Shiino", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "40--48", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Masui, F., Tsunashima, T., Sugio, T., Tazoe, T. and Shiino, T.: \\Analysis of Lengthy Sen- tences Using an English Comparative Structure Model\", System and Computers in Japan, pp.40{48, SCRIPTA TECHNICA Inc., (1996).", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "Transformation of tree data in the GDL system" | |
| } | |
| } | |
| } | |
| } |