Erik Peterson Computational Linguist, Applications Technology Graduate Student, Georgetown University A Chinese Named Entity Extraction System ABSTRACT For many information applications, being able to identify the proper names and other entities in the text is a vital step in understanding and using the text. For example, in a Chinese-English machine translation system, if a word is identified as a person name, it can be romanized, rather than being translated as a regular word. Other entities, such as times, dates, and money amounts, are best translated by modules with special knowledge of these domains. This paper discusses the development and testing of a Perl-based named entity identification and extraction system for simplified Chinese text. This system can serve as a first stage to other Chinese language processing systems. Entities that are identified include locations, person names, organizations, dates, times, money amounts, and percentages. The system uses a segmenter and a specially created pattern matching language to identify these named entities. Useful criteria for finding each of these entity types, along with the major problems in finding them, are discussed. Scores are reported for the system as run on a test corpus and possibilities for improvement are proposed. INTRODUCTION ------------ For many information applications, being able to identify the proper names, among other entities, in the text is an important step in understanding and using the text. For example, in a Chinese-English machine translation system, if a word is identified as a person name, it can be romanized, rather than being directly translated. Other entities, such as times, dates, and money amounts, are best translated by modules with special knowledge of these domains. Fortunately, these entities have common characteristics and occur in regular contexts that make it possible to write patterns to identify them. To utilize these common patterns, I have written a complete system for processing Chinese text and identifying named entities. This system is written entirely in the Perl programming language, making it portable to a variety of computer systems. It runs on texts encoded in the GuoBiao (GB) computer code set for Chinese, as used in the People's Republic of China and Singapore. An important part of any Chinese language processing system is tokenizing the text into character compounds. How the text is segmented is influenced by the needs of the subsequent processing stages. The segmenter used by this extraction system uses a simple Maximal Matching Algorithm, augmented with routines to find numbers, transliterated names, and Chinese names. Development of the segmenter occurred in close coordination with development of the extraction system. For example, compounds that would otherwise be considered valid words are removed from the segmenter's lexicon if they cause conflicts in finding a named entity. For example, the compound 英国人 (England + Person = Englishman) was removed since it interfered in finding "England" by itself as a location. Decisions on what to consider as a person, location, organization, etc. are mostly drawn from the tagging guidelines established for the Message Understanding Conference sponsored by the United States government [1]. Each extraction type below includes a summary of what was included and excluded during name finding. English text mixed in with the Chinese is not processed at all, as this is outside the scope of a Chinese named entity extraction system. Once a word or phrase is identified as belonging to a given entity type, its offsets are stored and can be a useful input to further language processing, machine translation, and information retrieval systems. The rules developed for this system also reflect the target documents on which the system was run. For development and testing, I used simplified Chinese news summaries from the Voice of America broadcasts to China. The writing style used by a mainland China newspaper or a Singapore newspaper would lead to the modification of existing rules and addition of new ones. PATTERN MATCHING ---------------- Although the Perl programming language provides a rich set of pattern matching operators, these operators are designed to work on the character level rather than on the word level. To facilitate the matching of named entities, which are usually one or more words, I created a simple word level pattern matching language which is the basis for all of the entity identification done by the system. My general pattern matching language allows for the identification of entities based upon the internal characteristics of the entity (the entity itself) and the external context of the text in which the entity occurs. For each entity type, I wrote several rules. A rule has four ways of matching a word: 1. The word (or phrase) is a member of a set of words. The system has several word lists that it uses to identify entities and parts of entities. The name of the set will start with a percent sign. The different sets currently used by the system are: %geonames: Name of locations %geotypes: Suffixes that indicated place names %surnames: Common Chinese surnames %notname: Words that do not occur in names %titles: Person titles e.g. 先生 Mr., 大夫 Dr., 参议员 Senator %persons: Known person names %orgnames: Known organization names %orgwords: Words that are commonly used in organization names %orgtypes: Suffixes that indicate an organization e.g. 公司 Company, 有限公司 Ltd. %dates: Known dates; e.g. 今天 today, 去年 last year %times: Known times; e.g. 夜间 tonight, 早上 morning %currency: Currency names; e.g. 美金 dollars, 日圆 yen 2. The word meets a particular test or has some desired feature. The system can run tests on words. The results of the test are passed back to the top level rule. Word tests are indicated by an ampersand at the beginning of the test name. Some word tests used by the system include: &allforeign: Is the word entirely composed of characters that are commonly used in transliterating foreign names into Chinese? Useful for finding locations and Western person names. &isPossibleChineseName: Could this word be a Chinese name? Looks for a Chinese surname and an acceptable given name. &isChineseName: Is this word definitely a Chinese name? &allnumbers: Is this word entirely composed of Chinese and/or Arabic numerals? Useful for finding times, dates, money amounts, and percentages. 3. The word is an exact match for a word in the rule. These words are indicated by being enclosed in double quotes. e.g. "地区" (district) 4. Using "ANY" in the rule will match any word. Each rule has the form: (TYPE (... BEG ... END ...) ADD) where TYPE is the type of entity to find with this rule (LOCATION, TIME, etc.) BEG and END delimit the part of the match that is actually identified as belonging to TYPE. Anything outside of BEG and END is context that still must be matched for the entire rule to match. If ADD is included at the end of the rule, then whatever was matched by the rule will automatically be matched elsewhere in the text. Before and after BEG and END can be any sequence of the four types of word matching mechanisms listed above. Finally, an element in the rule can have an asterisk "*" or plus sign "+" appended onto the end to signify repetition. In the case of the asterisk, it means the match can occur as many times in a row as possible, or none at all. For the plus sign, the match must occur at least once, but can match as many times as possible. For example, '&allforeign+' would match all the words in a row that are composed only of characters for transliterating foreign names. Not yet implemented, but potentially very useful, is an aliasing mechanism for entities identified by the system. This would be useful in those cases where an abbreviated form of the entity is used after the term is first introduced in the text. For example, it is common in Chinese to use the first character of a country name as abbreviation of the entire name. A country alias system would take the first character of the country name and then look for it elsewhere in the text. Other possible alias mechanisms are listed under the extraction types below. EXTRACTION TYPES ---------------- LOCATIONS Locations as identified by the extraction system are countries, states, provinces, cities, towns, bodies of water, islands, and named geographic features (mountains, valleys, etc.). Other types of locations that are tagged include military bases, buildings, and other immobile structures. Not tagged are movable structures (planes, ships) or regions denoted by compass directions ("the southern state", "the west half of the state"). Examples - 美国 America - 以色列 Israel - 海牙 the Hague Location Rules '(LOCATION (BEG &allforeign+ %geotypes END) ADD )', '(LOCATION (BEG &allforeign+ END "地区") ADD )', '(LOCATION (BEG %geonames %geotypes END) ADD )', '(LOCATION (BEG &allforeign+ END "附近") ADD )', '(LOCATION (BEG ANY END ANY "两国") ADD)', '(LOCATION (BEG ANY END "两国") ADD)', '(LOCATION (BEG %geonames END))' A productive way of identifying locations involves looking for a likely foreign name followed by a geographic suffix. In my system the majority of the locations were found through look-up in a large location name lexicon. However, while this may work well for international news text where only a limited number of locations are seen on a regular basis, it would need to be adapted for local areas. Each time text from a new area is processed, its locations would need to be added to the lexicon. In the near future, I also will add rules that find a location based on its context ("at ...") in the sentence. Location names that are a foreign name followed by a geographic suffix can be aliased by removing the geographic suffix and just looking for the foreign name. Countries can be aliased by looking for just the first character of the country name, as described above. PERSON NAMES Persons are the names of people, either real or fictional. Person names do not include titles (e.g. "the Pope"). Person Rules '(PERSON (BEG &allforeign+ "．" &allforeign+ END) ADD )', '(PERSON (BEG &allforeign+ "." &allforeign+ END) ADD )', '(PERSON (%titles BEG &allforeign+ END) ADD )', '(PERSON (BEG &allforeign+ END %titles) ADD )', '(PERSON (%orgtypes "长" BEG &allforeign+ END) ADD )', '(PERSON (%geonames "人" BEG &allforeign+ END) ADD )', '(PERSON (%titles BEG &isPossibleChineseName END) ADD )', '(PERSON (BEG &isChineseName END) ADD )', '(PERSON (%titles BEG %surname ANY END) ADD )', '(PERSON (BEG %persons END))' Chinese has two main ways of expressing person names. The first method is for native Chinese names. Chinese names have a one character surname (or rarely, two characters) that comes at the start of the name. This is followed by the one or two character given name. Surnames in Chinese come from a limited set of possibilities but there is not a limited set of given names. Complicating name finding is the fact that surnames can serve other functions in Chinese, as can the characters used in given names. My rules try to improve identification of names by looking at the context in which each occurs. If a word has a title next to it and has the form of a possible Chinese name, it is matched as a Chinese name. Examples - 徐文立 Xu Wenli, Prominent Chinese dissident - 李鹏飞 Li Pengfei, Chinese smuggler - 林海 Lin Hai, convicted of computer crimes The other common type of name is drawn from names for non-Chinese. In most cases, this is simply a transliteration of the sounds of the name into like sounding Chinese characters. Fortunately, this is usually done with a small set of characters, but unfortunately, these characters are also commonly used elsewhere in the language. Finding a word made up of all foreign characters that is in the right position in the sentence for a name will identify the word as a person name. Examples - 苏哈托 former prime minister Suharto (sounds like soo ha twoh) - 克林顿 President Clinton (sounds like kuh leen dwoon) Regions around China that have been influenced by Chinese culture sometimes do not fall into either above category. For example, Korean names are usually three characters, like Chinese names, but can vary in the surnames used. Japanese names are often four characters long and use a different set of (usually 2 character) surnames. Vietnamese and other Southeast Asian names all have their own peculiarities that make identifying them difficult. Japanese Name Example - 小渊惠三 Keizo Obuchi, Prime Minister of Japan Person names can be aliased several ways. For Chinese names, the system could look for instances when just the surname is used later on in the text. For foreign names which include both a surname and a given name, the system could look for either name separately. ORGANIZATIONS Organizations can include a wide variety of types, and as such can be one of the hardest types of entities to identify. Organizations include company names, official governmental bodies, educational institutions, political parties, and military divisions. * Examples - 美国国会 U.S. Congress - 中国民主党 China Democratic Party - 联合国 United Nations - R-J-R 纳比斯科控股公司 RJR Nabisco Corporation Organization Rules '(ORGS (BEG &allforeign %orgtypes END) ADD )', '(ORGS (BEG %geonames %orgtypes END) ADD )', '(ORGS (BEG %orgwords+ %orgtypes END) ADD )', '(ORGS (BEG %orgnames END))' The main way of finding organizations is to look for a person or location name followed by an organization suffix. While imperfect, this finds many organizations. It is hard to distinguish between "American Oil Company" and "American oil companies", since both would be the same in Chinese. It is hard to determine the start of an organization. Organizations that are the same as people names are hard to distinguish from regular people names. Organization names are often aliased by taking one character from each word used in the full name. However, which character is selected from the word is not standardized, making it difficult to build a generic organization aliasing algorithm. DATES Dates include specific decades, years, months, dates, weekdays, or combinations of these. Dates in Chinese have a standard, accepted way of ordering the different parts of a date expression where each part is listed from largest time unit to smallest. This makes identifying dates in Chinese an easier task than finding dates in English, which allows a wide range of variation in constructing dates. The system also finds a few simple relative dates, such as "tomorrow", "next year", "last week", and "March of last year". * Examples - 1998年 1998 - 1970年9月28号 Sept. 28, 1970 - 明年 Next year - 本周 This week Also identified as dates are holidays and special designated days, such as World AIDS Day. * Examples - 圣诞节 Christmas - 世界艾滋病日 World AIDS Day However, not identified as dates are time ranges or relative dates with no fixed reference. For example, "3 weeks", "the past 4 days", and "5 years after the election" would not be found. Date Rules '(DATE (BEG &allnumbers "年" END))', '(DATE (BEG &allnumbers "年" &allnumbers "月" END))', '(DATE (BEG &allnumbers "年" &allnumbers "月" &allnumbers "日" END))', '(DATE (BEG &allnumbers "月" &allnumbers "号" END))', '(DATE (BEG &ismonth &allnumbers "号" END))', '(DATE (BEG &allnumbers "月" &allnumbers "日" END))', '(DATE (BEG &allnumbers "月" END))', '(DATE (BEG &allnumbers "月份" END))', '(DATE (BEG %dates END))' The date rules find various combinations of year, month, and day. Work needs to be done to identify other dates such as centuries, decades, and parts of months (China divides a month into three ten day periods). TIMES Times include specific time references (e.g. 4:17 pm) and time periods during the day, such as "morning", "afternoon", and "evening". Time Rules '(TIME (BEG &allnumbers "钟" END))', '(TIME (BEG &allnumbers "分钟" END))', '(TIME (BEG %times END))' The first rule finds a number followed by the Chinese version of "o'clock". The second rule finds periods of the day, such as "afternoon" and "night". The limited number of times mentioned in my test corpus limited the development of these rules. Future work needs to be done to find other shorthand ways of expressing time in Chinese, such as the equivalents to "a quarter after 5" or "half past nine o'clock". Also, the words for evening and morning can serve other uses in Chinese. A method for distinguishing when each is used as a time or as an adjective is needed. MONEY AMOUNTS Money amounts are a given amount of a legal currency such as dollars, yen, pounds, or the soon to exist euro. Other financial instruments, such as stocks or bonds, are not included. Stock market indexes and points are also not included. * Example - 十亿美元 1 billion U.S. dollars Money Rules '(MONEY (BEG &allnumbers+ %currency END))', '(MONEY (BEG "$" &allnumbers END))' Finding money amounts is as simple as finding a number followed by a currency type. The second rule will also find money amounts expressed with a dollar sign. While this finds the majority of money references in newspapers, greater coverage of the different currency signs (such as Britain's pound sign or the Singapore dollar sign, S$) of the world are needed for complete coverage. Further questions to ponder might include how to handle money ranges (e.g. "5 to 6 dollars"). Would this be tagged as one unit or would "5" and "6 dollars" be tagged separately for future processing. PERCENTS Percentages are expressed in a standard way in Chinese. A percent, or more generally a fraction, is a number, usually 100, followed by the word "分之" (parts of), and then followed by another number. A percentage expressed as 100分之30 would literally be "30 parts of 100". Percents can be expressed using the percent sign "%". The system will also find percent amounts expressed as percentage points. * Examples - 百分之三 3% - 百分之19 19% - 三分之一 33.3% - 两个百分点 Two percentage points Percent Rules '(PERCENT (BEG "百分之" &allnumbers+ END))', '(PERCENT (BEG &allnumbers "分之" &allnumbers END))', '(PERCENT (BEG "％" &allnumbers+ END))', '(PERCENT (BEG &allnumbers+ "％" END))', '(PERCENT (BEG &allnumbers+ "百分点" END))', '(PERCENT (BEG &isPercent END))' The percentage rules closely follow the description listed above. No major problems were encountered in finding percentages. RESULTS ------- Development and testing was done on Chinese text drawn from news reports on the Voice of America web site [2] during the months of November and December. These files are in the public domain, making them easier to use and distribute. The files are also scripts for radio broadcasts and consequently are closer in style to spoken Chinese. This works well because written Chinese is often much more compact than spoken Chinese. The news reports also provided a rich store of entity names and introduced each entity as it was used. These factors made the VOA corpus ideal for development of the system. Identified entities in the output of the system were marked using SGML tags as adopted by the Message Understanding Conference. These tags are listed below: TAG TYPE START TAG END TAG ---------------------------------------------------- PERSON LOCATION ORGANIZATION DATE TIME PERCENT MONEY In addition, I also found a small corpus of hand-tagged sentences made available on the web by some of the organizers of MUC. This "Little Grove" corpus consists of 100 sentences picked apparently at random from various mainland Chinese news sources[3]. As such, the sentences have little context for the entities within. Also, they use the more telegraphic Chinese newspaper writing style. These two factors made scores on this corpus much lower than the VOA corpus. The "Little Grove" corpus was used as test data; no reference was made to it during development. The VOA files were split into two sets, test and development. These 10 test files were hand tagged using MUC-style SGML tags and then put aside. No further reference was made to them. Ideally, development of the test set should not be done by the person doing development, lest develpment be biased by a limited knowledge of the test set, but this was not possible in the time limits of the class. As part of the system, I also wrote a scorer in Perl to measure the accuracy of the system. The two measures of system performance are precision and recall. Recall measures how Recall and precision scores for the two test data sets are listed below: VOA Corpus Total Test Corpus Tags: 1117 Total Correct Corpus Tags: 514 Total Machine Corpus Tags: 971 RECALL : 46.02 PRECISION: 52.94 Little Grove 100 Sentences Total Test Corpus Tags: 254 Total Machine Corpus Tags: 147 Total Correct Machine Tags: 43 RECALL : 16.92 PRECISION: 29.25 CONCLUSION ---------- Initial development of the named entity extractions system has been promising. Major problems in extraction category were identified. I also wrote a scorer to obtain the recall and precision scores for the test data. Finally, I created an infrastructure that will significantly simplify future system development. Unfortunately, while developing my word-level pattern matching language will save a great deal of time in the long run, it took away time from actually improving the entity matching rules and hence scores are currently low. It is anticipated that two or three more months of part-time rule development would bring scores for precision and recall into the mid-80's. This would be near the scores for the professional systems tested by MUC. Other work that could improve scores would include improving the Chinese text segmenter that the rest of the system depends on so heavily. This system can be obtained from http://epsilon3.georgetown.edu/~petersee/cgi-bin/chinesene.zip or chinesene.tar.gz. After unzipping this file, the system can directly run on any computer with a Perl interpreter. The main file is chinesene.cgi. The system can also be run via the web at http://epsilon3.georgetown.edu/~petersee/chinesene.html. The web interface allows users to either paste in text to be processed or type in the URL of a web page that is downloaded and processed. REFERENCES ----------- 1. MUC-7 Tagging Guidelines ftp://ftp.muc.saic.com/pub/MUC/MUC7-guidelines/guidelines.NEtaskdef.3.5.ps.gz 2. Voice of America Chinese Radio Scripts http://www.voa.gov/chinese/mindex.html 3. Chinese Treebank's 100 with Named-Entity Tags http://umiacs.umd.edu/labs/CLIP/nrdnetaA.html