AN EXPERIMERTAL SYSTEM FOR A~OMATIC RECOGNITION OF PERSONAL TITLES

Casimir Borkowski Thomas J. Watson IBM Research Center

Yorktown Heights, N.Y. Smry.

Natural language seems to sublanguages (e.g., personal titles, its own structure which relative to is quite simple. contain various special-purpose personal names) -- each with the total structure of language

An ability to generate and to recognize automatically words and word strings belonging to various languages may prove to he very useful since role in indexing ~ in various systems for tributing information.

This paper (i) describes some of volved in automatic recognition of personal newspaper texts, (2) outlines some rules of to perform this task, (3) presents statistics rithm's accuracy and exhaustiveness obtained of the algorithm to texts, (4) discusses and interprets some of the results, and (5) suggests some applications for computer programs capable of recognizing personal titles and names. Motivation for the Exl~eriment.

One of the major questions of the day is the extent to which a computer can be instructed to identify various parts of texts written in plain, ordinary language. In trying to answer this question, we set ourselves the preliminary limited objective of developing an automatic procedure for identifying personal titles and personal names in English-language texts. i special-purpose subthey play an important extracting and disthe main problems intitles and names in

an algorithm designed

concerning the algo-

in manual application ) places a frequent source of personal names but also because many localities were named after people, as for example, Elizabet~ New Jersey andDallas, Texas. And to make matters worse, hotels~ business firms, universities, etc. can be named after people~a~d are often referred to by an abbreviated name which is that of a person (e.g., "He is staying at the Hilton", "He graduated from Stanford" "Ford was hit by a strike last week "~ UT~m of them cllmbedthe -~ ~ e~e ~st ''~ ) . As for personal names like "HelenaRubinstein" and "Max Factor", they designate persons as well as business firms, while "Philip Morris" is the name of a person, of a corporation, and of a brand of cigarettes.

Yet another difficulty arises in case of names of persons (e.g., "Madison") when they perform a naming function with regard to something, say an avenue, (e.g., "Madison Avenue"). Presumbly, it would be worthwhile to distinguish automatically references to persons from references to things named after persons.

Further difficulties in automatic recognition result from the co-occurrence in texts of names belonging to different name strings (e.g., "John Byron" as in "Estelle gave John Byron's Don Juan", "Mary Jane" as in "For Mary Jane had nothing but sympat-~" I ~xander Montgomery" as in "According to Alexander Montgomery was slow in exploiting successes".).

Other difficulties in :recOgnizing personal names result from the fact that personal titles are not unfailing aids in identifying and disambiguating personal names since titles themselves can be homographic with other types of words. For instance, "General" is a military rank in "General Mobutu", but not in "General Motors".

Further difficulties result from the fact that some titles are homographic with given names. How is an automaton to tell that "Dean" is a.title in "Dean Wiesner" but a name in "Dean Rusk", that "King" is a title in "King James" but a name in "James King", that '~arl" is a title in "the fourth Earl Russell" but a name in "the Chief Justice Earl Warren"? Some Reco~nltion Rules.

Our recognition algorithm was intended a~ a frame of reference in an investigation of the trade-off between the efficiency and the complexity of a series of algorithms.

To be able to investigate the trade-off between the efficiency and the complexity of a series of algorithms, we avoided taking as our point of departure a strong theory about the structure of language, memant~cs, pragmatics, etc., and about the amount of syntactic recognition required for successfUl identification in texts of personal titles and names. Instead, we sought to discover • \ what ~ssumptions and what information about natural l~a@e~ahd text~ may be pertinent .to.;.the resolution of the limited problems in recognition which we set for ourselves.

Stronger assumptions and muze elaborate techniques of analysis can be built into subsequent algorithms if required and as required. For instance, since parsing may be helpful in identifying sequences of names each of which is followed by its title (e.g., "The President nominated John Gordon Ambassador to Guatemala, William T. M. Beale Jr. Ambassador to Jamaica...") future recognition algorithms may parse sentences containing: (1) double-object verbs (e.g., "nominate") and (2) strings consisting of personal names followed by titles.

At a later time, parsing and/or other types of analyses may be extended to sentences, paragraphs, and articles containing other kinds of words and phrases. However, since parsing and other types Of analyses may be expensive, it would seem advisable to apply them only when they can reasonably be expected to provide economic solutions to valid problems.

Our rules describe the arrangement in the sentence of the words, phrases, and punctuation marks which are pertinent to recognition of names and titles. Generally, the description starts with the first, that is, the leftmost pertinent element and termi I nates with the last, or rightmost pertinent element. Recognition rules were given a "left-to-right" format because rules expressed in this way are easy to implement on an electronic computer.

For greater ease of understanding, the rules are'expressed here in narrative form. For the sake of brevity, only some recognition rules are listed here. A more complete description of identification rules is available elsewhere.(1) -

Our rules for recognizing names of persons take advantage of the style rules of The New York Times. We would conjecture that whereas details of name recognition rules may vary from newspaper to newspaper, their general pattern will remain fairly stable and independent of editorial conventions.

The rule for identifying personal titles which was selected as a reasonable first approximation states that a word or phrase in text is a personal title either: (i) if it matches a word or string of words on a list of titles or

(2) if it matches a word or a string of words which is on llst of words and phrases which commonly combine with titles "Acting", "Assistamt", '"~ice") and is followed by a personal (e.g., "Acting Mayor,, "Acting Assistant Vice President") a (e.g., title or

(3) if it is a personal title followed by a wor~ or a strim~ of words which is on a list of words which commonly combine with titles (e.g. t "-elect ~, " at Large", "pro tempore") as in "Senator-elect", "Ambassador at Large", "President pro tempore". or

(4) if it is a title designated by a llst, like say "Commissioner", followed by the word "of" and any capitalized word (e.g., "Co~nissioner of Parks"). or

(5) if it is a word beginning with a capital letter and followed by a title designated by a list (e.g., "Co~zmlssioner" as in "Police Co~,issioner" ). ~'

A preliminary (and a highly tentative) rule specifies how titles concatanate. This rule permits distinguishing some strings such as "Prime Minister, Sir" as in "Prime Minister, Sir Alec Douglas-Home", "Rev. Dr." as in "Rev. Dr. Martin Luther King", "Mr. Chairman"3"Mr. Counsel", "Mr. Chairman, Ladies, and Gentlemen:", and so forth from titles followed by names.

The present set of rules for identifying homographic (that is, ortographically identical) is relatively simple. It divides ambiguous titles words of Class I (e.g., "King", "Pope", "Prince") be titles (e.g., "King John", "Pope John") unless:

(1) preceded by either personal titles designated by such as "Mr.", "Dr.", "M. Sgt.", "General", and initials in various combinations (e.g., "Dr. Pope", "John Pope", "John M. King'!, "J. or (2) followed by such postnomial elements as "& Bros." (e.g., "King & Bros.), "and Company" (e.g., "King and Company"), and so forth.

titles which are with other words

into four classes;

are assumed to a 1/st

Occasionally, words of Class I are followed by capitalized words or phrases which designate various institutions, establishments, locations, and so forth which are frequently named after persons (e.g. 3 "Drug Store", "College", "Avenue", "Theorem"). Although, the llst of such words and phrases is open-ended, its most frequently occurring members can be discovered and listed quite easily. Furthermore, there is some evidence that many or most such phrases can be identified by means of recogniti6n rules. Words which are members of Class I are assumed to be names when they are followed by words such as "College" o~ phrases such as "Drug Store" . etc., or by "Mr. King", M. King") given names "John King", "Sr." (e.g., "King, Sr."),

Words which are members of Class II (e.g., "Kaiser", "Chamberlain", '~arl") are assumed to be personal names., Commonly occurring exceptionsto this rule (e.g., "Kaiser Wilhelm", "Lord Chamberlain", '~arl of" (if followed by a word beginning with a capital letter)) are listed.

Words which are members of Classes III and IV (e.g., "General", "Principal", "Justice") are assumed to be titles. Commonly occurring exceptions to this rule are listed ("General Assembly", "Major Medical Plan", "Principal Investigator", "Justice Department" ).

As our rules become more sophisticated, the need for lists of exceptions will diminish. However, it is likely that listing exceptions will often be an attractive alternative to rendering a rule more complicated. Personal titles in the plural are recognized by means of a simple rule. It the plural (i) it

(2) its Generals"

(3) if designates (e.g., "Collector" as in "District Collector of Internal Revenue").

A procedure in the plural is used case (e.g., singular: similar to the one for identifying titles to recognize personal titles in the possessive "Major General's", plural: "Major Generals'").

Our rules assume that the capitalized word or string of words and initials which frequently follows a title is the name of a person (e.g., "President Nkrumah", '~r. Paul-Henri Spaak", "Governor Nelson Rockefeller").

If a title is followed by a word beginning with a lowercase letter or by certain punctuation marks, this indicates that the title is not followed by a name. However, occasionally personal titles arm followed by names beginning with lower-case letters (e.g. I "President de Gaulle").

On occasion, titles are followed by capitalized words which are not names. This happens in particular when a title is followed by a capital~zed word or phrase which designates an states that a string of characters is a personal title in if: is recognizable as a personaltitle and if either final word is followed by the letter "s" (e.g., "Major ) or else one of the words of which it is composed and which a llst

as the stem for the plural is followed by the letter "s" institution, an establisb~ment', a site, and so forth, nsu~ed after a personal title (e.g., "Ambassador Bar", "Archduke Trio", "Emperor Concerto"~ "President Hotel", "Queens County", "Viceroy Lumber Company" ) •

Although -- as mentioned earlier -- the list of such words and phrases is open-ended, its most frequently occurring members have been listed and are consequently identifiable. Co~nly occurring exceptions to this rule are also listed and are therefore identifiable. In addition, we have some simple preliminary rules for identifying phrases whose designata often bear as names words or phrases which are personal titles (e.g., "President Radio Repair Shop", "Viceroy Lumber Company"). However, since the identification of phrases which designate such namesakes has been given little attention, these rules are very tentative.

Words which are generally names of weekdays when they occur after personal titles (e.g., "They saw the President Monday") are of course listable and therefore identifiable.

Occasionally, titles are followed by capitalized words which are not names and for the recognition of which the rules make no provisions (e.g., "British" in "The Prime Minister, British sources said, will arrive on Monday." and "New York" in "Mr. Stevenson prefers Washington and Mr. Rusk, the Secretary of Stat% New York."). Constructions such as these are, however, quite rare.

Occasionally, prepositional and other phrases intrude between a title and a name (e.g., "the French Ambassador to the _United States, Herve Alphand", "the Foreign Minister of France, Maurice Couve de Murville"). Prepositional phrases of this sort and other adjuncts are identified by various rules of the "brute force" type whose statements are constructed as follows: If a personal title (e.g., "Foreign Minister") is followed by a phrase consisting of the preposition "of" and of the name of a country, such phrase is part of the title.

In general 3 in a string of words consisting of titles and names t titles precede names (e.g., "the Secretary of State, Dean Rusk I the Foreign Minister of the Federal Republic, Gerhard SchrSder, the Foreign Minister of France, Maurice Couve de Murville"). Sequences of names each of which is followed by its title (e.g., "Dean Rusk, the Secretary of State, Gerhard SchrSder, the Foreign Minister of the Federal Republic, ...") are rare. (Ordinarily, iz. a construction of this type, each title is set off from the name which follows it by a semi-colon.)

In spite of counter-examples such as the ones above, one can reasonably assume that if the capitalized word or string of words and initials which frequently follows a personal title is NOT an identifiable word like "Hotel", "Garage", "Street" I "Monday", etc., or a phrase li~e "Barber Shop", '~rug Store", "Meat Packing Company", etc., then it is a personal name.

While counter-examples to this rule and to similar rules are easy to invent, the inventors of counter-examples usually miss the point that rules such as these are statistical observations a~ that in actual application to texts they hold up rather well. As state~ earlier, among the goals of an investigation of this type is to obtain experimental evideuce as to how well the rules hol~ up an~ what amendments are required to simplify them, ~to render them more accurate, and to expand their scope.

If a personal title (e.g., "President") is conjoined to the titles "Mrs." or "Miss" (as in "the President and Mrs. Johnson"), the capitalized word or string of words which follows the conjoined titles is generally the name of a person.

In general, the name of a person acts distributively with regard to the preceding titles, that is to say, a phrase like "the President and Mrs. Johnson" decomposes into "President Johnson and Mrs. Johnson".

Occasionally, a personal name does not act distributively with regard to conjoined titles which precede it (as in "an agreement between the Cardinal and Mrs. Johnson", "a meeti~ between the President and Mrs. Luce"); however, the present set of rules makes no provisions for recognizing such cases.

Generally, in newspaper ~rticles, capitalized words and initials which (a) frequently follow a personal title in the plural (e.g., "Senators") and (b) which ar~ not followed by other titles (e.g., "Senators, Congressmen, and Generals") are strings of personal names (e.g., "Senators Javits and Kennedy", "Senators Jacob Javits, Robert Kennedy and George D. Aiken", "Presidents Johnson and Lopez Mateos").

Of course, "Ambassadors Bar and Grill" is not a title in the plural followed by two names. However, commonly occurring phrases such as "Bar and Grill" are listable and therefore identifiable.

As a rule, the first name string -- which is often separated from the title by a comma -- terminates before the first conjunction "and" or before the next comma; the second name string begins after "and" or after the comma; etc.

If two name strings which follow a title in the plural (e.g., "Senators Jacob Javits, Robert Kennedy, ...") are separated by a comma, then -- generally -- the end of the second name string is marked by a comma or by the conjunction "and".

If two name strings which follow a title in the plural are separated by "and", then -- generally -- the end of the second name string is marked either by punctuation marks such as sentence period I a colon, a semi-colon, etc., or by a word beginning with a lower-case letter (e.g., "arrived" in "Presidents Johnson and Lopez Mateos arrived today."). However, if the word beginning with a lower-case letter is a name conjunction (e.g., "de", "yon"), then such word does not mark the end of the second name (e.g., "Presidents Lyndon Johnson and Charles de Gaulle").

Occasionally, the string of words which follows a title in the plural consists of both names and prepositional phrases (e.g., "Senators Javlts of New York and Fulbright of Arkansas", "Senators from New York, Javits and Kennedy"). Our rules permit identifying some prepositional and other phrases which may intrude between titles in the plural and the names which follow them.

Generally, the title in the plural acts distributively with regard to the names which follow it, that is to say, a phrase like "Senators Javits and Kennedy" decomposes into "Senator Javlts and Senator Kennedy".

Generally, the end of a name string which may follow a personal title in the singular is marked by punctuation (comma, sentence period, dash, semi-colon, colon, exclamation point, apostrophe t three dots, left or right parenthesis, etc.) or by a word beginning with a lower-case letter.

However, a lower-case letter does not mark the end of a name string if the word which begins with it is either:

(1) a name conjunction (e.g., "de" as in "Attorney General Nicolas deB. Katzenbach") or

(2) the last element of a hyphenated Chinese given name (e.g-, "lai" in "Premier Chou En-lai") or

(3) the one-letter Spanish word "y" (e.g., "President Jose Bustamante y Rivero")

(4) if it and so forth al-Kadir", is one of the Arabic words "ibn", "el", "al",

(as in "Abdul-Assiz ibn-Saud", "Abd-el Eader", "Abd-er-Rahman" ).

The end of a name string is often marked by its ~ment (e.g., "Jr." as in "Rev. Dr. Martin Luther King Jr.", in "Douglas MacArthur 2nd", the Roman numerals "I", "If", as in "King Idris I"). Cases in which Roman numerals are middle of a name are rare and can be treated as listable exceptions (e.g., "King Gustaf VI Adolf"). NOTE: The present rule makes no provisions for distinguishing Roman numerals "I", '~", and "X" from the first person pronoun "I" and from the letters "V" and "X" since contexts in which these ambiguities may cause error in name recognition seem rare (e.g. I "Malcolm X", "Pope Leo X", "Idrls I", "May I leave?"). "er",

"Abdlast ele-

"2nd" as "III", etc., in the

Ordinarily, a left parenthesis or a left bracket are among the punctuations which mark the end of a name string. This, however~ is not the case when a person'~s title and given name are followed by his nickname in quotation marks, as for example in "Gen. Howell ("Howling Mad") Smith", "Adm. William ("Bull") Halsey", etc.

Similarly, whereas ordinarily a left bracket marks the end of a name string, occasionally, when quoting someone, newspapers supply in brackets the part of name which the original state~nt omitted (e.g., "M~ agreement with Senator ~qichard B.\] Russell..."); sequences such @s these do not mark the end of a name.

The rule for identifying nicknames which was selected as a reasonable first approximation states that "strings of words in parentheses and quotes which occur immediately after the title and/or given names and before a surname are nicknames"..

A parallel rule serves to identify names in brackats which act as amplifications of original statements.

The preceding section has stated in considerable although by no means full detail some rules for identifying titles and names. We hope that this form of presentation indicates the vast amount of detail involved in rules for automatic recognition without, however, overburdening the reader with a multitude of minute points of information. Results of the Experiment.

Since our identification rules were embodied in dictionary entries and flow charts which were sufficiently detailed to permit an accurate manual execution of identification procedures, it was decided that our identification system would be tested out by hand on a sample of The New York Times texts.

Identification procedures were applied manually to some 40,000 words of texts. Altogether eighty-eight articles from eleven issues were selected and processed. Only newsartlcles were included in the sample. All materials found in the special sections such as (1) entertainment, (2) food-fashions-family-furnishings, (3) social events, (4) necrology, etc. were omitted. Materials in the sample consisted of only texts of newsartic±es; picture captions, advertisements, italicized lists of various sorts, charts and diagrams, etc. were excluded from the data.

Our 40,577-word sample contained 806 occurrences of ~ s of persons. Of the 806 occurrences of names of persons, 46 or about 6% of the total were missed. In addition, 47 words and word strings were mistakenly identified as personal names or personal titles. Figure of merit F for the results of this identification lO system was computed by means of the following formula: F=

C 2 (C+M) x T where C is the number of correct identifications, M is the number ofmistaken identifications, and T is thenumber of names of persons in the sample. (2) For T = 806, C =746, and M = 47 F = 7462 = .87 (746+47) x 806 Ar~l~sis of Ma~or Errors.

Twenty-six misses (out of a total of forty-six) and thirty mistaken identifications (out of a total of forty-seven) occurred in attempted identifications of words, word stems, and word strings which perform a naming function vis-a-vis some namesake (e.g., "Grumman Aircraft Engineering Corporation"). This source of misses and false identifications would be eliminated if in the future the automatic identification system was not required to decide whether words, word stems, and word strings (e.g., "Grumman") performing a naming function vis-a-vis some identifiable namesake (e.g., "Aircraft Engineering Corporation") are names of persons.

We also need more effective rules for phrases (e.g., "Aircraft Company") and personal man Apprentice") from their respective elements "Company", "Fireman", "Apprentice"). computing namesake titles (e.g., "Fire(e.g. I "Aircraft",

In addition, we need to caused by the assumption that all ambiguous words such as "General", etc. are names of persons. prevent or capitalized

"Justice", c~eliminate the errors words occurring after

"Major", "Principal",

We also require more effective rules to distinguish strings of titles (e.g., "President, Secretary of State") from titles followed by names. In addition, we need more effective rules for distributing a title among all ~ s of persons which follow it in the text (e.g., "Senators Vance Hartke and Birch Bayh of Indiana and Eugene J. McCarthy and Walter F. Mondale of Minnesota").

In ad@ition, we ~y requi~e rules which ~m0uld uheCk onHthe did ones : rather than supersede them. The new set of rules would be applied to words and phrases which were identified as names of persons by the old set of rules. The new rules could indicate the degree 3_I of confidence with which the al6orithm identified a wor~ or a string of wor~s as name of a person s or as a personal title followed by the name of a person, etc.

The advantage of this procedure consists in not having to revamp the algorithm in order to accomodate new rules. New rules would simply be tacked on to the old ones. New rules might che~k whether the elements of a string of letters, punctuation marks, spaces, numbers, etc. which the old rules had identified as a can be (a) words of the English language and (b) names of persons.

Whether a string of characters is a personal name could be decided by probability tables constructed along these lines: Is this string of letters an English word? Can this string of letters be a personal name? Is this string of letters the name Yes ~(es Yes Unknown No Yes No Unknown

The new rules should be relatively The question "Is this string of letters an English word?" could be answered by means of (a) a lookup in a dictionary based on some desk dictionary -- say Webster's Collegiate, and (b) si,~le rules for identifying affixes of the plural, the past tense, the gerund, the negation, etc. The question "Can this string of letters be a personal name?" could be answered by means of (a) a lookup in a dictionary based On a large telephone directory -- say the Manhattan Telephone Directs,and (b) simple rules for identifying the plural (e.g., "es" and "s" as in "the Joneses" and "the Weinbergs") and other affixes.

Improving the automatic identification system may require several subsidiary investigations. For instance, we may be well advised to determine the relationship -- if any -- between, on the one hand, the length, the date, the place of origin, the subject matter, the authorship, and the type of newspaper articles processed through the system~nd on the other, the effectiveness of the algorithm. 12 of a person? Probably yes It ' s unlikely Yes it ' s very likely easy to implement. Discussion and InteI~retation.

Automatic classification of words and phrases of the type described here can be regarded as a particularly simple case of machine translation. However, the goal of this type of machine translation is not translation into another natural language but TEXT REDUCTION: certain words and word strings are identified as "pertinent" (e.g., personal titles, personal names~ place names 2 street addresses, numbers and measures, dates and other time phrases, company names, trade names, chemical formulas, etc., etc.)and others as "not pertinent". Pertinent words and phrases are retained and labeled, and all others are suppressed.

Even this simple goal requires rules which are rather complex. However, because many word strings which the algorithms such as this one attempt to recognize have simple structure ("phrase structure"), they can be generated and possibly also recognized with a reasonable degree of accuracy by simple automata ("pushdown storage") or by a combination of linguistic and statistical techniques.

More generally, it may be useful to view natural language as a macro-language containing certain special-purpose micro-languages (or "sublanguages") -- each with its own structure which relative to the total structure of language is quite simple. It may be of some practical and theoretical interest (a) to investigate the structures and the inter-relations of such sublanguages and (b) to construct algorithms for identifying in texts words and word strings belonging to such sublanguages.

An ability to produce and identify automatically words and word strings belonging to various special-purpose categories (i.e., sublanguages, each with its own set of rules) may prove to be very useful in information retrieval because they play an important role in various systems for extracting and distributing information.

It would appear that along with researching and developing methods for high-quality fully automatic classification of words in texts, it may be advisable to set up efficient procedures for (a) manual classification and tagging of words and word strings in texts, and (b) subsequent automatic extraction of data from texts which were recognized either manually or automatically. One procedure for manual classification of words in texts would require computer-legible texts which can be projected on TV-type tubes (hereafter, "display screens") and either lightpens or cursors for writing on display screens. It may look approximately as follows:

A newspaper article would be copied from some type of machine-readable tape into a suitable computer. The computer would ±L~n project the article on a display screen. A clerk would then scan the display screen and locate various types of words and phrases in the article (say, names of persons, names of organizations, dates, addressesj and so forth).