NASA-CR-19779& Information Processing & Management, Vol. 30, No. 5, pp. 631-645, 1994 Copyright © 1994Elsevier Science Ltd Pergamon Printed in Great Britain. All rights reserved 0306-4573/94 $6.00 + .00 0306-4573(93)E001 I-C // J. MACHINE-AIDED INDEXING AT NASA JUNE P. SILVESTER, MICHAEL T. GENUARDI, and PAUL H. KLINGBIEL NASA CASI, 800 Elkridge Landing Road, Linthicum Heights, MD 21090-2934, U.S.A. (Received 8February 1993; accepted in final forrn 16 November 1993) Abstract-This report describes the NASA Lexical Dictionary (NLD), a machine-aided indexing system used online at the National Aeronautics and Space Administration's Center for AeroSpace Information (CASI). This system automatically suggests a set of candidate terms from NASA's controlled vocabulary for any designated natural language text input. The system iscomprised of a text processor that is based on the computa- tional, nonsyntactic analysis of input text and an extensive knowledge base that serves to recognize and translate text-extracted concepts. The functions of the various NLD system components are described in detail, and production and quality benefits result- ing from the implementation of machine-aided indexing at CASI are discussed. INTRODUCTION The National Aeronautic and Space Administration's (NASA's) machine-aided indexing (MAI) system is fully operational and cost effective. It is a third generation of a system designed by Paul H. Klingbiel for the Defense Technical Information Center (DTIC). This article describes the NASA system, which was developed as part of a concentrated effort to speed up the indexing of scientific and technical reports and cut costs. The system func- tions within normal NASA time constraints and workloads and is used in conjunction with electronic input processing. Although NASA has conducted a number of tests to evaluate its MAI system, measures were restricted to those that would not slow production. The best proof of the success of MAI is that indexers handle more work than ever before, they like MAI, and there has been no adverse effect on retrieval, as evidenced by user or retrieval analysts' complaints. NASA's system can be used for batch processing. More often it is used interactively during online document processing. NASA's MAI is designed as a tool for indexers, and all output is expected to be reviewed. To make processing fast enough for an online sys- tem, NASA replaced DTIC's method of machine selection of phrases and expanded the knowledge base (KB), which is used as a kind of conceptual network. This is described at greater length in this article. DTIC's phrase delineation method, which is the method that NASA originally em- ployed, is based on discovering noun phrases. This system uses a recognition dictionary to assign syntax to each word encountered in text and a Machine Phrase Selection (MAPS) pro- gram to string words together according to specified grammar rules (Silvester et al., 1993a). The object of MAPS is to identify the noun phrases that exist in natural language text. The first DTIC system required that the entire phrase identified by MAPS be located as a key to an entry in their Use Reference file, which they called the Natural Language Data Base (NLDB). This file became very large and cumbersome, but Klingbiel discovered that its contents could be reduced greatly. DTIC's current system has replaced the NLDB with a more condensed file called the Lexical Dictionary (Klingbiel, 1985), which was the pattern for NASA's original knowledge base. In the syntactic system described earlier, any word that is not likely to be part of a noun phrase becomes a stopword and is considered to be without indexable content. This includes verbs, since verbs are not part of a noun phrase. For DTIC's system, about Correspondence should be addressed to NASA CASI, 800 Elkridge Landing Road, Linthicum Heights, MD 21090-2934. 631 PlIEQ[DII_ PAGIE ELANK NOT FILMED 632 J.P. SILVESTEeRl aL 50e/0 of the words occurring in text are classified as stopwords, and NASA's original MAI system had more than 76,000 such words. NASA's present system is not based on grammatical parsing. It is semantically based. Bonnie Jean Dorr describes several arguments for choosing a semantic-based design over a syntactic one (Dorr, 1988), such as (1) the large number of rules required for a syntactic- based system to handle different meanings of context sensitive words, (2) the enormous amount of information needed to disambiguate words, and (3) the attention of syntactic systems to form rather than content. NASA's primary reason for using a semantic-based system was processing speed for online MAI. NASA's present system is based on the co- occurrences in parts of a sentence of what Alan Melby calls domain-specific terminology (Melby, 1990). This refers to words and phrases that are not broad in their meanings but that have (or suggest) domain-specific, semantically unambiguous, indexable concepts. These text words and phrases are matched against a list of text words and phrases that are generally synonymous to NASA's thesaurus terms. They are linked to NASA's thesaurus terms in the knowledge base. Susan Artandi describes this as an "inclusion list" approach (Artandi, 1976). Such a list can accomplish synonym control by switching from text words to specified index terms, and the input text is not limited to noun phrases. It may contain verbs, adverbs, and even nongrammatical word combinations that contribute to the rec- ognition of indexable concepts. When grammatical parsing was abandoned as the way in which to identify indexable concepts, it was possible to capture indexable concepts previ- ously discarded because they were not expressed in noun phrases. This improved the qual- ity of the final output. The semantic-based system significantly reduced the text processing time. It was also possible to reduce the stopword list to about 250 statistically selected words. These 250 stopwords and punctuation are used to break the text into word strings. The current online NASA system is constructed as a three-component system (see Fig. 1): • The knowledge base (KB), which is the dataset used for MAI; • Application programs; and • Access-2, a modular program that • Constructs search keys by concatenating words within established boundaries; • Looks up the search keys in the KB; and • Returns the candidate NASA Thesaurus posting terms and any other reports to the application program for output to the user. These components are described in greater detail next. THE KNOWLEDGE BASE The lists that indicate which thesaurus terms to use for any given input have been referred to by several names during the past 10 years. In 1982 at NASA, the dataset was called the NASA Lexical Dictionary (NLD), after the corresponding DTIC dataset. Soon the NASA Lexical Dictionary (NLD) became a system with three Use Reference files: • One for mapping DTIC Thesaurus descriptors to NASA's-a process that we call Subject Switching (Silvester et al., 1984; Silvester & Klingbiel, 1993); • A second dataset for DOE to NASA Subject Switching; and • A third dataset for mapping natural language to NASA thesaurus terms, referred to (after the separation of the original dataset into three) as the Phrase Matching file. The Phrase Matching file contained about 66,000 entries when NASA began to use it operationally. It has since developed into a knowledge base that, as of July 1993, con- tained more than 115,000 entries. The ultimate goal is to continue to expand the KB and to unify these three datasets insofar as it is feasible and compatible with natural language MAI. The KB could reach 250,000 entries before significant stabilization occurs. The growth of the KB is shown in Fig. 2. This, as we have stated, is the dataset that provides Machine-aided indexingat NASA 633 / / TEXT INPUT _UGGESTED PROGRAM TERMS1 BROKEN INTO ACCESS-2 FROM EUMNATE WFNOOOURTNDDS/ Fig. I. Overview of NASA'sonline machine aided indexing system. the translations from natural language to NASA's controlled vocabulary (thesaurus terms). Two fields are essential: • The Key field of each record, which is unique and serves as the computer address to the entry in the KB, and • The Posting Terms field. The unique Key field consists of one of the following: • Any word followed by a semicolon and three nines. Nines are used because they sort last in NASA's IBM-4381 mainframes, on which MAI is processed.* (A single word followed by 999 must sort last because that entry is the default lookup, which is of interest only when other combinations beginning with the initial word are not found. The word combinations beginning with the same initial word are searched sequen- tially in the computer's sort order. Sort order on the IBM mainframe begins with spaces and symbols, followed by alphas, and ending with numbers 0 through 9.) • A combination of two or more-words separated by semicolons. *At present,indexers haveaccessto the mainframe from a 3270-type terminal. 634 J.P. SILVESTER et al. 120I 1101 100 m 90!" 8o" ;/ E'o ;o / ! 40 Z 30 20 = c,_ i 101 n 1 1982 Ic.34 =1986 1_88 1990 1992 1983 1985 1987 1989 1991 1993 YEARS [] NO. OF ENTRIES AS OF JUNE 30 YEAR ENTRIES YEAR ENTRIES 1982 14,000 1988 98,300 1983 41,000 1989 101,800 1984 41,700 1990 107,000 1985 45,000 1991 110,700 1986 64,800 1992 111,500 1987 86,000 1993 115,400 1083-1984SNOWSEFFECTOFREDUCTIONINRESOURCES 1991-1992ALSOSHOWSEFFECTOFREDUCTIONINRESOURCES FORTHISPARTICULARTASK Fig. 2. Knowledge base growth. • A combination of two or more words separated by semicolons that becomes a unique combination of characters by the addition of ";999". See the examples in Fig. 3. Posting terms are known by a variety of names. In this document we have tried to restrict our terminology for that concept to either posting terms or NASA thesaurus terms. The Posting Terms field contains • One or more thesaurus terms that are equivalent in meaning to the key; or • Two zeros (00); or • An asterisk (*). See the examples in Fig. 3. The Key field and Posting Term field are a more robust rewrite system than that of Use References in a standard thesaurus. The usual Use reference types (with the addition of semicolons and the 999 flag) occur as keys, but a new and powerful concept is added- a rewrite to 00 in the Posting Term field. Linguistically, this deletion is a zeroing rule. As a conceptual network, the knowledge base contains not only entries that map natural language words and phrases to controlled vocabulary terms, but also entries that represent decisions regarding the relevancy of particular concepts (Genuardi, 1990). For example, within the aeronautics domain, the concept AIRCRAFT is much too broad in meaning to be a useful indexing term for most instances in which the word aircraft appears in text. In this case, specific entries in the knowledge base would initiate a search for a muitiword semantic unit such as A-320 AIRCRAFT, which describes the specific vehicle in question; or AIRCRAFT STABILITY, AIRCRAFT CONSTRUCTION MATERIALS, or AIRCRAFT CONFIGURATIONS, which indicate the particular aeronautical aspect of interest. Other entries in the knowledge base serve to disambiguate certain words such as matrices, which might refer to either mathematical matrices or material matrices. 635 Machine-aided indexing at NASA Keys PostLng Terms INDIRECTLY;999 O0 INDIUM;OXIDE INDIUM;OXIDE;COATING COATINGS,INDIUM COMPOUNDS INDIUM;OXIDE;COATINGS COATINGS,INDIUH COMPOUNDS INDIUM;OXIDE;999 INDIUM COMPOUNDS,METAL OXIDES INDIUM;999 INDIUM LIQUID;PHASE LIQUID;PHASE;EPITAXY LIQUID PHASE EPITAXY LIQUID;PHASE;999 LIQUID PHASES LIQUID;ROCKET LIQUID;ROCKET;PROPELLANT LIQUID ROCKET PROPELLANTS LIQUID;ROCKET;PROPELLANTS LIQUID ROCKET PROPELLANTS LIQUID;ROCKET;999 00 LIQUID;999 00 Note that in the Key field there are no spaces, and in the Posting Term field multiple thesaurus terms are separated by comas with spaces occurring only between words in a multiword term, In a file as large as the NASA knowledge base, this practice not only saves storage space, but more importantly, it saves computer reading time. Any procedure th=t saves computer processing time is vital to an online HAI system. Fig. 3. Examples of keys and posting terms. The process of identifying KB entries is similar to the one described by N. Vleduts- Stokolov for specifying "concept codes" from word co-occurrences in the BIOSIS data- base (Vleduts-Stokolov, 1982). The current method of selecting KB entries is based on a statistical analysis of the single- and multiword phrases that occur in large volumes of text (Genuardi, 1990). These phrases occur in text that (1) resides in the NASA database, (2) is indexed to a targeted thesaurus term, and (3) contains the candidate words or phrases with relative frequency. The phrases selected are phrases that would be used by the NASA MAI process as search keys. In general, the procedures selected for an MAI system's initial phrase delineation and analysis define what kinds of information need to be represented in knowledge base entries and how large an operational file will need to be. For example, the use of word stemming or phrase normalization could reduce the number of required entries. Likewise, the strat- egies used for disambiguating words and for analyzing relevancy can define the level of complexity required for knowledge representation and ultimately may dictate the kind of data structure that is used. In the particular case of the NASA knowledge base, when the trade-offs were considered, it was decided to keep all rules as simple as possible to keep 636 J.P. S[LVESTEegt al. the system's online response time as short as possible. By rules, we mean "if... then" statements. For example, if "In-102" is encountered in the title or abstract, then provide the Thesaurus term "INDIUM ISOTOPES" as a suggested term for indexer review. Or, if a word is hyphenated, then look in the knowledge base for the hyphenated form; if it is found, then read the Posting Term field; else (if it is not found) drop the hyphen and treat the hyphenated word as two separate words. Most rules in the NASA MAI system consist of rules that specify (1) if the search key is found and the Posting Term field con- tains NASA Thesaurus terms, then suggest the NASA Thesaurus term(s) for review by the indexer; or (2) if the search key is found and the Posting Term field contains an asterisk, then add the next word in the five-word array to the search key and look up the new search key; or (3) if the search key is found and the Posting Term field contains two zeros, then no translation to NASA thesaurus terms is wanted for that word or word combination. Some MAI systems have more numerous rules that will examine instances of capitaliza- tion of words in the key or look for specific words in close proximity to a word in the key as part of the "if" statement (M. M. J. Hlava, personal communication, October 13, 1992). For example, if the word titanic occurs, and if it begins with an upper-case T, and if the word ship occurs within four words of Titanic, then return the term U.S.S. Titanic for indexer review. The NASA system designers chose to forego such details in the interest of minimizing the reading and writing required of the computer and thereby maximizing the speed of processing. APPLICATION PROGRAMS Each new use for MAI requires an application program that • Identifies the source of the text to be processed; • Delineates word strings found in natural language text by establishing boundaries or parameters; • Removes parentheses unless they are embedded (i.e., unless there are characters on both sides of a parenthesis); • Checks any word with an embedded hyphen (-) or virgule (/) against the first words in the KB keys and, if not found, drops the embedded symbol and treats the word as two words; • Calls Access-2; • Receives the MAI output for that particular application; and • Writes out the suggested terms plus any reports specified. Input can be any pertinent text. Usually it consists of titles and abstracts; however, it can be the subject terms or descriptors from another organization's controlled vocab- ulary, material from indexed or unindexed documents, the first and last sentence of des- ignated paragraphs, an executive summary, or any other text specified by the user and identified by the application program. The program uses a table of about 250 statistically selected stopwords (see Fig. 4) and thought-ending punctuation such as colons, semicolons, and periods. As punctuation or stopwords are encountered, the string ends and it is ready for processing by Access-2. Note that the stopword list does not contain either a or the. To include these as stopwords would preclude the ability to recognize valid thesaurus terms and Use references that contain these words, such as BOMARC A MISSILE, A STARS, VITAMIN A (Use RETINENE), OVER-THE-HORIZON RADAR, and LOGISTICS OVER THE SHORE (LOTS) CARRIER. ACCESS-2 Access-2, which is a modular program, never acts by itself. It is always called by an application program. Access-2 was designed to replace the syntactic analysis which char- acterized the original Machine Phrase Selection program. In addition, it was designed to shorten the overall machine processing time by minimizing I/O (the transfer of data Machine-aided indexing at NASA 637 ABOUT DEMONSTRATED I.E PARTICULAR SUGGESTED ABOVE DESCRIBE IF PAST SUITABLE ACCOUNT DESCRIBED IMPLEMENTATION PERFORMED SUMMARY ACHIEVED DESCRIBES IMPORTANCE POSSIBLE TAKEN ACROSS DESIGNED IMPORTANT PREDICT TESTED ADDITIONAL DETAILED IMPROVE PREDICTED THAN AFTER DETERMINE INCLUDE PRELIMINARY THAT ALLOW DETERMINED INCLUDED PRESENCE THEIR ALLOWS DETERMINING INCLUDES PRESENT THEM ALONG DEVELOP INCLUDING PRESENTED THEN ALSO DEVELOPED INCREASE PRESENTS THERE ALTHOUGH DIFFERENT INCREASED PREVIOUS THESE AMONG DIRECTLY INCREASES PREVIOUSLY THEY AN DISCUSSED INDICATE PRODUCE THIS ANY DOES INDIVIDUAL PRODUCED THOSE APPROPRIATE DUE INTEREST PROPOSED THROUGH APPROXIMATELY DURING INTO PROVIDE THUS ARBITRARY E.G INTRODUCED PROVIDED TOGETHER ARE EACH INVESTIGATE PROVIDES TOWARD AROUND EFFICIENT INVESTIGATED PROVIDING TYPES AS EFFORTS INVOLVED RECENT TYPICAL ASPECTS EITHER INVOLVING RELATED UNDERSTANDING ASSOCIATED EMPHASIS IS RELATIVELY UNIQUE ASSUMED EMPLOYED ISSUES REPORTED UP AVAILABLE ESPECIALLY IT REQUIRED UPON BASIS ESTABLISHED ITS REQUIRES USED BECAUSE EVALUATE KNOWN RESPECT USEFUL BEEN EVALUATED LESS RESULT USES BEING EXAMINED MADE RESULTING USING BEST EXAMPLE MAJOR RESULTS VARIETY BETTER EXAMPLES MAKE REVIEWED VARIOUS BOTH EXISTING MAY RTOP VERSION BUT EXPECTED MEANS SAME VIA CAN EXPERIMENTALLY MORE SELECTED WAS CARRIED FEW MOST SEVERAL WE CAUSED FOUND MUCH SHOULD WERE CERTAIN FULLY MUST SHOW WHEN CHARACTERIZED FUNDAMENTAL NECESSARY SHOWED WHERE COMPARED FURTHER NEED SHOWN WHICH COMPLETE GIVEN NEEDED SHOWS WHILE CONSIDERATION GOOD NOT SIGNIFICANT WHOSE CONSIDERED GREATER OBJECTIVE SIGNIFICANTLY WILL CONSISTS HAD OBSERVED SINCE WITH CONTAINING HAS OBTAIN SOME WITHIN CONTAINS HAVE OBTAINED STATUS WITHOUT CONVENTIONAL HAVING OCCUR STUDIED WOULD CORRESPONDING HERE OTHER STUDIES YEARS COULD HOW OUR STUDY DEFINED HOWEVER OVERALL SUB DEMONSTRATE IDENTIFIED PART SUCH Fig. 4. List of stopwords used with ACCESS-2. between the central processing unit and the KB or any other file). This is achieved primar- ily by eliminating lookups to determine parts of speech for each word encountered and lookups to find the appropriate grammar rules to follow-that is, by eliminating parsing. Speed is also achieved by keeping rules as simple as possible and reducing the number of unneeded empty spaces in a record that must be read by the computer. The fewer times that flies must be accessed, and the fewer times that information must be written to a file or to a monitor, the shorter the response time. With the original MAI system, including the Recognition Dictionary, the Machine Phrase Selection program, and Access-1, a title and a 250- to 300-word abstract could be processed in approximately 1.5 minutes. How- ever, syntactic analysis was found to be unnecessary for quality performance of the system. With Access-2 the response time averages 6 to 7 seconds for the same amount of text. When strings have been delineated, Access-2 identifies semantic units contained within the string. The semantic unit in the NASA system is normally limited to a maximum of five words to ensure grammatically correct word associations without parsing; however, the system can handle longer units if the words are consecutive. Search keys of fewer than five words must be created from within a five-word segment of the machine-selected string. 638 J.P.SILVESTER etal. This five-word proximity limit was established empirically and represents the best trade- off between identifying the most semantic units while limiting the risk of inappropriate word concatenations. When Access-2 accepts each word in an input string from the application program, and before it begins to identify semantic units, it places each word into its own array cell. Identifying semantic units is done as follows: • The computer-selected strings are examined, from left to right, in five word seg- ments, beginning with word one and word two. The first word of every word combination is checked against the KB to see if it exists. If it does not, the word is stored in a list of "Words Not Found As First Word in a Key" and printed out for indexer review. • If word one followed by word two is found in the KB as a key to an entry, the post- ing term field of that entry, which contains the equivalent NASA thesaurus term(s), is read. There are three possibilities (see Figs. 1 and 3): • The posting term field may contain two zeros (00), which will generate no NASA thesaurus term or terms; or • The posting term field may contain one or more thesaurus terms that are equiv- alent or slightly broader in meaning to the key and that will be provided to the indexer as suggested indexing terms; or • The posting term field will contain an asterisk (*), which causes the program to look for an additional word (within the five-word segment) that, when added to the two previous words, will match the key of another record. • If word one followed by word two has an asterisk in the posting term field, and this combination followed by word three, or four, or five does not find a matching key in the knowledge base, then the program adds 999 (which sorts last) in place of the final word and tries that combination as a key. If that is not found, the final word in the candidate key is dropped and replaced with 999. This procedure is repeated, if necessary, until the key is reduced to the first word and 999. • If word one followed by word two is not found in the knowledge base, then word one is looked up with word three. • If word one has been tried with each other word in the five-word segment and no key leading to a thesaurus term is found, the computer looks up word one followed by 999 to see if a thesaurus term is provided for the single word. This is possible for single words that represent specific indexable concepts. • When the process has used or rejected word one, the five-word segment is again measured off, beginning with word two. • Once a word is found as part of a KB entry, it is "poisoned"-that is, it is stored with a flag appended to it until the processing has passed that word. A poisoned or flagged word may not be used again unless an unpoisoned word is added to it. (See the following example of AERODYNAMIC CONFIGURATIONS.) • If word one and word two are found in the KB and word three is and or or, the last word in the key is dropped and the first word is combined with the word that follows and or or to form a new search key. For example, consider the following five-word segment: "aerodynamic configurations and properties that..." Look up search key "AERODYNAMIC;CONFIGURATIONS". Find the thesaurus term "AERODYNAMIC CONFIGURATIONS". The program "poisons" (or flags) "AERODYNAMIC" and "CONFIGURATIONS". These words may (now) be used again only if combined with a new word. The next word is "AND". The program drops "AND" and concatenates the word "AERODYNAMIC" with the next word, which is "PROPERTIES" and which has not yet been poisoned or combined with "AERODYNAMIC." Machine-aided indexing at NASA 639 Look up the search key "AERODYNAMIC PROPERTIES". Find the thesaurus term "AERODYNAMIC CHARACTERISTICS", poison the word "PROPERTIES", and conclude the processing for word 1, "AERODYNAMIC". The next five-word segment is counted off, beginning with the word after the conjunc- tion, and the process begins again with "PROPERTIES" (now poisoned) as word I. A more complete example of processing text with Access-2 isillustrated in the follow- ing example. Given the following title and sentences from an abstract of a document: Helicopter Noise Acoustic data for a 40 percent model MBB BO-105 helicopter main rotor were obtained from wind tunnel testing and scaled to equivalent actual flyover cases. It is shown that during descent the dominant noise iscaused byimpulsive blade-vortex inter- action (BVI) noise. In level flight and mild climb BVI activity isabsent; the dominant noise is caused by blade-turbulent wake interaction. KB file entries needed to process the sample input, and some related KB entries have been extracted and are listed in Fig. 5. Word strings are delineated by means of stopwords or any thought-ending punctua- tion such as a period, colon, or semicolon. This delineation process produces the follow- ing word strings from the foregoing title and abstract: 1. helicopter noise 2. acoustic data for a 40 percent model MBB BO-105 helicopter main rotor 3. from wind tunnel testing and scaled to equivalent actual flyover cases 4. descent the dominant noise 5. by impulsive blade-vortex interaction (BVI) noise 6. in level flight and mild climb BVI activity 7. absent 8. the dominant noise 9. by blade-turbulent wake interaction Key Posting Term ACOUSTIC;DATA ACOUSTIC;DATA;CAPSULE ACOUSTIC PROPERTIES ACOUSTIC;DATA;999 ACOUSTIC PROPERTIES BLADE-VORTEX;INTERACTION BLADE-VORTEX INTERACTION BLADE-VORTEX;TURBINE TURBINE BLADES BLADE;VORTEX BLADE;VORTEX;INTERACTION BLADE-VORTEX INTERACTION BLADE;999 00 BO-105;HELICOPTER BO-105 HELICOPTER BO-IOS;HELICOPTERS BO-105 HELICOPTER CLIMB;999 CLIMBING FLIGHT DATA;999 O0 DESCENT;999 DESCENT HELICOPTER;NOISE AEROACOUSTICS,AERODYNAHIC NOISE,AIRCRAFT NOISE HELICOPTER;ROTOR HELICOPTER;ROTOR;NOISE AEROACOUSTICS,AERODYNAHIC NOISE,AIRCRAFT NOISE HELICOPTER;ROTOR;999 ROTARY WINGS HELICOPTER;ROTORS ROTARY WINGS TURBULENT;WAKE TURBULENT WAKES TURBULENT;WAKES TURBULENT WAKES TURBULENT;999 TURBULENCE WIND;TUNNEL WIND;TUNNEL;TEST WIND TUNNEL TESTS WIND;TUNNEL;TESTING WIND TUNNEL TESTS WIND;TUNNEL;TESTS WIND TUNNEL TESTS WIND;TUNNEL;999 WIND TUNNELS WIND;TUNNELS;999 WIND TUNNELS Fig. 5. KB file entries needed to process the sample input. 640 J.P. SILVESTER et al. In the following example of Access-2 processing, references are made to the input array and the KB entries just shown. For a flow chart showing the processing logic, see Fig. 6, which explains how search keys are formed from word strings. It may also be helpful to refer again to Fig. 1, which provides an overview of the system. Processing Descriptions and Outcomes Mark off five-word array in title. Outcome: Only two words exist; therefore, the array is "Helicopter Noise." Look up search key "HELICOPTER;NOISE" in KB. Outcome: Key found. Post- ing term(s) "AEROACOUSTICS,AERODYNAMIC NOISE,AIRCRAFT NOISE" returned. No more words exist in the title. Move to the first string in the abstract. Outcome: The first MAI-selected string in the abstract is "Acoustic data for a 40 percent model MBB BO-105 helicopter main rotor." Fig. 6. Process for forming search keys from word strings.